We were WRONG about RAM - Or were we?

[@MetsariSika]

This format where many staff members are expressing their opinions was really refreshing and enjoyable, wouldn't mind to see more!

[@tw751]

Yeah, i can only agree.

[@JrockProject1]

Agreed. They all have different tech backgrounds and they are all probably right

[@ThKolle]

"Their Opinion" For me it does not make a difference who is reading the script.

[@SevenHunnid]

I’m 20, grinding hard to get out the hood, I can’t be 30 with nothing to show for so i smoke weed on my youtube channel 🤦‍♂️g

[@murilovs3827]

@@ThKolle exactly. Interesting to see some people fall for it.

[@88porpoise]

So, the testing seems to show: You need your RAM, CPU, and GPU fast enough to keep up with each other and any one falling behind drags the entire system down. In other words, balance your components for best results for the cost.

[@DrakoonLP]

They don’t need to be balanced, the weakest link just needs to be as fast as you want the system to be

[@88porpoise]

@@DrakoonLP Then you are wasting money and/or throwing out performance.

[@no_nameyouknow]

@@DrakoonLP no they don't need to be balanced but unless you're planning on upgrading the slowest part you don't want to go too far above that on your other parts or you're just wasting money. And if money is no issue then you just want to get the fastest parts you can for all of them.

[@DrakoonLP]

@@no_nameyouknow wasting money isn’t a good argument, or are you using your computer at 100% CPU and GPU usage 24/7? There will be always performance wasted, because it’s not constantly necessary, so what’s the difference if you have a weak link?

[@anousenic]

Not necessarily That would only be true if all components had identical costs (relative to their potential performance contribution across all used applications and across all metrics). And if you don't plan to ever upgrade any individual component. Depending on the costs of the parts you may get more performance for the same money when deliberately building an unbalanced/bottlenecked system. Especially if you plan on upgrading certain parts later (effectively shifting the bottleneck to another component). If you had a perfectly balanced system (which doesn't exist in reality, because you don't use only a single app while never installing updates for os/driver/app), then the only upgrade path would be throwing it away completely and buying a completely new system. Not exactly cheap either.

[@Volatar]

My personal experience: I had a 5800X and a 2070 Super and went from DDR4-2066 to DDR4-3600. That was the only part I changed. Overall framerates went up by about 2fps in most games, but my low fps points went up by a lot. In the AC:Od benchmark I went from a 15fps low spike to a 30 fps low spike with the RAM change. I felt it was worth it for that frame time consistency alone.

[@AarPlays]

That's actually really interesting and might push me to replace my ram. I get really bad 1% lows with my 12700 and 3080. While it could be because the task scheduler is still having issues with e cores I wouldn't be surprised if the ram was at fault too. I'm at 2666 (gen 1 ddr4 basically)

[@hmst5420]

@@AarPlays you definitely need to upgrade or overclock it

[@pirojfmifhghek566]

I really want to see them redo this test with some modern AMD cpus as well, because RAM speeds make a massive difference in CPU performance. The infinity fabric can only go as fast as the RAM sticks.

[@taiiat0]

​@@AarPlays yes, you should. the gains will be not amazing if you're just looking to use XMP (rather than Manual settings), but you can still get a worthwhile difference. while supply is going down as they're out of Production lately, a Kit such as... BL2K16G36C16U4B is a good value for newbies. any Kits higher performance than that will start to likely need you to set Voltages manually. if you can't find that Kit, the next cheapest one (that's good)on US PartPicker atm is F4-3600C16D-32GTZN. expect to need to manually increase System Agent Voltage a little bit though.

[@Kimbalhota]

@@pirojfmifhghek566 I got surprised they didn't mention Infinity Fabric and the clock parity in this video.

[@PunktBlah]

This is awesome. Acknowledging your own limitations and making them transparent while showing how your knowledge has improved is an awesome point. And it's far more credible than not standing behind your past tests and methodologies. Awesome video!

[@DoubleMonoLR]

Personally I'd be interested to see how much difference RAM speed makes on an IGP, as it's sharing the system memory. Considering they'd be more common in cheaper computers, the difference in cost of faster RAM would also be more relevamt.

[@reesesman8821]

Dawid does tech has a whole bunch of videos about it check it out it does matter

[@xBruceLee88x]

I think it was always obvious to not get the SLOWEST memory of a particular type, but 9 years ago it was more that you wouldn't always see a significant difference or rather feel a significant difference, in theory. Good to see you guys decided to revisit this for more recent systems.

[@DJSekuHusky]

I feel like the only reason my DDR3 system has been able to keep up with the DDR4 builds 'til now was thanks to the 2133MHz RAM kit. It allowed my 1st gen i7 Extreme to compete with up to 8th Gen i7 on equal computing output power, despite lacking modern instruction sets. Even then, I'm comparing 2133 against 3200 and 4000.The frame curves might as well have been a direct overlay with the data presented in this video.

[@rodryguezzz]

Exactly. Even CPUs 9 years ago were all similar to each other. i3s had 2 cores, ofc, but i5s and i7s were all quad-cores with a few hundred mhz clock difference. i7s had 8 threads, which were useless in most games. Most CPUs were enough for all games and performance was similar.

[@Szydelski]

@@DJSekuHusky I don't know. I am still using E5-1680v2@4.2 with 64GB of DDR3-1600. All combined with RTX3090 FE since two years ago. No problem with framerate in any game.

[@st0nedpenguin]

9 years ago it still made a huge difference to minimum framerates. LTT just dropped the ball on this one, memory has mattered for quite a while.

[@anonym3017]

it boils down to there being a "fast enough" speed. being under it hurts performance. going above it just wastes money.

[@punkcanuck69]

Games that have large RAM footprints are heavily impacted by RAM speed, Factorio, Stellaris, Dwarf Fortress. etc.

[@Intelwinsbigly]

Pretty sure dwarf fortress is fine on old ram. It was being made back in the early ddr3 days.

[@MikeDawson1]

it's the access pattern that matters

[@Milo19970]

And Minecraft😂

[@trapical]

Factorio is fine for the majority of players. It's the 1% with insane game-breaking factory sizes where RAM speeds all of a sudden are worth their weight in gold. So yes, it helps, but those players are the exception rather than the rule. Average, or new players, shouldn't spend extra on RAM just to play Factorio.

[@hilligans1]

@@trapical wont be long until they turn to the 1%

[@kaelwd]

I'd have liked to have seen this with Ryzen too, I've heard it has a more significant effect than with Intel because infinity fabric and ram clocks are linked or something, so faster ram also means faster cache and inter-core communication.

[@louisfriend9323]

hope they repeat this for Ryzen yes, especially with Nvidia offloading dispatch to the CPU where Radeon has hardware acceleration for that

[@TheSpiritof76]

That was true for the first gen Ryzen only IIRC, Intel actually uses ram better nowadays if I'm not mistaken.

[@Heymisterbadguy]

@New Moon why the per-ccx corr count matter in this case? Thats curious. I was sure they were going to at least mention how ryzen is affected by mem speeds usually, and the diffs between each gen, but they didnt :(

[@imadecoy.]

That's no longer true. That info is at least 3 years outdated at this point.

[@superneenjaa718]

@@TheSpiritof76 it's true for all Ryzen. New or old. Fabric clock or fclk is tied to ram speed so there's better inter core communication with faster ram.

[@jannikl.8542]

I absolutely love the structure of this video! Great concept, would love to see more of that, it really benefits from the fact that there are so many brilliant minds employed at LMG

[@Frosty2]

Anthony I believe has gotten progressively better at making videos and it’s so refreshing to see. It’s so awesome seeing this channel grow and it’s employees showing their true talent and passion for videos.

[@issaciams]

🥱

[@nelsonbutcher1]

AMD said this in a presentation a few years ago. Right now the issue isn't making factor CPU cores, it's feeding the ones we have to reduce wasted clocks waiting for data.

[@sagegeas5198]

And this is basically part of why they included in the instructions for cpu's like the FX series to use higher than 1600MT DDR3. 1866MT or higher to be exact, if I remember correctly from back when I had a FX8350.

[@Postman00]

Yeah, that's why the majority Zen 3's architecture improvements over Zen 2 are almost all in cache, and why the 5800X3D exists. I suspect it'll be the same kind of development for Zen 4, and how they can design the architecture around the faster DDR5 memory to work with cache to improve IPC.

[@sagegeas5198]

@@dekjet If by critical problems you mean the zero day bugs like specter and other stuff that only works because both Intel and AMD got sloppy; then we might have a different definition of 'critical problems'. Here's an example. My 3600XT does just as well using just its 6 cores without any 'hyper threading' as other 'better' cpu's with more cores but less cache. Why? Because Postman is right, and you... well. you aren't wrong, but you are definitely on the wrong track. My point here is that we wouldn't need those virtual cores in these cpu's if we just made the 'real' cores properly in the first place by supplying them with an overabundance of resources. THEN if a person really wants to use all the other fancy gizmos and tricks that also allow for all sorts of security exploits, they can do that; while the rest of us get a superior product that works fine without all the need for patching zero day bugs that wouldn't exist if we just didn't include those fancy tricks. You can say I am wrong all you want, but at the end of the day, my computer works better than others, and is more secure, because I have more cache per core than similar cpu's with same cache but more cores. All because I don't use the 'fancy tricks'. I hope that makes sense. I'm not feeling well today, but your comment irks me because it's wrongheaded. AMD and Intel's only job is make CPU's work well, and do it securely. Which is why I have a problem with your comment. If they were actually concerned about 'critical problems' like you say, then all those zero days wouldn't be a thing.

[@andreewert6576]

@@sagegeas5198 no special instructions there. A bit of tuning for the memory controllers, then some binning. But yes, fast-ish memory was needed to keep all the cores fed in some scenarios, just not gaming. Fast-ish becaise of course when 1866MT was finally supported, the market already offered 2133MT+ kits.

[@sagegeas5198]

@@andreewert6576 Actually... my FFXIV game didn't stop being a mess of black frames and stuttering until I increased my ram speed. I got a lot of flak over this from know it alls back then, and I still get some to this day when I bring it up. But back then FFXIV was basically pushing the limits of some systems, and the 8350 wasn't exactly as great as they sold it as, at least until properly equipped. DOOM was another one. P.S. I only knew to buy the 2133mhz ram cause of the instructions... so... I really don't know what universe you all come from, but this one I live in had instructions saying to use that ram...

[@Guiny]

the technicalities within these videos, especially the research done on the topics have improved soo much over the last few years! Incredible and impressive stuff from the LTT team!

[@chickenfried12]

Can't believe you watch ltt!

[@mrincrediblememesuper257]

2nd also didn't know you watched LTT

[@vill2980]

cant believe you still suck with better hardware /j

[@illogik]

@@chickenfried12 they literally have over 14 million followers

[@bringbackdislikes3195]

@@illogik wow literally? damn I literally didn't know that.

[@declangoncalves9373]

I dig the format of expressing and testing multiple team member's theories sequentially!

[@Kenjis9965]

There's always a sweet spot in ram pricing. Going for the absolute fastest ram available probably isn't worth it but it's probably a terrible idea to cheap out on your ram. Like I don't feel the 2666 was worth the money back in the day (iirc it was a considerable amount more) but 1600 and 2166 weren't too bad.

[@Derzull2468]

Overall that's true for every component. I doesn't make sense to go cheap on all but one component or best in slot for all but one. It's better to be consistent to minimize bottlenecks as much as possible.

[@joshuasomeones9618]

I agree with both of you. Just one thing to add. From my personal experience cheap RAM generally doesn't perform terribly as much as it just doesn't last, or you get duds more often. Ether way, same deal as both of you said already... Just as a general rule of thumb, once prices start jumping for smaller amounts of performance, you are paying for the "testing" of newer technology. I generally don't have the money for high end machines, but I still pump out some nice graphics for even some more intensive games. Just not getting that high frame rate or 4K res.

[@srpenguinbr]

The absolute cheapest usually isn't that much cheaper. I was surprised to seee how little price differs between DDR3 and DDR4, last time I checked

[@nostrum6410]

ddr4 4000 is pretty good value now

[@Someone-jf3mb]

When running games, there are a lot of information needed to be passed to the GPU from RAM, not just models or textures. Stuff like positions, rotations, etc... for models/objects each frame are needed to be sent to the GPU from RAM. In the past, the RAM is always running faster than the GPU even at the lowest speed (or GPU takes longer time to process than sending data from RAM to GPU) Meaning it is GPU bottlenecked. And now, because the GPU can run faster than the RAM and process things so quickly, it is RAM bottlenecked instead. The same thing actually applies to CPU but on a less scale because sending data from RAM to CPU is always faster than sending data from RAM to GPU. So therefore you can see CPU does have an affect but is far less than the GPU.

[@gamingmarcus]

Yeah that seems to be the primary factor. Fun fact: total RAM latency basically hasn't changed since DDR1. We just increased bandwidth at the cost of CAS latency. Going from DDR3 1333 to DDR4 3600 isn't even a 3x increase, but we've seen around a 10x increase in GPU power depending on where you draw the line.

[@Owl90]

Better explanation than the video.

[@oginer]

This is not true. There's something called vertex buffers for this exact reason. All model data resides in VRAM, not in system RAM. The same with textures. So there's not that much data sent every frame. A lot of computations in modern games are done directly by the GPU and all the data resides in VRAM (see how RAM requirements have barely changed in the last 10 years, while VRAM requirements have consistently increased). And in the case a lot of data needs to be transferred from RAM to VRAM, the PCI-e bus is the bottleneck, not the system RAM. Unless you're running a weird configuration *couch* RX 6500 XT * cough*, system RAM is much faster than the PCI-e link, so RAM to VRAM transfers are bottlenecked by that.

[@the9red9one]

Actually pretty reasonable thing to do using statistics in minitab or some other software. Not sure what has kept people from really modelling this

[@philistine3260]

Effect* Sorry, I couldn't help myself

[@lordofpolls]

lets be real; we can only be correct about tech for about a month

[@hellsoul0]

yeh almost any 9 year old computer tech video is going to be outdated and or flawed in todays term lol.

[@Wurldly]

@@hellsoul0 so now they can go back and redo videos. More content lol.

[@hellsoul0]

@@Wurldly infinite content loop bayybeeeee

[@1SOSORBIDE]

yes, you are all wrong..

[@urkent4463]

thank god thats not the case

[@Neoxon619]

I guess RAM speed is one of those things that became more significant as load times became more of a factor for newer games, especially with current generation consoles pushing that with their SSDs.

[@MrMOGHammer]

SSDs are still the bottleneck even NVME 2. Faster memory helps for already loaded stuff but not when it has to search on disk

[@12coco100]

I think data and moving of data is gonna be another method of judging power for gaming. The unreal 5 city demo proves why the current gen console run as good as they do

[@killertruth186]

@@MrMOGHammer Even faster than direct storage? Or is it the fact that both Nvidia and Microsoft had switched from old method of storage into direct storage?

[@Mr.Morden]

Speak of SSDs, this video should be revisited when games that use DirectStorage become available.

[@seldompopup7442]

@@killertruth186 direct storage is just DMA from Nvme don't go through CPU memory space but direct to GPU VRAM (correct me if I'm wrong). DDR 3 1600 gives ~12GB/s continues read/write. So basically 6 GB read to VRAM and 6 GB write from Nvme DMA. Typical consumer pcie gen4 Nvmes can get 5 GB/s read.

[@ebsolas]

Idea for Labs. On the website a “bottleneck calculator” would be helpful. Being able to slot in my CPU and GPU and figure out what ram I needed would be nice. Or for GPU or for CPU

[@methos1024]

Just look at some of the thousands of Benchmarkst an look at what you can get with which Ram speeds. You are most likely totaly fine with ~3200 to 3600 MHz DDR4 RAM, Buy good, but not excessively expensive RAM. If you would put that excess of spent money instead in a better Graphics Card you will gain more FPS. So Just use the usual Benchmarks of different RAM speeds to check, that you don't buy too slow Ram Modules. And also dont forget to give every Channel of your RAM Controller a seperate RAM Module (dont missinterpret Memory Bank with MemoryController Channel. The most Desktop Memorycontroller have 2 Channels) That is way more Important than the pure MHz. Only Using one Channel vs two is already half the Speed for throughput. Also Bottleneck Calculator seems silly :D and nooby kinda/missleading. You hav always One Bottleneck and if it is a CPU, GPU, RAM or Software Bottleneck (FPS Limit or VSync etc.) heavily depends on what you configure in your Game Settings AND which Games you use (or Software)

[@taiiat0]

there's so much more to DRAM than just Memory Frequency.

[@deivytrajan]

Bad idea. There was already a website to calculate bottleneck and it was horrible

[@amiltonfcjunior]

There's no way to precisely calculate a bottleneck. Hardware and software changes all the time. Given the application though you can "guesstimate" how good or how bad it's going to be.

[@theParticleGod]

Calculating bottleneck.... bleepedybloppidyblipblip.... Bottleneck located between keyboard and chair. Response time: 150ms Optical throughput: 8.75mbit Max speed: 19Km/h

[@t.c.b4722]

It's seems odd to leave out Ryzen here, especially since it's well known to be a lot more sensitive to RAM speed.

[@lucidbarrier]

They were trying to figure out what was wrong with their previous methodology. Their previous tests were done on an Intel platform, so switch to AMD technology would add so many different variables to the equation. They needed to replicate the previous tests with modern modular components: ie, swapping out the old GTX 660 with the 2080 to see if the tests were GPU bound. It wouldn't make sense to swap to a Ryzen CPU in the middle of a test. They would have to go back and replicate their tests with AMD hardware at the time, with FX processors (that were notoriously bad with memory) and then compare them with modern stuff. That would be an entirely new video.

[@PickeringSamuel]

Ryzen isn't as sensitive to ram speed as it was in the first generation.

[@LucidStrike]

He says at the end, "Radeon or rising may scale differently." Not every study covers every possibility.

[@simanova837]

@@PickeringSamuel its very sensitive, thats why ryzen 5800x3d is on market

[@johndicus123]

I just love the progression of this channel and its people in all these years. Way to go folks!

[@ZadesLegacy]

I'm honestly shocked they didn't talk about the architecture of infinity fabric. Or how Integrated GPUs (Like AMDs APUs since like... forever) are far... FAR more dependant on RAM speeds.

[@Th3Rom3]

Seems weird to me, too. I am no expert but it seems like a big oversight.

[@ryue5626]

They spend too much time kissing their sponsers ass and promoting their products

[@dqpowered]

That made me expect "Sponsored by Intel" at the end, but yeah....

[@Keptains]

True and i agree, but im sure they know about it as Anthony said at 8:38 and somewhere else aswell that ryzen or radeon can be a different story aswell. Maybe they felt it would be to much compressed content or they probably want to dedicate a own video regarding this topic

[@Zuriki09]

Infinity Fabric won't have much impact on with regards to system memory. Integrated GPUs will benefit massively from higher speed memory, but that isn't a useful benchmark as the system memory is substituting dedicated video memory, so really you're only benchmarking a form of GPU bottleneck that tells you nothing about the role system memory plays in a typical gaming desktop configuration.

[@MikeKrasnenkov]

There are vastly different from typical AAA game genres with different RAM performance scaling. At one time you used to have Factorio benchmark, which bottlenecks on RAM and CPU speed on large bases.

[@zackarysemancik5491]

A fellow Factorio brah! Go Factorio (but stay there)

[@nathanlowery1141]

I’ve been curious about factorio. Haven’t bit the proverbial bullet on it yet.

[@zackarysemancik5491]

@@nathanlowery1141 free demo on steam. I recommend trying that out if you're on the fence. iirc, the demo is just the tutorial for the game, so it'll help you learn some mechanics too!

[@Splatahking]

@@nathanlowery1141 Don't do it. You'll suddenly gain consciousness a few weeks later with a massive headache and 250 hours into that game. Unless?

[@DarkNexarius]

@@Splatahking Unless your PC doesn't have enough performance.

[@am53n8]

I think we should see videos like this more often, challenging and sanity checking our old beliefs to see if they still hold true. Tech moves fast, and a lot of things can change

[@shreyaskul]

I have a simple analogy for this: Imagine the RAM being as a fuel to a car, the software being the driver of the car and the hardware being the engine and the car itself. Now, if you put high grade fuel (high speed RAM) into a honda civic (Low grade hardware) you're not going to see any significant difference at all. Now if you put high grade fuel into a racecar but the driver is a amateur (i.e. unoptimized software) again you won't see any significant difference as the driver wouldn't be able to achieve full potential of the car+fuel. If you achieve perfect combo, racing car, best fuel and a trained driver it will greatly amplify any improvements which exist.

[@lordspalse0062]

Thanks Prof

[@TaylorPassofaro]

This video was really great, love everyone chiming in on stuff, and Anthony is always just the absolute boss on presenting complex stuff. I'd love to see more dives into topics like this, I'm slowly getting a computers 101 from this channel and these videos are the heart of it.

[@TheDude50447]

DDR5 relativises this to a degree because its basically dual ranked by design. A few other tweaks also allow for considerably higher data throughput even at the same clocks to DDR4.

[@Donnerwamp]

My personal experience is that it's really depending on the game and the CPU. Ryzens love that Infinity Fabric Clock x 2 RAM speeds with as tight a timings as possible and if the games are already CPU bound, the faster RAM will help to squeeze a few more FPS out. And well, as you've seen yourself, GPU bottlenecks are the best friend of slow RAM. And slow CPUs.

[@tomaszszupryczynski5453]

well ryzen isnt connected to ramspeed but ratio. cpu must synchronize with other devices, multi memory channels will only work with multitasking, its always changing but you cant access memory close to each other or in same cache line, that will cause stall. strange that nvidia dumps recently resizable bar, with amiga and pc, playing with planes or swithing 64kb range even on 256kb card was nightmare even in 32bit, with 16bit that was understandable

[@ethans4783]

I had the same thought as Alex, the IPC limits of the CPU was the bottleneck, where the system memory couldn't stretch it's legs and reap the benefit of the faster bandwidth

[@annieworroll4373]

Love to see that you look back on old content and give out appropriate updates.

[@gothmaugen]

an outstanding multi-faceted assessment! as someone who has gone through dozens of gaming rigs from 90's to present day, playing every genre i really loved this!

[@nikoszatcarnyi4028]

However, its the subtimings that really, REALLY matter if changed correctly.

[@reallunacy]

I'll be honest, I assumed it was the processor that changed how important ram speed was after all the media talking about ram speed affected the infinity mesh on the first Ryzen.

[@saccaed]

I've found timings to have about as much an effect on system performance as speed. Where a huge benefit can be had is for systems that run integrated graphics that share main memory. Outside of render or compile operations, memory speed typically only effect overall system performance by single digit percentages. Far more important that the CPU and GPU have sufficient onboard memory than for a system to have fast memory to make up for lacking memory on the CPU or GPU.

[@gazj]

im running 3866 at cl16 its trade offs against the same ram clocked at 3600 cl14 on my amd 5600x system

[@st0nedpenguin]

Memory speed, and more importantly latency, has had large effects on gaming for years.

[@CotyRiddle]

I have found that having ram over clocked (speed only) don't have that much of a affect on integrated graphics (intel side AMD might be different) but having tigher timings at the same speed gave a significant bump in minimal FPS.

[@saccaed]

@@st0nedpenguin Yes, however main memory does not have nearly as much effect as cpu cache memory or GPU onboard memory when gaming is the concern. When hardware specific memory is not enough and main memory is used as an extension, that is the best case for main memory to make big performance differences. Latency and bandwidth dominate, speed is mostly just a simplified metric as it's derivative of latency. The same dominates all memory. It's the reason why spending a bit for a GPU that has better performing memory and or increased capacity is going to be a better performance investment than spending the difference on slightly better performing or higher capacity main memory(provided that sufficient main memory is already present). Same goes for CPU cache capacity and performance(though performance really isn't something that can be shopped for). A large CPU cache by itself often will yield more performance than can be found upgrading from middle to performance main memory.

[@saccaed]

@@CotyRiddle I've had similar results. Unless the base speed is outrageously low, timing optimization always has yielded more performance than speed adjustment. Typically a good amount from tuning to the longest refresh the system can reliably tolerate, then a bit more from adjusting the access timings.

[@iwantmypot]

Anyone else remember when you tuned your memory to properly sync with the clock speed of your cpu? Even if your ram could run faster, you tuned it to a multiple of the clockspeed of the CPU (usually either 1x or 2x), so every time the CPU made a call to the RAM, it didn't cause any hiccups because the RAM was mid cycle.

[@dezpotizmOFheaven]

You mean if a CPU does 3.6 GHz my RAM should be a 3600 MHz (so basically also 3.6 GHz) in the best case?

[@dmitripopov8570]

It is great that more team members are participating in video hosting! Great job, fellows!

[@Savitarax]

PLEASE PLEASE do more content on ram, so many games benefit from ram timings and ram speed. ESPECIALLY COMPETITIVE GAMES. Games are not as GPU bound as people think, they are CPU bound because your ram is so god damn slow. Lots of my videos I publish discuss this topic and I like to show it In gameplay.

[@skilletborne]

But the video just proved that's not the case. Having very poor RAM impacted performance, but anything mid-range or higher wasn't making any real difference. Even on highest end CPUs, it's pretty obvious that the current standard of 16gb ddr4 3200 CL16 is still overkill even in CPU bound competitive titles.

[@Savitarax]

@@skilletborne because 99% of all tech reviewers touch just the primary timings. Which don’t do anything. It’s the secondary and tertiary timings that do a whole load of helpful stuff for cpu games. It’s really frustrating for reviewers to adjust the frequency and nothing else. Because it’s just one SMALL Piece of the puzzle. 2133 is not “high” nor is 3200 These are becoming bog standard achievable on every piece of ram in the last 4-5 years. And to make matters worse. When you increase the frequency the bios loosens the timings even more. So that’s why 3200+ seems “useless” it’s cause your latency skyrockets. You may have 3200 cl 14 or 12 or low primaries. But they are completely mitigated. By the huge latency.

[@jesusdesanto432]

I would like to see this test done with Ryzen. It's my understanding that ram speed matters a lot with second and third gen Ryzen in particular.

[@naufalap]

also ram latency please, people say cl16 is twice as fast as cl18 but I haven't seen anyone testing it maybe because I haven't looked at it

[@timmy7201]

@@naufalap I know the struggle... I picked 2x32 GB 3600Mhz CL18 for my Ryzen 5900X. Took me about two weeks of browsing forums to decide what memory I would pick... Higher clock-speed or better CL timing? I'm stil unsure if it was the correct choice. That said, the rig has been performing very good for the past 3 months ...

[@synatis4950]

@@naufalap isn't it the other way around? the lower the cl the faster it is?

[@naufalap]

@@synatis4950 oh yeah I forgor, edited

[@matthewday7565]

For Ryzen, it's the Infinity fabric speed... which equals RAM speed until you go too high and there's a divider

[@JoeBlow-ub1us]

This makes me even more interested to see the next gen gpus with next gen hardware.

[@qm3ster]

I loved how the imprecisely recorded specs of the old rig were put in sarcastic quotes. Best part of the video for me.

[@colinrobinson9858]

I’d love to see benchmarks using large Minecraft mod packs- it’s a great way to test frame rates using ram as their mod packs are crazy ram heavy

[@mglsj]

I think an AMD bench should have been used because Ryzen is known to scale its performance based on RAM speeds

[@jubuttib]

Yeah, and arguably at least is fairly severely held back by slow RAM, rather than gaining a whole lot with super fast RAM (at least Zen 1-2+). IIRC the last time they tested it, 3200 MT/s was only slightly slower than 4000 MT/s, but 2666 MT/s was significantly slower than 3200 MT/s.

[@Thelango99]

Yes Ryzen also bases its infinity fabric interconnect speed on RAM speed.

[@ABaumstumpf]

"because Ryzen is known to scale its performance based on RAM speeds" And that has been done to death and yield no different results.

[@creaturedanaaaaa]

Ryzen does not scale with ram speeds, it scales with FCLK clock which is tied to ram speeds.

[@game_now_4927]

The funny thing is that Intel still scales better than AMD with RAM speed. But beacause there is such a nice explanation for AMD people seem to think it scales better than Intel.

[@TaPaKaH1]

Single-rank memory vs dual-rank memory affects performance no less than RAM clockspeed and timings do.

[@markoulas2]

yes, because dual channel memory setup has double the bandwidth of a single one. With bandwidth = speed, I remember the good old times where triple channel setup was possible on x58 motherboard

[@7beats]

@@markoulas2 He's talking about ranks, not channels.

[@albertlevins9191]

Memory ranks make a LARGE difference if you need ram speed. I just had to buy new memory because my dual rank memory was too slow to be useable in my setup. I had to get single rank memory and overclock it to finally get fast enough transfer speeds to feed my GPU.

[@ruediix]

The interesting part is the difference between high end and top-of-the-line memory. This difference is very minor, at best, and it's at the top, while beyond what actually matters. It is meanwhile, more than worth going from basic to standard gaming grade memory. This costs only slightly more and is not only a much bigger difference but in the range that it actually makes a difference.

[@user-fd7tl4xg6y]

Precisely!

[@taiiat0]

if by 'top of the line' you mean like those 4500+ DDR4 Kits and 6600+ DDR5 Kits, true, those aren't worth the Price. but certain expensive-ish Memory is a great value for those that are going to do more than out of the Box. not so much if one isn't though.

[@ruediix]

I am talking the kits that are $400 vs. the kits that are $100-200 for the same capacity. Patriot doesn't even make that price range, only Corsair and G.Skill do. They typically have one or two better timing cycle better at normal speed and are capable of the maximum OC speeds of current chipsets, and typically run over-voltage by default.

[@Chloiber]

Yes, agreed - I think the video or conclusion is not really correct or they make the issue larger than it really is. Nothing much has actually changed. Of course, if you go down to 1600MHz now, it will be way too slow. Isn't that obvious? But still, comparing "regular", good value RAM with top-of-the-line RAM still means: RAM speed barely matters.

[@taiiat0]

@@ruediix for DDR4, yeah. you can already get the best quality stuff at ~$200-250, up close to $400 (for 32GByte) doesn't really change anything.

[@mamaharumi]

very cool, there are actually quite a few older LTT video's that need updates/revisions like this

[@mooncrow2447]

That last bit of “RTX 1080 it” was golden. Guess I was playing on a rtx this whole time.

[@proesterchen]

I seem to remember Digital Foundry showing RAM speeds mattering in some games in the Haswell era.

[@combosx]

Haswell era??? Mine is still going with a 4670 🙂

[@CanIHasThisName]

Yeah, the problem was that back then almost nobody would get anything faster than 1600MHz

[@federicocatelli8785]

@@CanIHasThisName Past 1866 gains were minor

[@realzakariax]

Now one other thing I'd like to see a video about (mainly because I don't have a pc yet, just laptop for now) is how ram impacts the performance of a laptop with INTEGRATED gpu's I'm looking to get more ram and I want to see if I can cut corners on bandwidth or if it will make a huge difference getting the max bandwidth and double the size

[@saccaed]

Main memory greatly effects integrated graphics(on cpu). Not to be confused with a discrete graphics setup though(dedicated graphics). If you plan to primarily use a laptops integrated, on cpu, graphics, memory speed and timings can effect performance quite a bit.

[@CanIHasThisName]

If you're using integrated graphics with shared memory, there's only one rule: You want the fastest RAM with the lowest timings you can reasonably get. If you can't increase the frequency of your RAM, at least spend the time on tightening the timings. Three problems: 1 - It can literally take a week of work to find the lowest stable timings 2 - Most people would never bother even if it was easier 3 - A whole boatload of laptops don't allow you to tweak the memory in any way So if you can do it and want to do it, definitely go for it, there isn't a scenario where the iGPU wouldn't benefit from it in gaming.

[@Nachokinz]

From what i've seen in my personal ram purchases; budgeting for higher capacity than one thinks they'll need (eg 64gb instead of 32gb) and a couple speed tiers below the top end has brought the best value and a build that ages more gracefully over time.

[@deivytrajan]

amm, for gaming this is the worst recommendation ever. 16 gb ram still running all games no problem and going 32 gb gives no fps boost

[@Nachokinz]

​@@deivytrajan As someone who has run builds with that mindset, i've run into a constant question. Do I see myself running the same software and games in a few years? With the constant increasing demand of software as a whole, more ram means a smoother experience outside of games and you don't have to be as mindful of your ram usage. Everyones use case differs, however I always encourage one to build with some headroom in mind to have "a system that ages more gracefully over time." Calling my recommendation the "worst recommendation ever" is unproductive and unhelpful for those genuinely attempting a build themselves.

[@kermitthegaminghermit8736]

That is very noble of you guys going back and reviewing your video providing further validation. Keep up the great work!

[@danbiohack5353]

I REALLY like how you all got to throw your hats in the ring on this one. Kind of a genius format. Keep crushing it

[@stefanosstamatiadis740]

I miss Anthony 😢😢😢

[@LindonSlaght]

He's emily now and you're a bad person if you disagree 🤡🤡

[@stefanosstamatiadis740]

I didn't know that!

[@LindonSlaght]

@@stefanosstamatiadis740 I wish I didn't lol. Same guy different do.

[@finderrio]

​@@LindonSlaghtwhy do you even care

[@LindonSlaght]

@@finderrio *posts a 10 minute video pleading for attent*on* omg bro why do you even care?

[@stevenclark2188]

My theory on another factor: Ryzen arrived and AMD systems benefit from increased bandwidth more than Intel, even when they aren't APUs.

[@taiiat0]

not the Bandwidth so much as the Fabric Clock since it's tied to the Memory Clock.

[@CryptoJordanVR]

One area where I would have loved to see you guys test is a scenario when you've gone over your VRAM limit and your GPU has to feed off system memory. In such a scenario, faster RAM "could" be ASTRONOMICALLY helpful! An answer to this kind of question could be very helpful to people who are on a tight budget and can't afford a new graphics card but faster RAM is within their reach.

[@h1tzzYT]

the issue with that situation is that once you truly run out of vram, no matter the main system memory speed you will see massive performance degradation including massive stutters, now why it wouldn't matter? Because when gpu can access data locally aka from its vram, it will be multiple times faster, so no matter main ram speed it will be unplayable experience regardless.

[@danl7756]

"I can tell you nothing changes about these segways to our sponsor..." that was ironic, relevant, and and so damn spot-on. You win sir. Thumbs up before I even watched the rest of the video.

[@SteelRatVT]

0:21 thumbs up for baby face Linus!

[@FatNorthernBigot]

I want some of that invisible RAM!

[@THEMithrandir09]

I wanted to shout Ryzen at the screen but learning that it made a (albeit small) difference with Intel was interesting in and of itself. But infinityfabric was the point RAM speeds really mattered if you had one.

[@theoldpcgamer77]

Low end Alder lake with ddr4 samsung b die OC'd on both speed and timings make a huge diff. I swear these main stream channels have zero clue about what they are doing.

[@st0nedpenguin]

@@theoldpcgamer77 Faster memory and tighter timings made noticeable differences even on Haswell. It's taking SO LONG for tech channels to catch up.

[@MetaDrow]

I'm kinda with Alex's theory that we a lot more faster cores to feed data although my theory is that faster RAM always did matter but to little to no benefit or was not worth it at the time. With the introduction of the infinity fabric tied to RAM speed from Ryzen, I think that's where RAM speed started to matter more and software followed to utilize it more.

[@saccaed]

Clock speeds within a generation of memory often matters less than the access timings when the clocks are similar. Increases in average performance often are going to be minimal, but increases in reliable performance can be huge(think minimum frame rate). Access timings are a multiple of base clock and different timings are going to react differently to base clock adjustments. When base clock is increased to just before the point of system instability, individual timings often can be adjusted more aggressively still. Alternately, memory timings can be adjusted without touching the base clock and can reliably beat out a simple memory clock boost. The process is just very involved and only really can be simplified a bit when some spread sheets are employed. Actually plotting timing reliability in nano seconds, then converting back to clock multiples is a way to automate some of the process of dialing in on peak memory overclock.

[@phattjohnson]

Best and most easily digestible content for PC builders / enthusiasts on the web currently! Top-tier production and information here gents!! I've not built a PC in 4 years but I feel I know exactly what I'm walking into for my next one now!

[@colortreo]

This was a really cool video! Thanks for the different staff and comparison of old results and what is happening now. Good job

[@MarcheRutolski]

Great video, glad to see the last few months have had some content changes that are more deserving of your audience. I think giving Jake, Alex, and Anthony more screen time also helps strengthen the brand, because while Linus is great for tech tips, sometimes the other guys are willing to explain details in a way that enhances Linus's perfect introductions and basics on a topic.

[@WarriorsPhoto]

Very interesting findings. I am not surprised since modern systems do have a lot more resources. We should expect 32GB to be the standard eventually. Just like 4, 8, and 16 were before.

[@carlosquinteronoda]

Would it be possible that the changes of OS have an impact too? I mean, could you make the comparison of one of the tests but when using Windows 7 versus 10? I guess it is worth to check if the most modern OS makes a better use of the communication between CPU, RAM, and GPU. Thanks for the great video.

[@Jules.11011]

also a good point yes

[@matthewcooper3774]

The format with different cast members was awesome. Keep up the great work LTT team!

[@alistairblaire6001]

I'm convinced the typo on the iFixit Minnow package was intentional to make us rewind and watch again :)

[@aresonair8186]

This test needs to be redone with Ryzen CPUs, everyone knows that Ryzen performance scales DRASTICALLY with RAM speed increase.

[@tyjuji]

They made that video long ago, there's even a screenshot of it in this video.

[@kalapahar]

I would like to see the similar performance benchmark done on Ryzen setup.

[@915Boss]

I think it'll have similar results but could vary a bit..The theory doesn't change..

[@monke2361]

Not going to be that different

[@AdamKafei]

Might it be worth doing an in depth analysis with a variety of of CPU/GPU combinations? Also with platter vs SSD vs NVMe might be interesting.

[@TheMikeyb86]

Anthony, the camera loves you and so does the mic. His voice is so easy to listen to.

[@britishmick]

I wanted to build my own rig as I normally do, but bought a pre built as it was close on price. When I got it home I realised where they “cheaped out”… psu and ram. I was getting substandard fps on my 3070 as well as occasional hitching, I questioned my ram speed to my knowledgeable brother who said “it’s not your ram”, I checked all my configs and eventually could do nothing else but to change the ram dimms from 2400 to 2933. It was drastically better?? I can post my set up for you guys to test as I do not think you will see a more clear example of faster ram benefitting a system.

[@pushingboundariesyt]

Been waiting for this! Ram speed is definitely a contributing factor to game performance nowadays

[@GonnerMeLeggies]

I think a lot of this is not only the gpu but on the CPU and the ram set you have. My system runs much better at 3600 with slightly lower timing than the rams bog standard "overclock" of 3200

[@alanhonlunli]

Fallout 4 was the first time I realized RAM speed mattered. The number of downvotes I received cuz of Linus...

[@starboy5177]

I remember seeing that video to build my rig back then. Dude has been around since forever!

[@svegetax]

One thing overlooked is memory controller placement and later on fabric clock ratio's. On an older ROG AM3 the memory controllers and pcie lanes were still offloaded onto the co-processor/NB(North Bridge). Now the AM3_x6-1100t did have a memory controller on chip. But was still bound by the NB. And without a fabric clock ratio to bind you. 1600 ddr was the common OC on a 1333 ddr3 system. Also when adding 2gig per core was met. Performance took another small jump. With how this episode ended. You're on your way to explaining how this still holds true today. Culminating everything discussed with Linus's observation. You can see why SLi died out. Aside from sync issues the limited bandwidth and pcie-lane limits we hit now removes it's viability on anything other than Server/Workstation/Farm/Threadripper. I'm guessing this series will be heading into the ability to push past fabric clock seeking the new sweet spot between ram speed and amount. Looking forward to the findings and Vids.👍

[@Catsincages]

That's okay kiddies, we have GN for actual tech advice.😼

[@TheGoombles]

always wondered if I screwed myself by buying a kit of ddr4-3200 vs something faster

[@HotdogSosage]

Nah its golden. 3200 is about the sweet spot where extra speed is diminishing returns. You'd be paying a fair bit more for an extra 1 fps in many titles, if you went for more

[@francoisgagnonlemieux3135]

I found out that right now after 4000 you are throwing money out the window. 3200 is nice because it can still easily achieve 3600.

[@FatalDreidel]

I mean for the longest time all you could find was 3200 at a decent latency back when everyone claimed latency was more important. Took a while before we saw some faster speeds with same latency available at the big box stores.

[@ezecskornfan]

@@HotdogSosage right, at least in my case where I use mid high ryzens.

[@sterpumihai]

I can tell you my own experience but you won't like it. 3200 mhz cl17 ram, 2080 super, 12700k, everything was smooth and gpu bound. This year I decided to upgrade to a 4090 and man.. as soon as you go up with the details stuttering begins. The more complex the game scene is the more you observe this. Upgrading to 4800 Mhz cl17 soon to fix this problem. Don't listen to Internet people saying 3200 is enough - for a 4090 it's not!

[@PokeWaffles]

I watch every Anthony video the moment it comes out

[@NerdClick]

In Call of Duty games, Modern Warfare 2019 and after, especially on AMD you get a big boost with high MHz / low CL.

[@DEMENTO01]

tbh imo the main thing of why ram speed started being important is ryzen, like yeah as we can see on these numbers on monolithic cpus ram speed can matter by a lot but ryzen? especially the first gen back in 2017? we went from "matters a bit" to "you need 2900mhz+ ram or else youre wasting ur cpu" real quick

[@RK-ej1to]

Would have liked to see an overkill speed like 3600

[@photlam9769]

I have 32gb of it, can confirm it is not all that it is cracked out to be

[@dezpotizmOFheaven]

3600 and overkill? You mean 6000 and above...

[@curie1420]

@@dezpotizmOFheaven you mean 69000ghz and 120000mhz

[@donkeyy8331]

As a computer scientist I don't think there was ever a time when memory speeds weren't important, it's actually one of the worst problems for engeniers to solve nowadays, ram speeds are way too slow.

[@thelifeemery]

Watching this video at 1:30 and realized I DO have a 660 ti in my closet. I have something Linus doesn't. This will be the first and last time.

[@hardlytech9072]

I'm glad other channels are finally starting to acknowledge that RAM matters. How much it matters depends on your task load, of course, but it matters! This was a pretty cool format, I hope to see more like this. Great work, team!

[@LBXZero]

If you want another theory, I have a suggestion for a benchmark, multitasking. What else is your PC doing in the background? The old games were single threaded. The main loop needed to have time to poll for changes, process the game mechanics, update graphics and sound, and repeat. How long after CPUs became 2-core and 4-core did games switch to multithreaded engines? Think about what gamers are running on their PC. A simple benchmark will concentrate on the user only running the game. When I play online with my local friends, we use Discord to live stream our screens to each other. Of course, live streaming your screen will add additional load onto the PC, but it is a realistic load. I can have firefox open showing data about the game, discord in a voice channel live streaming my game while displaying 2 or 3 of my friends' screens, and playing my game at the same time. Having all of this running at the same time should really choke the system in a realistic way.

[@mandavaler]

Well even if it had a lower effect than i thought at least at 6,000mhz ill never have to worry about a memory bottleneck lol

[@francoisgagnonlemieux3135]

For now

[@jamesburke2759]

I remember going from 4gb 667 to 4gb 1066 DDR2 and there was a massive response time difference.

[@thegenxgamerguy6562]

Actually (I'm a senior software architect) the reason is: RAM and CPU speed are diverging more and more, also we need much more bandwidth between L3 cache and RAM because we have more cores putting load on the bus. In (especially "managed", like C# or Java) software engineering we see the effects very clearly when discussing the impact of blocking garbage collections, which are heavily memory bound operations.

[@csmith9684]

Great info So speed is one thing but now (gaming) we are seeing ram requirements / recommended over 16g /32g !

[@creationsmaxo]

As an indie game developer specialized in gameplay mechanisms & game optimization, I can add my bits of info. One of the things that makes the RAM important in more modern games is the caching of reusable data. What is caching about? Well, it's about storing data that is still being in use over a period of time until it's not used anymore and, then, gets removed from temporary storage by the cleaning process which usually goes by the name of "the Garbage Collector". For the most part, the GPU handle 98% of the rendering-related data including its caching, hence we're far ahead from the time when the GPU was only about rendering and the RAM was about storing the data loaded. Especially since game engine are nowadays made to handle the GPU-oriented caching automatically for the most part unless programmed to do otherwise. The 3 biggest RAM sensitive usages you'll find nowadays are for the following data: - Physics stored data such as collisions, positions, orientation, etc. in a 3D or 2D scene. - Streamed data such as audio data and, in some cases, textures data not in use, but required on hand. - Online-related stored data such as player data, server access and on-hold data and other things. In the cases of some of the consoles, you might also have to consider that the RAM and VRAM is basically the same thing as the custom GPU (with eRAM dies) and CPU might be using the same RAM chips to store cached data. Just to point it out. One example of a game where more RAM might change the experience would be games with generated worlds like No Man Sky. All the modification done in the environment around the players is stored in the RAM, during game time, and momentarily stored on the storage via the periodic saves. So, the more a game requires to store temporary (but highly important) data, the more RAM it might be using. For example, in No Man Sky, when you have lower RAM, the game will reset changes done on planet faster as the game is made to make use of a % of the available RAM and, as the limit is reached, the game automatically determines the data by time stamps and distance to be erase to store the new data. As such, the game is not only relevant to the CPU and RAM capacity, but also speed (as data has to be processed, deleted (destroyed) and new data has to be stored anew constantly.)

[@albuslee4831]

Anthony's acting level have reached pro. Emotional expressions is fluent and natural, better than an acting school freshmen.

[@billy65bob]

Don't forget about the Ryzen quirk. Not sure if things changes with the revisions or the follow up generations, but it required the memory to be clocked at a similar frequency to its infinity cache for good performance.

[@JGComments]

You’re ALWAYS bound by SOMETHING, based on your budget or available tech. That means that practically, there will be a lot of other components for which additional performance won’t matter.

[@ChristopherHallett]

I'm still on a 3770K and 16 gigs of DDR3 1600 from 2012, and yet I can run Far Cry New Dawn buttery smooth at 1080p with my GTX980Ti. Same with RE2 Remake. I upgraded from my original 2012 GTX680 in 2017/2018 after I bought DOOM because it simply couldn't run it above about 5 FPS. Thankfully a friend was getting into VR and they had just upgraded to a 1080Ti from the aforementioned 980Ti and sold it to me cheap.

[@grone1778]

Man, Anthony’s camera presence has grown like an evolved Pokemon. Guy is downright charming

[@zinxderobo]

Anthony episodes are the only ones I'm not immediately sceptical of.

[@KimoKimochii]

where did he go?

[@NotTheCIA1961]

The format of this video was really entertaining, keep it up and well done

[@nikolafilipovic5506]

Thanks for this video! Funny enough you guys looked at my current system build back in 2013 and im looking at upgrading really soon. So the timing of this video is amazing.

[@danielv6906]

Wow, if only mainstream media had the same ability to question itself and do an honest investigation. Kudos to the LTT team!

[@STUBZx]

Currently having issues with my ram and this video popped up while troubleshooting and man.. I miss Emily. COME BACK TO THE CAMERA PLEASE! When you're ready, of course, but damn you're missed.