The CORRECT way to gain FREE FPS with any AMD GPU

[@ThaWiz11]

I like how he just casually uses a 65 inch monitor

[@dylanherron3963]

Right, my man just "Hangs out" in 8k res.

[@Ligmaballin]

I too use a TV on my gaming desk, i can switch from games to movies with alt + tab and it's wonders

[@carloareaser]

@@Ligmaballin TV screens are bad for gaming. only good for watching movies and doing office work.

[@Ladioz]

I think it's great for KZbin presentation

[@Kanoog]

@@Ligmaballin Mon > TV

[@drewmangan8542]

I've been an undervolt person for ages. My focus is trying to get my GPU to perform as it did at stock at lower temps. I'm not willing to lose performance, but I want to optimize what it takes to get there.

[@1337Ox]

Same here! I managed to undervolt my RX580 from 1150mV to 1070mV, while keeping the stock clock of 1411MHz, but the max power usage dropped from 170W to 130W, and the card never reaches 60°C, fantastic!

[@bassdeff8819]

Same. I like running things efficiently and quietly. Lower Temps means lower fan speeds so better performance per watt is always my goal.

[@Brabant076]

Can it help with reducing coil whine? I want to buy the 7900 XTX but I've read story's of very loud coil whine.

[@Habixus]

@@Brabant076 it can help you can't know. I was lucky to get a 3070 and 3080 right after release and the 80 had massive coil "rattle". One of my friends wanted to buy either card from me for the price I paid and I almost gave him the 80 due to the noises it made! I then used the video from Optimum Tech about undervolting and reduced the 3080 clockspeed from 1950 to 1790 while dropping the voltage from just above 1000 mV to 830 mV. The card basically went fully silent! Performance loss is 1to1, so 1790/1950 = 8% in theory. BUT the games that ran into the power limit and would've needed that performance made the clock speeds drop to 1750 MHz anyway... So in the end I gained 40 MHz in the power limit while making the gpu silent and drop the average Watt when not at the limit about 120W from 330W to 210W!

[@SagoFanHD]

@@1337Ox I did the same thing with my old rx 570. 990mV at 1284MHz (Stock) instead of 1150mv, which was too much.

[@danepoitra9146]

Glad you revisited this topic so quick. I’m used to only overclocking nvidia, so it was interesting seeing the amd side.

[@AdamantMindset]

@@standarsh8056 most just watch him for entertainment, I highly doubt anyone not having amd GPU is specifically going out to search for this topic.

[@erwinmatic5062]

Ancient gameplays is the youtube channel you want to watch when it comes to AMD overclocking and adrenaline software stuff.

[@bretttanton328]

Same!!

[@yourmomsleftcheek1471]

Fr

[@oscar21781]

@@erwinmatic5062 Ancients has all the AMD info for sure

[@dondiegotorres3093]

Just had this video on my home page and decided to watch it. Overclocked my RX 7900 XT and got the same score at the Auto OC on the 7900 XTX. Granted, it's been a year since this video came out, and AMD has been doing really well pushing out driver updates to improve performance. I got my 7900 XT july last year for $700, and I couldn't justify spending another $300 to go for the XTX. Seeing how my overclock went, I'm glad I didn't go with the XTX because I'm getting the same performance with my overclock!! Been watching your videos for years, and I'm not the one to comment much. Love the videos. Keep it up man!

[@TheNiteNinja19]

Infinity fabric speed makes a ton of sense, that's a great observation. Maybe when you're pushing the memory too far it gears down to 1:2 or something like that just like the CPUs.

[@heatnup]

That’s not it. It’s the new desynced shader clocks they are using this generation! They are super power intensive so when you take PL headroom away those start downclocking in the background!

[@LazyEngineerYT]

No! Infinity fabric works between GCD and MCD here not like regular Zen cpus.

[@longjohn526]

That's what I said in his last video ..... Technically it's Infinity Cache on a GPU but it works kind of the same as Infinity Fabric on a CPU. Chiplet design does NOT GAIN YOU PERFORMANCE, it only lowers the cost of manufacturing but you are going to take a hit to performance compared to a monolithic design. In AMDs own literature they said they had to raise the GPU clock frequency 200-300MHz to get back some of the performance lost with the chiplet design. That should tell you that AMD is already pushing the clocks as high as possible out of the box. Much like Ryzen CPUs it's going to take another generation or two for AMD to tune the Infinity Cache to get as much out of it as possible, perhaps get it so 100 MHz increase in clocks evens things out. Even as good as Infinity Fabric is on Ryzen CPUs there is still a limit on RAM speed and Intel can do a few hundred MHz more than AMD because of their monolithic dies

[@PRiMETECHAU]

probably controller clock speeds downclocking but you can't see that value because AMD said no... maybe we will get access to that value in future.

[@MLWJ1993]

@@longjohn526 You can technically gain more maximum performance if you combined multiple GCD's, but that's not something they did this time, mainly because the interconnect isn't fast enough for that (not enough bandwidth). With a monolithic die you'd have so many defects at such large sizes that it's going to be close to impossible to match such a chiplet design (if you're capable of manufacturing a GPU with multiple GCD's).

[@Sir_Defyable]

For those who are familiar with PBO tuning, you gain more boost clocks when you can keep the temps lower. The “curve optimizer” is a good example of how that works, since you allow higher clock speeds at lower voltages, and get less heat from the lower voltage so you also boost higher. The GPUs work the same, where you just have to keep the internal boosting move the clocks by finding ways to lower temps and/or voltage

[@bpdave8223]

Hi, I'm trying to remember how Can I Do OC scanner with MSI afterburner on my RX 480... Newer MSI AFTERBURNER can't gave that for me at this point, so I' might have to ditch it. So If you can tell me how to Apply PBO thing to my RX480 that would be sick.

[@rustler08]

It's not always about temp. On watercooled GPUs, you'll see silly-low temps but still benefit from undervolting.

[@LivingTheBestDayOfMyLife]

​@rustler08 Agreed. I have been using a 3080 FTW3 with watercool kit. It runs ice cold with the undervolt and I can push it surprisingly far beyond stock. Performance is maybe 5%-10% less than overvolting at 400W. With the undervolt it runs around 250-315W and WAY cooler.

[@nsuinteger-au]

I think there is misconception about undervolt. The whole point for me is to run the system cooler with lesser wattage while getting stock or slightly above stock card’s performance. Which has been the case with nvidia gpus so far.

[@darnon2031]

And in regards the 'why not just buy a lower tier card if you're going to undervolt' sometimes the higher tier card is still faster. A 3080 at 280W (83% power) and 5% less framerate handily outclasses a 3070 Ti at the same power draw.

[@thesaltlick4057]

I gained ~9% fps by undervolting my 6900xt to 1065mV

[@xXSinForLifeXx]

The thing with AMD cards is undervolting gets you more performance tho

[@scimitar4211]

@@thesaltlick4057 can you share other settings you changed? I have a gigabyte 6900xt gaming oc

[@TheAnoniemo]

That is the idea yes, but with the current GPU boost algorithms an undervolt will result in lower temperatures and lower power consumption, leading to higher boost. And with NVIDIA (at least before 4000 series) the cards were power limited, so undervolting would drop the power for a certain clockspeed, allowing a higher clockspeed before hitting the limit. Thereby also increasing performance.

[@chrispritchard7977]

I'd love to see you experiment with water blocks on these. -Removing the fan power draw leaves a bit more power on the table -Eliminating temperature concerns I'm curious as to what your results will be.

[@MrPhyrce]

From my own experience. The Liquid Devil comes with an EK block and its almost identical behavior as the Red Devil. Same board/bios as the Red Devil just under water.

[@goldenx86]

Downclocking of GPU and VRAM is an issue with RDNA series cards, it affects us a lot with yuzu, losing over 40% of performance if you don't manually force a high VRAM clock speed (this is caused by the variance of an emulator workload with a lot of overhead). You can enforce max VRAM clock speeds with video encoding, set instant replay and record desktop to enabled, the encoder will enforce higher VRAM clocks for the video buffer.

[@rufusaa]

The quirks of the new gpus and their overclocking potential is certainly going to be interesting to learn about over the next couple months.

[@gm08351]

this guy needs to script and edit his videos. it's impossible to watch him stutter and ramble

[@bjornna7767]

@@gm08351 H'mkay, I like it the way it is.

[@TeddyRekk]

@@gm083513.62m followers say otherwise

[@zartul]

Well, on 5000 series, the advanced showed percentage, but the fine tunning options in fact gives you a graph as values, and basically you can redo their voltage/fan curves. And a thing I personally found out, is that you can do the Auto Overclocking/Undervolting, then switch to manual, and it will keep the setting of the auto that you just did, but you can now fine tune it and control the fan curve.

[@W3NNIS]

I got a 5700xt, can never get any stable OC settings, can I ask you what you're running and you're settings?

[@saman1820]

​@@W3NNISI'm running 1087mv 2050 core 1850memory

[@kravenfoxbodies2479]

@@W3NNIS Biggest thing about the RX 5700 and XT is how many 8 pin connectors it has, my reference XFX RX 5700 had two 8 pin connectors and it took the XT bios and would still auto overclock stable to 2049mhz, reference used the highend XT boards with better power phase design. not all RX 5700 XT are the same power design. Someone showed a reference RX 5700 XT in a video and it had one 6 pin and one 8 pin connections, that will never do for overclocking and makes me think it had weaker silcon on board and why the 6 pin connection on an XT.

[@frostfieldgaming4725]

I UVd my 6800xt just 25mV while setting max to 2600.. my Timespy score tanked by 2000pts compared to stock!! WHY?

[@raresmacovei8382]

@@W3NNIS 5700 XT doesn't have a lot of headroom, neither on the memory, neither on the core due to the 1.2V limit. My card can only OC reliably to 1830 MHz on memory and 2080 MHz on core. Some games will accept 2120hz, but not all. Going beyond 1.2V allows higher clocks. For undervolt, I'm usually maining 2000 MHz at 1.112V.

[@b00marrows]

Glad you got to see/show the AMD driver recovery. As i mentioned yesterday, i believe AMD's own built in OC tools are targeted are being simple yet functional. 1 button overclocking for the every day user as an option. I will also note that there hasn't been any actual driver releases since before the 7000 series released. Meaning the users are running on press drivers AFAIK.

[@chillhour6155]

They're software is very sensitive to crashing though, I played cyberpunk 2077 with a OC using afterburner with zero crashes on my first playthrough, then tried Amd's OC software hoping for less overhead and maybe slightly better performance, but the game crashed(flatlined) for the first time within 10 minutes of play

[@Rov-Nihil]

@@chillhour6155 not just that, forcing reverting to default settings after some failure... meanwhile the only thing i changed is the fan curve lmao

[@TheAnoniemo]

For the "perfcap reason" and other sensors, they do show in HWInfo64 along with a bunch of other stats. You can also view graphed data in it. I highly prefer HWInfo64 over GPU-Z.

[@1337Ox]

Agreed, you can even see Memory Errors if you push your VRAM clock too high

[@demonicdonut22]

HWinfo is definitely superior.

[@Mauser1965]

I was just trying to remember while watching, "Isn't all that shown in HWiNFO64..?"

[@DeathRow1985]

No not really. For example, the radeon 6600 non xt from powercolor gives you nothing.

[@ToxicAntidote]

I'm glad I learned about HWInfo64 right now.

[@mrtwisty9262]

Always appreciate the thoroughness and follow up from your site. If you find a problem that clearly as it shouldn't be, you dig to find out why instead of just saying it's trash and ignoring it afterwards. This helps tremendously as I just switched from a EVGA 3080 to a 7900XT Red Devil, my 1st Radeon card since the early 2000's.

[@davidellis1355]

This is actually really helpful for me, built my first gaming PC in about a decade, definitely out of the loop, after all this info I've decided to just use the default OC and leave it at that. Cheers Jay

[@Pisscan]

How is AMD treating you? 💀

[@ElementalDonnie]

@@Pisscan Totally not warranted. Please take your childish trolling/bootlicking else ware. *coming from a 3080 user*

[@CabbageShooter]

@@Pisscan Probably good? Been treating me good for years as has Nvidia until these 2 grand gpus lmao

[@Pisscan]

@@ElementalDonnie I asked the AMD user for their personal opinion. Please stay in your lane.

[@Brabant076]

@@Pisscan Lmao 💀

[@theshadowcult]

Hey Jay, any chance you could write up a step by step guide on undervolting cpu and gpu for dummies? Something that explains the process at each step, not just tells you what to do.

[@wilihey1425]

theres many guides online already for all sorts of cpus and gpus

[@FoShizzleMyPizzle]

@@wilihey1425 That's not what he asked.

[@wilihey1425]

@@FoShizzleMyPizzle there are literally guides for undervolting cpus and gpus step by step

[@mikidof]

I mean, just do what Jay is doing in this video. Start off with default settings, then try lowering your voltage while doing test runs. You should be able to lower voltages at least 25-50 mV and you should see some improvement in temps as well as clocks. The more time you're willing to spend, the further you can push your tune.

[@DRSKINGDAVID]

@@wilihey1425 And what that mean???? He obviously trust what Jay is doing thus the reason for asking him to do it.. I hate seeing comets like yours, he didn't ask you.. most people don't trust everything they see online.. he asked who he felt comfortable with watching.. move around dude..

[@bmgjet]

Im working on a bios editor similar to the ampere bios editor I released. So far iv found there is 3 main power targets and logic. If you dont touch the memory or core clock speed then it will target a total board power limit. You manually push the core clock like what you have been doing then it will prioritize the core as getting the main power budget. And like wise you manually push the memory clocks then it will prioritize the memory as the main power budget. Which is also why you cant enable auto overclock on both of those at the same time.

[@speedreaper9378]

To properly overclock, start by sliding the power slider to its max setting. Next, set the max frequency to 3200, go to vram next. Enable fast timings, set vram clock to 2700. Then undervolt to 1050. DO NOT TOUCH minimum frequency. Now you can go back through and tinker with each setting until you find stability in all games you play.

[@Farrcast7]

And if its not stable then what do you do?

[@speedreaper9378]

@@Farrcast7 add voltage,reduce max frequency or lower vram settings. Each card will be different and you will have to find the right settings for your exact card.

[@poolshark28]

Interesting result. To me it seems like the firmware is lowering (throttling) the workload sent to the memory controllers to keep the power level in check, even though the clocks are high. It's a complex algorithm with many different factors... There's also 24 channels, six memory compute dies so maybe the average or the max die clock is high, but the active die(s) are lowering their clocks and not represented. I'll bring this up with the power team :)

[@WTFZOMG]

maybe at those clocks theres a bunch of correctable memory errors and the GPU is wasting a bunch of time correcting 1 bit errors, which would lower memory controller utilization and perf.

[@poolshark28]

@@WTFZOMG actually that sounds very plausible!

[@Laforge129]

@poolshark2880 Do you think there might be a bad timing chip somewhere on the board that is causing it not to show. I am thinking there has to be a hardware silicon chip that isn't up to spec and that is why @Jayztwocents is having issues with it not showing in the bar. It could very well be the firmware but my gut feeling is it is a hardware issue and not just software or firmware. I'd like to know what is really causing this but we may never know.

[@poolshark28]

@@Laforge129 it's not a silicon problem. I'll ask the team at work but not likely results that would be shared publicly ...

[@Gan_Gineandro]

@@poolshark28 Any new insights?

[@baumstamp5989]

the 7900XT is my first card after a 1070 and i am rly blown away. sure would have been blown away by a 4090 even more but i come from 1070 budget (which was 450 bucks back then) and didnt want to go quadruple the price.

[@lucidnonsense942]

Fast timings is just tighter memory timings, if you are running stock memory speed you'll be fine with fast, unless the silicon lottery really skunked you. If you're overclocking, the default will let you get to a higher frequency. You then have to test if the lower frequency with tighter timings is more performant than higher frequency with loser timings. It's usually one or two percent difference between them, except for some edge cases.

[@joshuaross8471]

This is why I love J2C so much, he is engaging with the audience. Asking the viewer for help and bringing them into the conversation. Jay you rock!

[@-opus]

Ancient Gameplays has a complete run through of these settings if you need a few pointers Jay ;)

[@jamezxh]

Fabio’s an AMD overclocking Guru

[@757Bricksquad]

He does a terrible job explaining those settings.

[@jvidia]

​@757Bricksquad who does?

[@TiagrajI]

Yep he did. That guy is a gift

[@lakecityripper]

He is great honestly the reason I never had the issues many had with the 5700xt.

[@kevthomas7293]

Reckon interlink between MCD & GCD or between MCD to VRAM is dropping from 1:1 to 1:2, hence Jay's video showing memory controller load halving from 32% to 16%. AMD built some load protection balancing for stability.

[@AnIdiotAboard_]

No they didnt they did what they always do, have always done and will always do, 20+ years of AMD and ATI Cards :) This behavior is in ALL AMD Cards and its caused by a lack to total power, AMD Do NOT like (nvidia) slap a 600 watt power stage setup on a 300 watt card, if they need 300 watts power stages provide N + 10%. Then again in my life time I've never really overclocked GPUs and thanks to mining i got undervolting down to an art form :)

[@phhan0]

RDNA overclocking can definitely be a pain, but as a general rule, you want to tune memory first (fast timings when possible). Then you can tune your speed by dropping your voltage. It’s definitely more complicated than that, but what Jay was doing in this video is a lot closer to what you want to do. Ancient Gameplays has a few vids on it, but it you want to try to fine tune it, you increase the uv, then move the min/max sliders to within about 100-200 mhz of each other until you get stability/ fps drops. You’ll hit your lax uv and clocks. Then you usually want to pull back from those maximums for your stable OC. You can drop your min clocks then too if you want.

[@panikk2245]

Is overclocking safe? And is it worth it when I'm already playing pretty much maxed?

[@Demmrir]

@@panikk2245 This rule applies to both overclocking and buying new GPUs: If you're already maxing out your FPS and resolution for your monitor and your settings for the games you actually play, there's no need to upgrade any more. You can if you just want to, of course, but there's no reason to.

[@moomeuh1342]

Yay Ancient Gameplays did great videos on this topic, he deserves a lot more subs Just be carefull when you use RayTracing GPU usually need more juice for it, so you can have a stable UV/OC in full raster that crashes with RT enabled

[@phhan0]

Exactly what Demmrir said. If you’re already maxed, don’t worry about it. Squeezing an extra 5-10% isn’t going to do much for you for the time it’ll take to fine tune your GPU. It’s only if you want/need to push the limits to get the performance you want or you’re an enthusiast like me and just want to see how hard you can push your card. As Jay said before, you should never buy a GPU for its overclocking potential. The AMD software is pretty good about stopping you from killing your GPU, as long as you don’t try a massive overvolt, power limit increase, overclock. Less is more for overclocking.

[@hydromisc]

Every time I try to tune the vram frequency on my 6600M it locks the max core frequency to 500 mhz. I can only overclock/undervolt core for some reason. Power limits and fan control are fine

[@michahojwa8132]

Jayz good to see you admit to messing up those oc and improve fast. Put power limit to +15% and then undervolt for higher performance. Set memory to fast timings. Test both rt and non rt game, the oc is different if AMD rt muscle is resting. Mine 6700xt did 2,8ghz and 2,9ghz oc on non-rt games. You need to oc memory a lot for small fps gain. Cheers

[@solisbane]

PREFACE: I have the XT and not the XTX I had the best luck maxing power limit no matter the settings i used after this. I found that i kept gaining in FPS and timespy (dont have port royal) score increasing memory until crash at 2750. 2751 produces instant crash on my card no matter what every other setting is 100% of the time. I left it at 2750 for the rest of this and tried decreasing to see if i would gain performance after the further testing. I then would increase max GCD clock until i would see a drop in performance/timespy score. I would drop 10mV which would let me push the clock harder with performance increase. Once i maxed out what i could push that way i tried dropping the memory clock by 10 and instantly lost performance which did not come back with more GCD clock speed. Overall i ended up at core max: 3150 | voltage 1010mV (minus 90) | memory 2750 | (again on the XT). This pushed timespy from 19936 to 26327 and the 2 main games i tested in CP2077:107>124 and RDR2:141>158 at 1440p Now i hope they fix the drivers to make VR better because that's the entire reason i bought the card. My 6700XT was fine for flatscreen stuffs :( PS: You're def my favorite youtuber mixing PC and gearhead stuff all the time. I feel like you would just be a cool dude to shoot the breeze with about either.

[@edyluchian]

Hi man, i have a 7900xt and a ryzen 7 7700x, i bought on the system a PSU corsair tx750m with 750W, can you help me with some advices how i can overclock a little bit this gpu to be safe considering the amount of power of the psu. Thank you very much!

[@RicoZaid_]

thanks!

[@solisbane]

@edyluchian I'm running an evga 750w with a 5800x3d and got those numbers. Remember results vary by card you'll have to test but the test methodology should stay the same

[@Symthisis]

Got my 7900 XTX to overclock to 3095MHz from just using their auto overclock option - pretty crazy.

[@TheHavocInferno]

Jay, a little explanation on the Minimum Frequency at the start of the video. Yes, pulling it up will seemingly force the card to run at higher frequency. But, it becomes a problem when the card runs into a power limit at which it can't sustain the set Minimum Frequency. That's when it will internally downclock/clockstretch anyway, degrading performance. The power management simply doesn't allow enough power draw to keep a high min clock stable. Leave Minimum Frequency relatively low, far enough down that it can still be sustained in a really power-heavy workload. I mean you can already see that at maxed power limit, the card can't really do more than about 2600-2700MHz in Port Royal. Trying to force it to run at a higher minimum clock than that can't work (unless you can disable the power limit). Intermittent load switches in the millisecond range probably even run below 2500MHz. On RDNA and RDNA2, you realistically could only push Min clock up if you gave the card vastly more power limit via MPT.

[@HeyItsKesh]

Literally just bought an AMD GPU yesterday. Appreciate ya Jay!

[@-opus]

Same here (6800), Ancient Gameplays has a lot of AMD videos

[@tarell676]

First thought I have is that the RDNA3 with the chiplet design is a first gen, so there might be some oddities in the beginning like we saw with Ryzen fist gen. Like you pointed out. So for my part, I'd love to see you revisit this issue in a few months when the drivers have matured a bit more and maybe AMD/Radeon could have found the source of the issue and corrected it.

[@colten6996]

Here we are a few months later and it’s still the same issue. Can’t raise minimum clock speed without tanking fps

[@nickelwrangler446]

Nah if its like ryzen first gen we'll have to wait til RDNA 5/6

[@anthonygoldie6961]

Awesome video thank you I just went and bought a 7900xt and was starting to look for videos on overclocking undervolting etc and was thinking there would not be any good reliable ones out for a while. Very useful 🙂

[@smokeyjoeruff]

Ancient plays have the best video with overclocking the 7900

[@legros731]

Jay still got it wrong You shoud overclock the vram first until you see a drop in performance after findings the max vram speed you can begin to raise core clock and undervolt You should aim for the highest vram speed to gain max performance But maybe try more core clock and less vram speed as well it could be better each card is different afterall

[@fist003]

just go watch Ancient Gameplays for RDNA overclocking guide. you're not getting it here.

[@FeartheBoxes]

Rx 7900xt (Not XTX) Best results so far. 1. Increase Power limit to the maximum, which is common 2. The memory clock operates at approximately -20Mhz higher than the set value. (eg: 2640=2626) 3. For undervolting, slide the maximum clock up and observe the core voltage displayed on Adrenaline. The lower the core voltage, the longer the higher clock will be maintained. 4. For reference products, the power consumption does not exceed 400W. 5. There are differences depending on the product, but PL+15%, maximum clock setting of 3300 to 3400, memory 2640Mhz, core voltage 980mV, and maximum clock around 3Ghz sounds the best so far. Cooling needs to be increased. (Fan speeds) 6.AIO products can get more 200~400Mhz here. 7. You can use MSI after burner since it's more aggressive.

[@Pisscan]

Buying a 7900xt was your first mistake 💀

[@airsoftmanic]

After watching the initial video, i went into my 7900XT settings and fiddled with it for a few hours. After coming from an Nvidia 1070ti having to undervolt felt very wierd, but managed to get the XT to 3.1GHz GPU CLK and 2.6Ghz MEM CLK at 960Mv under timespy looped. So watching this is definitely proof that AMD and NVidia GPUs behave very different to overclocking.

[@mikidof]

I have the 7900 XT Hellhound and was able to get similar results to yours, but it's not at all stable in MW2. So I personally find stress testing with Time Spy pretty much useless, if it can't handle a super popular modern title with the same settings.

[@airsoftmanic]

@@mikidof indeed Timespy is not a real world test, after playing for a while was getting weird issues, but dialed back the memory and it's fine now. Didn't see that much difference in perf vs stock so I've just gone the other way and tried to make it as quiet as possible while retaining the stock performance.

[@zengrath]

Thanks so much for making another video so quickly. After reading all the comments saying to under volt and is the trick with AMD GPU's, I was curious to know if they where right or not. I'm impressed to see some actual decent OC now with this card your testing, at least in Port Royal, for gaming it still feels silly to me since you only gain a few FPS so it would have to depend on how stable it is for me. I have always found that OC'ing to gain a couple FPS but increase crashes or wondering if a crash in the game your playing is due to your GPU OC or another issue was never worth it for me. But if I buy an AMD GPU I may give this a try anyway and see what happens. This also goes to prove that "higher number" for Mhz doesn't mean higher performance in an actual game as shown when you try running voltage and everything maxed. I wonder how many people will buy this card, OC it so it to a higher frequency but not actually bench mark it properly and play games clueless that they made it worst, haha. So important to actually bench mark your favorite game before/after to verify it actually benefits then just trusting what the OC tool numbers are showing you and running a tool like Port Royal is good for stability testing.

[@insiainutorrt259]

For my 5700xt at least short tests dont reach the full temp clockdown threshhold so they dont show the full benefit of undervolt...

[@khaychi]

I just hope AMD is working some more on drivers and features until 7000 series drops to realistic prices. I would totally go 7900XT or XTX this gen if one drops to 700 to 800 euro. A slider for performance % like Afterburner has would also be neat. I also hope Intel Arc catches up one day so we have 3 major players. Just put more resources into driver, software and feature optimization. 👍

[@chrisbruchak4519]

I mean you can get an XTX for like 950 right now. Just how cheap do you expect a card that can play basically every AAA game in native 4K at over 120 frames to be? There is really only a handful of displays that can even fully take advantage of what this generation of cards has to offer yet you seem to think its too expensive.

[@KraNisOG]

@@chrisbruchak4519 Yeah, 700 seems extremely unrealistic, the card is worth probably around 850-950 imo.

[@jwat8400]

Just got a 7900xt for $750, buy while you can !

[@legendofthefall7082]

Just picked up a 7900 XT for $720 around Black Friday

[@ozawashere]

This has been happening for awhile now i remember trying to overclock my 580 a few years ago and wondering why nothing was happening

[@skitzdaniel3314]

I bought a RX 6700 XT a few weeks ago. It is a good example that you have to have a case and cooling to go along with it. I put the card in my mid-tower case. I started to run games and the GPU was hitting 80c, but junction temps were toward 110c. I am an engineer, an worked on electronic products for 18 years. A lot of components were rated at 104c for highest temp before problems. So I looked at the case. I added one fan. Put all fans directly to the power supply molex plugs. The mainboard fan connectors were to slow for cooling the case. So I have 3 fans air intake, and two fans air exhaust (including power supply). There are grills in the case the intake air can also escape by pressure differential. I set the Graphics card fan profile to on early at 60c to get ahead of GPU, and Junction temps early. The card fans hit a steady 2600 - 2900 rpm and keep the junction temperature down between 80c - 90c. The GPU runs about 65c, and the CPU 70 - 75 c. I happy with this level, and I don't smell heat, like to hot anywhere in the case. At 1440p this card performs well, about 120 - 200 fps depending on the Battlefield games I am playing.

[@ZFshmN]

If the infinity fabric is not the fault of the loss in fps, I believe what might be happening where the fps is dropping (both frequencies remain the same reported: game clock and memory clock, but when the fps drops, the shader clock is getting the short end of the stick. Essentially, the game and memory clocks are both being reported and shown properly, but when manual overclocking is being done, the shader clock is unable to be changed or viewed by the user, resulting in seemingly unchanged clock speeds, but also a drop in fps. This might also explain the lack of memory controller usage.

[@1337Ox]

That's what I think as well, and it makes sense. Too bad we cannot see both clocks (shader clock speed AND front-end clock speed), maybe in the future it will be possible with some updates.

[@ZFshmN]

@@1337Ox I'm sure AMD will make those changes to their software at the very least. I think that it was something they just overlooked. Even if you cant change the clock, they will definitely let you view it at the very least

[@BaLLisTiKaLFPS]

As far as the fan control being locked in anything besides custom tuning, you can just enable Auto OC, Rage mode, or whichever preset you want... And mimick those settings in a custom tune. You can save different tuning profiles. This is what makes Radeon software so much better than GeForce experience.

[@colesym84]

It could be something like the IF ratio on Ryzen. Once you push the core clock past a certain point, it drops out of 1:1

[@AnIdiotAboard_]

Jay on the undervolting you wanna see something impressive, there's a tool for download that can force the card to 1v or lower, you loose the top 10% of performance, but you save way more in total board power, and the cooling changes are staggering! Ahhh brings back memories of forcing GPU's to 0.975 volts, don't mind giving up a few frames for a seriously cooler operating card

[@bradclapp4022]

Can't wait to see Jay hook up the AC to 7900 XTX! Curious to see how it reacts to being Colder!

[@onomatopoeia162003]

Be interesting to see what the water-cooled ones can do.

[@LetsSeeYourKungFu]

March 2024: just completed my first build. Used guides by two other creators i wont name; massive issues. Microstuttering. crashes. freezes. Terror. Followed this video. Fixed every issue immediately. Almost doubled my performance. Would love a revisit to this

[@SethOmegaful]

Cool video, jay. It would be great if AMD could explain in detail these behaviors to us.

[@ayushagarwal9530]

Hey Jay, so according to me issue is not the memory, its the core because if mem load is decreasing it means it is not the bottleneck there (if memory clock was decreasing than mem load will increase you can test it by doing the reverse i.e. overclocking the memory for same test, mem load will decrease). So the other part that i can think of is core as it was only reaching max of 2700 mhz (when min core clock was set to 500) i think thats the max it can do stably so forcing it further introduces core streching, try hwinfo for checking if thats the reason ( ig it was effective clock which shows the actual core clock). Waiting for your next video on this 🙃

[@damianstevenson1196]

This happens with 6000 series as well. I found it happens mostly when I increase VRAM clock. I thought it was errors occurring within the memory causing the drops.

[@NeededByNobody]

That's so weird... My 6800 XT doesn't have this behavioral issue.

[@damianstevenson1196]

@@NeededByNobody both my 6900xt's do it. They won't crash they'll just start dropping performance.

[@1337Ox]

This also happens with my RX580. I can push my memory to 2300MHz and while it is stable, the performance drops, and actually in HW Info you can see Memory Errors. And they do happen. So that's why. It's not only with 6000 and 7000 series. I guess you can think of it like ECC memory, where it tries to correct the errors and loses performance, but is stable. So everytime you overclock memory, keep an eye on the memory errors.

[@NeededByNobody]

@@damianstevenson1196 If I am not mistaken it's like a memory lottery of some sorts, a majority of cards have this issue, but a few subset does not. I was having core frequency performance loss on air though, 2,500 was faster than 2,700, but when I went on water the issue was eliminated... Probably relating to temps.

[@damianstevenson1196]

@Menti Capti what does yours do?

[@thematrixredpill]

So glad you admitted you messed up last time then just went back and started undervolting and tweeking with software and just got on with it. Well done

[@antiunholy8781]

I think the memory controller went to a 2:1 ratio instead of a 1:1 ratio just like the ryzen cpu's. Memory speed would still be up, latency would go up creating loss of performance but the interconnect would have less active throughput showing a lower memory controller usage because it's waiting due to latency. Just my thought process; they specifically mentioned infinity fabric several times in the press release so it would reason to be the same interconnect methods.

[@snickerdoooodle]

What a timely video, I just bought a used RX 480 for $50 to refurbish a PC to give for Christmas, and I had to tune the voltage down 90mv to get it a bit cooler and stop throttling itself.

[@kagetora03]

hi jay, it would be cool to see some game benchmarks after u overclock the gpu to see how much performance gain translates from portal royal to actual games

[@DuBstep115]

Well 1100 point increase is 7% and 7% in game that's running at 80fps is 5fps. As he said, you can expect a 3-5fps increase.

[@Pisscan]

@@DuBstep115 5 whole fps .. wow 💀

[@dagrimmreepa]

@@Pisscan lol yeah, totally worth the effort!

[@Ro_Ball]

Hello Jay. I believe it might be a driver reporting issue. If I remember correctly, RX Vega also had an issue where you could overclock it to insanely high frequencies and the driver was reporting those frequencies but in actuality the card was dropping down to it's 2D clocks. I don't know if 3D and 2D clocks are still a thing now with how much frequencies fluctuate but it might be something that is software related and we can only hope that AMD will fix it in the near future.

[@lukashehler9901]

My experience with overclocking my 6700xt is that I was able to just max the vram at 2150, max the powerlimit and set the max core speed to 2800. This is mostly stable and I have tried lowering the voltage but that gave me less performance and less stability

[@tenorsaxhero]

My gigabyte eagle 6700xt gets up to 2777. I could fiddle with some dials and get better results but i noticed a significant increase in cyberpunk framerate (37 up to 65) which was with the adrenalin utility on a 1440p display.

[@bradmiles1984]

Biggest uptick in performance with the 6700 you can get is change the memory to fast timings.

[@hando87]

Interesting. I've got my 6700xt red devil running at 2600-2920 mhz, mem maxed, powerlimit maxed, at 1060mV... any higher and it would throttle down clocks. It's about a 7-8% performance increase for a 9% power draw increase.

[@D3ATHCHAS3R]

I came back to this video because what Jay says at like 16:50 totally misses me. My GPU drivers crash, then I have to reinstall them nearly every time without fail. I wish I knew what I was doing wrong with my 6700XT. MSI Gaming X 12G Manual OC to 2600MHz@1.05V VRAM tuned to 2100MHz +5%Power. Fan Curves up to max around 50c. Replaced Thermal Paste recently to drop temps, which helped that but not the crashing.

[@jadams2289]

New to OC but after watching this I was able to increase my heaven score by 6% on my 6700xt, Thanks Jay!

[@dirrti3rsanchez387]

What did you put your settings at? I have the same card.

[@jadams2289]

@@dirrti3rsanchez387 I have a gigabyte aorus card. I let it auto oc and wrote down those numbers then put them all in manually. Then, I bumped the core up by increments of 10 while running the benchmark in the background. It was something like maxed vram and 2794mhz on the core with a slight undervolt. I ended up turning it off though because it got really wonky during screen recording and video editing. Turns out Jay maybe right and the AMD cards are a lot more delicate with their clocks. Can't go wrong upping your fan curve and power limit though.

[@AzziesPersonalRecordings]

Honestly, your last video, I was so exhausted and your voice was so relaxing I passed out about a third of the way in. I found the information interesting, but after spending half the day at the Dr's office and being sick, I had nothing left in me. Thank you so much for the wonderful videos though! If I can get one of the new cards from this generation I'll definitely be going back to rewatch every single one of these videos.

[@rickthomas9894]

Thank you! Appreciate all your and the teams work and effort to explain this stuff.

[@mike26037]

Well Jay you did it again. I moved the minimum frequency up to just under the max and my fps literally almost tripled

[@jomz8892]

How? Did you use his setting in this video?

[@oscarbecker7843]

I am pretty sure that you're right about the memory speed tanking related to bus-speed dependencies, like with the CPUs. Another cause could be that at certain overclocking values the memory falls into a asynchronous operation, that leads to huge memory latency and/or to many waiting cycles!

[@jediknight4316]

Hey Jay that was me who mentioned not opening the vram tuning section lol. Thanks for revisiting the topic and taking AMD overclocking serious this time. In DCS I get increased performance with the vram set to fast timing.

[@philrdaly]

My 6800xt never changes the voltage from 1150mv if I try to overclock the frequency regardless of whatever voltage I set so its nice to see that the you actually have control over voltage with the 7000 series. The only way to get the voltage down for me would be to down clock it or use MPT. Always love the content Jay, especially the water cooling videos cheers!

[@elysian3623]

6800xt is a card that's far from power limited, it's a 72cu card if I'm not mistaken, most of the time they're equipped with a similar 3 to 4 slot cooler as a 6900, generally with 6000 series a 6800xt is basically a defective 6900 so it's still very happy to boost, the way that AMD uses boost is to reference heat and power head room, if you have tons of head room in both then it will use the voltage available to it to ensure that what ever speed it can reasonably boost to will be safe and stable, obviously board manufacturer bios and which bios is in play on dual bios cards will impact how the boost algorithm decides what to do. I have a 6900xt, specifically an Aorus master, I have inline capacitors to maintain clean power even in transience and power fluctuations, lots of things add up to the performance of your card, your card may be pulling that voltage for a number of reasons, boost may be one, if your card is hooked up with a pig tailed power cable can be another, amperage tends to be the limit on that set up as too many amps cause a hell of a lot of heat so voltage is added to compensate as it's easier to step it down and regulate it, For reference my 6900 runs perfectly in and around 2800mhz @1120mv & 310w worth noting that my card has a 1200mv cap, memory has been a secondary concern as oc on the older 6x00 cards is worse than a 6x50 from what I've seen so is run conservative @2120mhz on fast timing, This is an example of a card that's had every precaution taken to ensure it gets enough and consistent power delivery, individual 8pins on a 1000w common rail psu, sounds over kill but my performance speaks for itself, my now defunct gpu is sniffing at the heels of some of the new card coming out like the 4080 admitted it's not a hard mark given it's relative performance to a 3090 but it certainly shows that some of the high end cards will last for years, it's enough to be bottle necked on a 5800x3d which is nuts at 2160p.

[@moonwalkerr87]

I have a 6800 (non XT). I have set the voltage to 950 mV and the frequency to 2330/2430 (min/max). Memory to fast timings and frequency to 2100. Power to 15%.

[@philrdaly]

@Elysian My 6800xt is in a custom loop so temperature is never a problem. The best OC I can get is 2550-2650 and 2130 on the memory and my junction is at about 60. Just always thought it was weird that if I try to undervolt it the voltage won't budge off of 1150mv while overclocking. I'm also considering a 4080 or 4090 but that would also warrant a platform upgrade as well since my 5900x will bottleneck them (definitely didn't see that coming). Really the 6800xt is enough for me but it's coil whine is obscene and while it does improve with undervolting and frame limiting it still never gets to a volume that I can't hear. I've seen that some of the 4080 and 4090 models seem to have very little coil whine and since I spent all this money on a custom loop to have a quiet build, my screaming gpu made it seem pointless.

[@philrdaly]

@Samir Ben Karroum I can undervolt it if I set those kinds of frequencies down to 980mv but if I overclock it the voltage won't change

[@MrDutch1e]

@@philrdaly better make sure it's your gpu with the coil whine and not the GPU. My 6800xt has basically zero coil whine on my sf600 or sf750 but I tried the cooler master sfx 850 watt and coil whine was bruuuuuutal from the PSU. Returned it the next day and bought the sf750 which was essentially silent.

[@seanryu5214]

With the new Amd gpus you actually need to undervolt in order to overclock. There’s videos (2 I’ve seen so far) that perfectly show this. The way I understood it is that, at least with the 7800 xt, the amount of power the card draws is locked. So by undervolting you’re giving more headroom for higher clock speeds within that envelope of power. Gonna test this when I build my pc with hellhound 7800 xt

[@logan5938]

Youre an incredibly valuable source of information, thank you Jay!

[@ElysiumAim]

Finally, someone has done this research. I've been trying to figure out what's going on there for a year now. Thank you!!!

[@jujuliolezesuiglaz1780]

Bought a xfx merc 6800 xt during black friday, first amd card I ever bought and very happy with it. I like the adrenaline software for overclocking, the ability to set a custom OC profile for different programs is a neat feature eventhough I just settled for a global overclock. My only gripe is that every time I restart my computer it goes back to default and I need to load my profile (might have to do with fart startup, haven't yet tried to disable it) I focused on undervolting for better temps ergo better longetivity, along with an aggressive fan curve (I just don't hear the card fans over the ones in my pc case when gaming). It is also nice to reduce the power draw with the increasing price of electricity. But even for people looking for better performances, undervolting yield better results since lower temps prevent thermal throttling. In my case I went with a stable OC going from 1150 mv to 950 mv (could go as low as 930 in benchmark and most games but would crash in cyberpunk). very happy with the xfx card, I also noticed that the hotspot temp is always very close to the overall temp; the fans and thermal pads seem to be doing a very good job

[@Rol4nnd]

Just bought the same exact card. How much did temp changed after undervolting?

[@jujuliolezesuiglaz1780]

@@Rol4nnd sorry didn't do a comparison ^^ grats on your card, this is actually the best air-cooled model for temps. Depending on how lucky you are with the sillicon lottery you might even undervolt a bit further than I did

[@altctrlscream]

Any chance you could share a screenshot of your config in the AMD software?@@jujuliolezesuiglaz1780

[@JBG84]

Credit to you Jay for the maturity and swallowing of pride to make this video. Showed great character 👏👏👏

[@iiredstar7592]

Awesome video as always Jay! But yeah what intrigues me. I have been using AMD cards for a while now too and the "Fast Timing" thing is pretty recent. I had RX 5700XT and it did not have it at least at start. I think I saw it first time in 6000-series cards. I really would like to know what is the difference between "default" and "fast-timing"

[@lucidnonsense942]

It's tighter memory timings, if you are running stock memory speed you'll be fine with fast, unless the silicon lottery really skunked you. If you're overclocking, the default will let you get to a higher frequency. You then have to test if the lower frequency with tighter timings is more performant than higher frequency with loser timings. It's usually one or two percent difference between them, except for some edge cases.

[@kaoe145]

Also some games will let you. Run both high Frequency and fast timings. While other games will crash with fast timings even at a lower frequency. You have to test on ever game you play. And adjust setting that best for each game. Than save it as a profile. Also after updating/changing drivers you might have to Tweak the profiles.

[@williammurray7520]

To take a guess at whats occurring, its taking power from the infinity fabric between the MCU's and the Core. As the total power budget is shared between not only the MCU's and the Core but the fabric as well. Hence why people are seeing higher idle power consumption as it takes power to facilitate the transfer between the MCU's and Core. Once the Core and MCU's take the entire budget theres nothing left for the fabric to use, hence a loss in performance that Jay has shown here. Super interesting and thank you Jay for showing us another angle!

[@BNLoken]

Based on your results when undervolting, the card is power limited. Even with the undervolt you were staying at 375w. It seems amd made some last minute decisions to lower the tdp on the cards due to the backlash of the 4090 power draw. I think if you added more power to the card via more power tools utility you could get well over 3300mhz and maintain if not overclock the memory. Something like 500watt would probably get the card to perform like it’s actually capable of and hit the rumored performance targets

[@jagvirsodi8192]

This was my situation, more power tool increasing max power limit was a godsend. But annoying that I couldn't do it in afterburner or amd's own software.

[@BlackV4You]

@@jagvirsodi8192 How do i use this that you are talking about? What is it and how do i get it? Thanks

[@Turktien]

Love this stuff, Jay. Interesting to see the progression... I'm very interested in this topic. I had an XFX RX 480 8GB and now have a 6700 XT. On my RX 480, AMD's software had separate voltage controls for the VRAM and GPU. On the 6700XT (and on your 7900 XTX), they don't. On my RX 480, I could up the GPU clock to 1400-1425 no problem. Slightly sigh on the VRAM clocks or VRAM Timings setting? Crash. Now move on to the 6700XT. Similar situation to 7900 XTX but, when overclocking, I've noticed similar but different behavior. On my 6700 XT, I have the max slider set to 2750 and the VRAM to 2100. Here's the difference in behavior I noticed on mine as I was trying to find balance: As I upped the clock, like you, it had very little impact (no negative performance but I'll get to that). If I didn't touch the clocks but upped the VRAM, it did improve slightly. At high enough "Max GPU Clock" in AMD's control center, I would get memory artifacting. GPU OC = VRAM artifacting didn't make sense. If I set it to "fast timings," it would cause a GPU Hang/freezee and force a restart. So, ultimately, what I think I oops'd onto was that memory/GPU thing you're seeing here EXCEPT my VRAM clocks never dipped but I got similar behavior. I either got a freeze or VRAM artifacting. Freezing/artifacting applied to toggling fast timings options, too. Lowering the "voltage" under GPU fixes the artifacting and freezing and allows me to maintain an OC on GPU and memory both with a net gain in performance and reduction in temp (both fairly marginal though). I also have a 4600H/5600M laptop that uses AMD's SmartShift thing where it dynamically changes power availability between CPU and GPU for efficiency. (More CPU = more power to CPU, less to GPU and vice versa. So, I have a few theories but not sure based on the three different generations of cards I've had for AMD but may just be guessing. My suspicion is that power delivery must not be regulated the same on post Polaris chipsets and, on RDNA3, is being shifted dynamically similar to how they've done their mobile CPU/GPU chips in laptops. (Mine's a Dell G15 5505 or something.) RDNA1 and 2 seem to lack the shifting feature so stealing power for GPU ultimately robs the VRAM. That leads me to my other suspicion. On RDNA2, I think the "default" timings must be dynamic and "fast" must be a locked setting. So, on RDNA2, my suspicion is that increasing power draw on the GPU won't lower the actual clocks on the VRAM but it's loosening the timings to lower the power requirement. On RDNA3, what you're seeing looks far more akin to the SmartShift technology on their mobile setups. I'd love to see more tests on this but I think this is something I'd love to hear more from AMD about. I don't think it's a bad thing, I just wish there was more information out there so it made more sense what's happening. Anyway, I could be stupidly off base but I thought I'd share my thoughts since I've had the opportunity to kind of watch the evolution from RX 480 -> 6700 XT and now 7900XTX thanks to you so it kind of paints a picture when you look at those changes and behaviors gen over gen. EDIT: Also to note, my current 6700 XT is the XFX QICK and is set to 500/2750 on the GPU clocks, 2100 on Memory with fast timings on, and +15% power limit, voltage at 1110mV (default 1200). This works stable with no artifacts.

[@colten6996]

I went from a gtx 1080 to a 6700xt and haven’t noticed much of a difference at all. Would you recommend these settings to get most performance out of my gpu?

[@bradmorri]

Have a look at HWINFO64 sensor panel. That may show more info than the other software. I suspect that the interconnects between the GCD and the MCDs (Infinity Fabric) clocks are having their frequency ratio messed with when you manually push the core/memory clocks up. The core clock runs fast, the memory runs fast and the connections in between are being strangled and cut in half.

[@senseofpermanence]

just went fully team red, followed this exact easy set up and got more performance and my temps are rock solid.

[@thepuppetarmory6262]

I love you content I just built my FIRST pc your videos helped me out alot!

[@raszelast]

Congratulations! Hope you had a lot of fun learning and building it, now you get to play around on it.

[@matthewcooksey5411]

I was game stable for a while @ 3125 core 1125mv but it ran warm and I actually gained 1000 points in timespy by reducing the clocks and going to a 1100mv undervolt. It also runs max 75c junction temp so I'm happy with it.

[@L0rd_0f_War]

Sadly lack of research by Jay is what caused the first video 'errors'. The OC tuning for 7900XTX has been very well explained by Techpowerup in its OC page for XTX cards. Basically, for most cards, just dial in 1080mv, 0-15% pwr slider, mem OC work between 2650-2750, and leave ALL ELSE ALONE. DO NOT TOUCH THE MIN MAX GPU CLOCKS (till much later if there is any headroom left). Edit: Very good tip by Jay on Mem load. Should help with memory OCs as well.

[@newls1]

good old school informative jay video... more of these please and thank you for these 2 videos on this subject. I hope you find the cause of it!!

[@floppyseizure8615]

I like turtles 🐢

[@JesusIsKing21888]

Same

[@chaz-e]

Lol

[@_Turtle_420]

Well I like u :3

[@tpjmadrigal12]

I love lamp

[@SwiftSilence]

Me too❤

[@bigmanoncampus3861]

manual fan speeds used to be a thing but they decided to take it away, I had my fans tuned when I wanted random temp spikes to go away while mining

[@HartenDylan]

Might be a bit of a P.i.t.A. but I'm curious to see if there's any BIOS settings to accompany these findings. With FCLK in the bios that can be modified on the CPUs, maybe there's a more labour intensive process of setting up the BIOS before Adrenalin. Could very well be wrong, but I'd love to see a deep dive. Thanks Jay! Also, I've been noticing a huge amount of salt from the tech community as of late, particularly towards you and Linus. I hope you guys remember that there is a community that appreciates the time, effort, and resources you guys put in to informing us. The people that are angry with the opinion of creators they have no obligation to watch will always break my brain.

[@Pilzkun]

Just wanted to say I'm glad you're finding cool things you want to make videos about in the tech space still!

[@zechariahbarrett9540]

I really like this kind of video. Its very informative for newbies and people that have been watching you and overclocking for years even on Nvidia cards. I was able to get another 10fps on my 2070 from information gained here. Thank you Jay and team.

[@Ronke01]

My GTX980M in my laptop does exactly the same thing and gave me crazy stutters in Escape From Tarkov with otherwise high FPS. It will drop memory clocks when power limit kicks in, rather then drop core clocks. Once I noticed this, my game ran perfect when running the card on stock clocks and even better when only overclocking the memory.

[@NeverAware]

1:24 i'm already lost, least i got snacks

[@SyeedAli]

This phenomenon of overclocking one place lowering voltage elsewhere would be a "voltage droop" or "voltage sag". I doubt it's a feature, as you surmised. If you watch old videos..

[@jpteknoman]

drivers are another important thing with AMD GPUs. i havent updated the divers on my 6600XT in almost a year now. i got the latest drivers yesterday and i saw up to 50fps gains in some games.

[@gencoserpen1260]

Adrenalin driver 22.5.2 introduced DX11 optimizations back in June so you are probably seeing a performance uplift on DX11 games. Greatest uplift for me was on AC: Odyssey.

[@TheExcelon87]

That studio turned out great btw, I love it. Great job!

[@TheTekknician]

I would appreciate a video about how to properly OC an iGPU like a 4750G Pro, 5600/5700G and so on. It plays along with your RAM of course, but I unless I fiddle with the subtimings, gain in FPS is absolutely marginal. Running a 4750G Pro at 4.4GHz and the Vega 7 at 2.3GHz (1.2v) and my ram 4x8GB @ 3866-CL16 (stock 3600-CL14), I get a 2-3 FPS more. Something comprihensible would be awesome Jay!

[@larryschnellope80]

My guess is that in order to reach the minimum clock you set well over 3000mhz, it’s putting the majority of the load on one of the gpu chiplets.. infinity fabric used to talk between the chiplets rely heavily on memory.. so seeing the memory load % drop from ~30% to ~15% is because the secondary chiplet isn’t being utilized as much as the primary chiplet.. think of the chiplet cpus and CCX.. when you push the clocks too high it’s difficult for both chiplets to hit those clocks together.. it looks like it’s limiting the frequency of one of the chiplets to supplement the other chiplet with the power necessary to hit those clock speeds.. so yes, while you are hitting over 3000mhz, it cuts the performance of the gpu in half because that clock speed is only on one chiplet, not both.. software doesn’t detect/read the individual chiplets so its appearing that the gpu core as a whole is reaching those high frequencies when in reality it’s just one chiplets, cutting performance by almost half…. That’s my theory… I hope if I am right I get some credit 😅

[@rcavicchijr]

Junction temp was up at 85. I remember Steve from GN had shown less pressure on some of the bottom MCDs on the reference model. I also remember their being different hotspots on the different dies. Maybe one or two of the MCD's are over heating, or maybe it's the fabric like you said. Also, like you said the front end, and shader clock are different. Maybe the front end clock is trying to hit 3.2 ghz, or maybe going less than the 200mhz difference between the two on the minimum slider causes a conflict. Too Much going on. AMD will have to weigh in here. Either way, I guess we know where AMD hit trouble in meeting their intended clock speeds.

[@ogosh5364]

@JayzTwoCents to your point about the interconnect/fabric clock, remember how on Ryzen if you increased the RAM speed past a certain point (was either 3200 or 3600) then it would cut the fabric clock in half? I wonder if something like that is happening here? Talking out of my ass here to make the numbers work: Maybe if the GPU clock goes to high, it has to drop the fabric clock and the effective memory frequency gets cut in half as a result. The reference card may be reporting effective frequency while the Red Devil is still reporting the actually memory frequency. I’m prob wrong but makes the numbers work.

[@kylerogers9046]

So appreciate both of the videos you’ve put out on for this. I’ve never been into OCing much solely for my lack of knowledge- but listening to you talk about it, it sounds like AMD has “Apple-fied” it. Apple does it all the time to so many processes (especially in their “pro apps”) Something under the hood is making adjustments/smoothing or like you said Robin-hooding (rob from one to give to the other). I’m hoping it’s a driver/software fix from AMD. And I do so wish they would just give an expert panel that showed every single stat about the card. Even if not control, but full transparency on every clock/core/whatever just so you can see what’s going on. Thanks again for your info and time. Don’t let the keyboard warriors get you down 👍 you do great work

[@bitsbfg1810]

Based on the deep dive Steve did for the new GPU structure id assume the controller from either the infinity to memory to cpu is the issue and it could also explain the physical power limit too. Look back at the substrate that AMD used between the 6nm memory chips and the 4nm cpu chips in peticular.

[@aalanmau]

thats not a monitor, thats a MONITORRRRR

[@rylodinson6100]

Ryzen 5800X3D [PBO2-30] & Sapphire 7900 XTX nitro+ min: 500 max:2900 1025mv 2764 VRAM power +10% -- 3DMark Port Royale : 17732

[@spakentruth]

The thing is AMDs current hardware relies on voltage and temperature to determine how much performance headroom you have. Lowering the temps/reducing voltage curves tend to leanto higher sustained performance from my experience.

[@stefensmith9522]

I've had good luck overclocking amd cards by also changing the MIN frequency to 100-200mhz lower than the max frequency for manual OC. It seems to help maintain a higher average mhz during a 3d mark run.

[@iRocks-Mak3rde]

@jayztwocents My reddevil 6950xt oc a lot better if I set MIN and MAX gpu MHz EXACTLY 100MHZ from each other. Undervolting under 990 results in lower fps. Also going to watercooling helped, not only with the Temps BUT also it seems to have a bit more power available because no leds and no fans ✌️

[@Ben-Rogue]

I can't imagine 990 would be stable for anything other than short benchmark runs. I've never been able to get any RDNA or RDNA 2 chip below 1050MV(ish) without stability issues with extended testing (multiple different games tested with hours of gameplay)

[@krisiscove]

If you are under water why undervolt instead of maxing power with MPT?

[@iRocks-Mak3rde]

@@krisiscove I haven't done that jet because the card preforms nicely and is already taking 400w under load. The oc I could reach was super close to 3ghz.

[@krisiscove]

@@iRocks-Mak3rde if it isn't broke. No need to push the extra power if your card is running the way you want it to!

[@TotallySlapdash]

Personally I'm in the 'buy a better card and undervolt it' camp. That way 2 years down the line I can overclock it and feel like I got an upgrade, while I enjoy the quieter system and lower power consumption while it's new (and stretching the cards lifespan).

[@DailyCorvid]

In two years time GPTChatBot will be building your RAM in a backroom for you :)

[@radugrigoras]

Could be a chiplet thing. You are basically trying to OC 4 GPU dies at once. Remember in SLI you could only OC to the lowest of the two cards, this could be the same thing. You could be crashing 2 chiplets out of 4, your freq would look normal even if only 1 chiplet was able to hit it because you don’t have individual CCD visuals like on ryzen. Driver doesn’t crash because whole card didn’t crash and switching it to default re-initializes all the chiplets. Just an idea I could be completely wrong. On the red devil it might not drop ram because they chose not to do that but that still wouldn’t stop a whole chiplet crashing or multiple. I would imaging the GPU could still run, with 1 chiplet it was designed to be like that, 1 chiplet, 2 chiplets, 3,4,5,6 I mean that’s the whole point of the design. You could test it possibly with a FLIR on the back of the card to see if any of the chiplets shut off.

[@hellsgate09]

The GCD is still monolithic. The MCD's are what got carved off into chiplets, and there are 5/6 of them - depending on the model. The interconnect needed in a GPU for GCD chiplets does not exist yet, and is probably a few years away based on what AMD's RDNA engineer was saying in GN' video with him. Given that there seems to be a direct correlation, at this point anyway, between the memory controller load dropping and the drop in performance, whatever is happening is in some way related to the MCD's and/or how they connect/communicates with the GCD.

[@RyTrapp0]

That's not how RDNA 3 is designed - the chiplets are only memory cache dies, the compute functions are still on a single monolithic compute die

[@jts0221]

just messed with my 6900xt while following along and now I have a -100mv offset with +5% power limit and am pulling 2700Mhz instead of 2540Mhz with LOWER temps (67 degrees) Thanks Jay!

[@grazziotinpedro]

i had a similar issue with the rx 6900 xt when it launched and mining later drivers fixed that minimum slider is forcing the card to run at X speed even though it dosent have power for it and since it cant drop core clocks to be withing the power budget it has to drop memory which is the next thotling state after core try using Hwinfo and check if the SOC isnt the one dropping out, when mining thats what usualy causes the hashrate drops also if you check MPT for 6000 series there is a thotler for memory and SOC TDC so thats probably whats being triggered im waiting for igors lab MPT to buy 7000 series till then they come with too much limits i usualy run my 6900xt at 2300mhz and 900mv it dosent pull over 200w when playing lighter games

[@JCTiggs1]

Glad the Google search helped and you finally figured it out Jay.