Would 3D V-Cache Help Intel CPUs? 14th-gen Cores vs. Cache

[@falcdragon]

In general these are pretty minor differences in cache, 33/36 is ~9% difference. The 24 vs 36 is a more substantial shift at 50% but still minor compared to 300% increase that X3D brings. If we go back to your original 5800 vs 5800X3D video there was an average 15% increase in performance for 300% increase in L3 Cache. If we assumed a fixed ratio 15/300 that 0.05% increase in performance per 1% increase in cache. At that level going from 33 to 36 is 9% more cache and assuming the linear relationship you could expect a 0.45% increase in performance. If we go from 24 to 36 that's 50% increase in cache, so again assuming linear relationship you might expect to see a 2.5% increase in performance. For this sort of comparison you should be looking at linear regressions between increased cache (either in MB or %) and observed performance change. Also given the predicted performance change it's likely within the run to run variability of your tests, which means you may struggle to be able to detect it. @Hardwareunboxed Not saying you are wrong in your conclusion (clock matters more for intel) but your evidence is insufficient to clearly show it, as it's not possible to test a comparable scenario.

[@falcdragon]

In fact for a number of the games you are seeing a larger than expected performance increase for the additional cache. Which would suggest cache is important and the effect is somewhat non-linear.

[@GewelReal]

Don't look at the % difference in size. Look at the total size, as instruction size is measured in MB, not %. Going from 3MB to 6MB would net the same increase as going from 33MB to 36MB, as you'd fit extra 3MB worth of instructions, despite one being 100% increase and other being 9%

[@falcdragon]

@@GewelReal It doesn't make much difference if you use per MB or percentage. It also doesn't really matter if the L3 is mainly cache or data. In both cases, increasing the amount of cache reduces the number of times you have to go out to the main memory and thus saves cycles of the cores waiting for instructions/data. If you calculate it out as a percentage performance improvement per MB of additional cache based on the 5800X3D and 5800 (and note that this doesn't account for the non-X3D's higher clock speeds), then what you get is 15/64 = 0.234%/MB. If you assumed the same benefit applies to Intel CPUs (which is reasonable the intel memory subsystem isn't massively better for latency or bandwidth than the AMD one) then you get the following predicted changes: 24MB > 33MB = +9MB, 9*0.234 = 2.1% estimated improvement in performance. 24MB > 36MB = +12MB, 12*0.234 = 2.8% improvement in performance. 33MB > 36MB = +3MB, 3*0.234 = 0.7% improvement in performance. Thus, in the tests the guys ran above, assuming linear scaling (which is an assumption), we'd expect to see 0.7-2.8% better performance, which certainly appears to be the case in a number of those games. In some cases, the difference is higher than that. The real question, though, is can the guy's test detect a 2-3% increase in performance, or is that within the margin of error from run-to-run variance?

[@jackleville546]

Fully agree with what you say. The increase in cache may not be enough for really cache-sensitive applications to store the data they directly need. And I would rather have liked to see the focus on games that really seem to benefit form more cache: Assetto Corsa, MS Flight Simulator, Borderlands 3...

[@GewelReal]

@@falcdragon The best CPUs to test cache scaling are X99 Xeons (to a point). in fact, I had 7 of them, and scaling is indeed dependant on the raw amount rather than % increase (architectural differences might make it less pronounced). Altough, Haswell/Broadwell Xeons somehoe stop scaling past ~30MB (the 50MB chip showed 0 increase in performance in all games over 30MB/35MB ones). But the best example for showing how anything below ~30MB is a massive disadvantage, as going from 1650v4 at 15MB to 2650v4 at 30MB INCREASED performance in all games despite 1650v4 running at 3.8GHz all core while 2650v4 only doing 2.5GHz. But then going from 2650v4 to 2673v4 at 50MB (2.6GHz all core) didn't actually increase performance (maybe 5-10% at most in 1-2 titles)

[@alltheotherhandlesaretaken]

Where’s the 5800X3D results?

[@Hardwareunboxed]

Haha 😄

[@user-wq9mw2xz3j]

intel 5800k3d

[@Collin_J]

Can't believe they didn't test the CPUs in 4K

[@frieza1016]

M8 didn't test in 8k, what a fucking w@nker!

[@seafafrage6934]

this joke is becoming cringy now.

[@EmanuelHoogeveen]

For *some* of these games, it does seem like cache makes a respectable difference. 36MB is about 9% more than 33MB, so when we're seeing a 4-5% performance increase that's still fairly decent scaling. Unfortunately there's only 2 data points with 8 cores so we can't extrapolate to say, 96MB - I'd expect it to plateau to some extent, but maybe nearly tripling it would still net a decent benefit. But I guess they'd have to lower the clock speed significantly to avoid frying the 3D cache so it probably wouldn't help overall for their current architecture, as you said. Edit: As others have said, there might also be workloads where you need to hit some magic number to see a large benefit - like 64MB or a little more on the side for smaller things. It's a pity that there's no real way to test for that except by looking at AMD chips (which have entirely different constraints).

[@WilliamJasonSherwood]

I'd argue that L3 would likely behave similarly to RAM or V-RAM where it kind of matters at thresholds, 25-36MB all behaves the same because they're all going out to main memory anyway. But if you got to say 50 MB or 80 MB, suddenly you've got enough L3 to be able to skip 80% of those main memory calls. You would need something stupid like the w9-3495X to be able to test it, and I don't even know if they'd support fafing about with cores & cache configurations. IMO that would be the way to test it, especially since with w9's potentially have even more L3 than the AMD parts with a maximum of 105 MB. BUT another consideration is that having the cache split up on 50 odd different cores while AMD has it all on say 8 cores, means the intercore transfer is pretty significant and perhaps THAT's what you are gaining on, perhaps EVERY core need access to say 25 MB of L3 to see these kinds of performance improvement, and say a shared L3 arrangement COULD perform the same.

[@conanpiggott9465]

I'm with you on this, there was clearly a difference in the 8 core samples and a mild difference in the 6 care samples.

[@CarputingYT]

Have a nice Saturday Mr. Hardware Unboxed man

[@AbbasDalal1000]

Man unboxed

[@notthedroidsyourelookingfo4026]

Tf you on about?

[@cocosloan3748]

Soon will be Sunday in Australia but ok...

[@yueqi7499]

its actually saturday night but good one

[@LlywellynOBrien]

​@@AbbasDalal1000that is their OF channel. They try not to mention it too much.

[@MarkHyde]

Can't get enough of this type of specialised CPU content. Thanks!!!

[@kevinerbs2778]

Hardware unboxed or someone should try testing this same theory with mutli-GPU setups vs single card setup. There's only few mGPU games to test this on anyways that actually use mGPU. I'm pretty sure clock speed is still better for mGPu than cache for it. As I'm someone who has RTX 2080 ti's in S.L.I & have noticed that my 5600x seemed to give better mulit-gpu performance from higher clocks (4.85ghz) over then 5800x 3D (4.45Ghz) I replaced it with. While the 5800x 3D seems to be happier pushing a single card better. Here is the list of mGPU games. 1. Ashes of the singularity = Natively supported for All cards even different vendors 2. Chasm the rift = Vulkan 3. Deus EX mankind Divided (DX12 AFR must have single card set in Nvidia's control panel.) 4. Rise of the Tomb Raider 5. Shadow of the Tomb Raider 6. Quake 2 RTX = Vulkan 7. Far Cry 5 8. Red dead Redemption 2 Vulkan 9. Gears of war 4 they'd probably only be able test this on Intel's Z690 boards with a I think a 12900K as that's the last boards on the intel side that support S.L.I/ while only about 5 x570 boards supported one being the MSI MEG ACE x570 board that I'm using right now. On side note that I think people don't know but All of AMD's RDNA line up support mGPU. All you do is with two of the same tier cards, find mGPU in the control panel & enable it restart & you're done. It does however disable Rebar in bios when enabled.

[@Ralipsi]

Now we do understand clearly why Intel are so focused on clock frequency. Thanks for the impeccable bench test and analysis! I absolutely love HardWare Unboxed. You guys know you stuff and do a great job. ...making it so easier for us to understand our hardware needs and what are the best gears to get.

[@RobBCactive]

A treacherous path to follow though.

[@alejandrocalori6298]

Wins on single threaded applications and lets people brag about intel just showing single core benchmarks, like some dumba** on twitter that I saw just now.

[@maynardburger]

@@RobBCactive I mean, it's not like they're designing specifically for clocks and nothing else. They understand the need to push IPC and whatnot as well. Alder Lake was a huge lift in IPC and I have no doubt Arrow Lake will see another decent lift here as well, especially if they have to lower clocks a bit for the new process a bit. Intel's first generation for a new process tend to not clock super well.

[@RobBCactive]

@@maynardburger but they've had to push both P & E cores high up the power wall to compete in benchmarks pushing frequency and using E for die area efficiency has caused sacrifices on AVX and we know to stability on the K series which try to compete with Zen4. The problem with IPC is games in general simply don't hit anything like the CPU theoretical values that numerical code can reach. I would not be surprised if the close difference between i7 & i9 cache is down to small size difference so they haven't found games with highly boosted hit % rate.

[@maynardburger]

@@RobBCactive Intel are still hamstrung a bit by being a whole node generation behind AMD at the moment(talking desktop here). If they were on the same or similar node as Zen 4, they could have relaxed things greatly in terms of power draw without sacrificing clockspeed. And whether or not any given game is fully utilizing architectural IPC improvements is kind of pointless to argue, because IPC still does play a significant part of improved performance in games. It was the main reason Alder Lake was as big a leap as it was in gaming performance.

[@GS-kh5se]

I understood how important cache was when my Pentium Pro with its fast L2 cache was still usable years after regular Pentiums had to be retired.

[@exscape]

I learned about it when I disabled the cache (only the L2 cache was disabled IIRC) on my Pentium 100. It's truly crazy how slow it became, tasks that used to take seconds could take minutes.

[@Sebastianino]

cache is a lot faster then any ram so yes better have more then less

[@dakoderii4221]

I had a bunch of different Intel chips back in 2014 with the same architecture but different cache amounts by disabling cores on high end i7s to match a lesser i7 or i5 or i3 but with more l3 cache available to them. I saw a marked improvement by just adding more l3 cache for each core. I was saying why don't just add a little bit more, at least on some models? I wasn't the only one. I've seen a few people talking about in videos and on forums years before it was announced by AMD.

[@SpoonHurler]

​@@Sebastianinonot just faster but orders of magnitude less latency... even if its a similar speed if you had something crazy like 8 channel ultra fast ram, the cache still is better for anything that fits in it... which is another important consideration for vcache, it allows more within it

[@pauldunecat]

The Pentium Pro was such a stud.

[@heeerrresjonny]

I'm confused by some of the conclusions here for certain games. For games where there does seem to be a difference based on cache, like in Dragon's Dogma II there does seem to be a notable uplift at times. DDII had a 4% uplift in avg FPS with only 3MB more cache. The X3D parts increase cache by much larger amounts, e.g. the 7800X3D is 96MB vs the 7800X at 40MB, a 140% increase. So, a hypothetical 14900K"X3D" might have something like 86MB of cache. At 6 cores, it is even more dramatic with a 14% uplift from 12MB more of cache and again a 4% uplift from 3MB of cache. That's about 1.2fps or 1.3fps PER MB OF CACHE lol. If you scale that up the way it'd be for an X3D part, a 14900KX3D might hit 130fps or higher (in theory). While clock speed seems to be a primary limiter, there is obviously always a limit where you can't increase clock speed further...and then more cache might indeed lead to even more performance. I feel like the conclusions for some games here aren't accounting for the variable amount of cache being added nor accounting for the much larger cache variation we see between X3D and non-X3D parts.

[@mgk878]

I agree, the FPS boost is actually decent in certain games. I am very curious where the scale ceiling is. It's plausible that a 48+MB 6-core chip would outperform the 14900K at least in DD2, while using slower memory, and running at lower power too. So why hasn't Intel made a chip like that? Maybe it raises cost too much, or they care too much about synthetic benchmarks that don't need much cache... Edit: looking at the higher clock results, I guess it wouldn't outperform it, but maybe just match it. Hm.

[@imglidinhere]

I think it's safe to say that Zen 3 was massively bandwidth-starved, with Zen 4 showing much the same result, and the true performance of that architecture can be seen once that problem was alleviated with the added V-cache. It would make sense given how heavily memory speed affected overall performance. I remember seeing as much as a 15% swing from using the slowest to the fastest memory, with Zen 3 at least. Intel isn't hobbled in the same way, clearly. The difference between using 4800mhz and 7200mhz memory is measurable, however minor, compared to AMD.

[@malcaniscsm5184]

Well I guess that's why AMD made a deal out of unifying the L3 for Zen3, so each core sees 32MB rather than 8MB for the Zen2.

[@naamadossantossilva4736]

You hit the nail in the head.

[@Angel7black]

Then the next thing i have to say is why do AMD know this about their product, but for a second time in a row now decided to first launch Zen 4 at marked up prices as a mediocre product that at best went nearly even with 12th gen Intel in gaming and way worse in productivity due to core count differences? Dont miss understand the point here im not being a fanboy and only comparing AMD vs Intel in gaming before Vcache to make AMD look Intel look better, im going after AND for willfully knowing they should be putting Vcache on all their chips but launching gimped for a year then putting out the chips the way they should be and jumping prices up again and charging damn near $500 for a 8 core 16 thread. AMD needs to stop milking the consumer base, it really doesnt even help them and Zen 5 needs tp be 3D vcache on ALL cars not this parked cores shit, that of learn how to do hybrid cores or make non 3D cores focus on background tasks while 3D cores focus on gaming similar to how big little cores operate. They still havent fixed that issue fully with the 7900X3D and 7950X3D parking or using the wrong cores fully

[@nepnep6894]

​@@naamadossantossilva4736Intel does still have great memory scaling, but speed does almost nothing on it's own. You need to tune subtimings to get any meaningful gains. AMD does also but a lot of the scaling is from fclk.

[@DeadAimKillah]

@@Angel7black Ummm you do know 3dvcache wasn't made for gamers right? 3dvcache is more efficient in there work making them suitable for server processors (look up Zen c cores) they didn't gimp there processors or hold back because they were better than intel. they aren't on all chips because they are expensive and they save most of them to go on their server c cores. this is a case of trickle down technology where we get a cool addition to our cpus because its something AMD thought we'd like. it was the same thing with HBM. but nvidia rendered that useless for gamers when they purchased the best memory chips to overclock them for GDDR5X

[@masterchi17]

I think the conclusion drawn is not entirely supported by the experiments. The cache increase is quite small from 24 to 36MB, while for AMD it is 32 to 96MB. So 64MB extra for roughly 15%, and we did saw for many 6-core tests a roughly 10% increase for 12MB increase. Also the 4-core results indifference to cache size can be explained that with less cores the game will launch less worker threads. And each thread has its own memory structures, so reducing the number of threads reduces also the pressure on the caches in terms of capacity. But apart from that, another excellent video answering the 6 to 8-core debate 😅.

[@glenndoiron9317]

Yeah, the 33MB -> 36MB cache increase isn't very big ( i9 means almost nothing for most games for 14th gen processors, and you probably should be snagging the i7 for a gaming system (assuming you don't want to deal with AM5 motherboards and/or 7800X3D stuttering issues while web browsing - if you're ok with those, then the 7800X3D is a better pick for gaming.)

[@aapzehrsteurer9000]

I was thinking the exact same thing. The difference between AMD's 3D cache and this is so huge that it may cross a threshold at which bigger data structures can be loaded into the L3 cache.

[@chriswright8074]

Intel architecture isn't memory sensitive zen always has been and eypc has great improvements with 3d cache

[@M3N04]

​@@chriswright8074💯

[@A.Froster]

​@@glenndoiron9317 First time i have heard of the 7800x3D stuttering issues

[@vulcan4d]

They did, the i7-5770C back in the lga1150 days. It was rare and expensive. I upgraded from a i7-4790k and whoa was it better. Games that stuttered, didn't stutter anymore...it was revolutionary. Unfortunately, they did not continue. Probably because Intel likes to stall back in the days.

[@twilightravens9798]

It was a bit different though, the L4 on the 5775C wasn't nearly as fast as the L3, and was on a separate die (was slightly faster than my DDR3 2400 but not by much). I used my 5775C for a while until the 4 cores became the bottleneck. Was a cool idea but wasn't really feasible for a more mass production run because 14nm yields at the time were terrible and Broadwell wasn't exactly known for hitting high clocks, with or without the L4.

[@Jason-ol1ty]

Some years way back, I had an old CPU where the cache on the chip failed, but the rest of the CPU worked. The chip was soooo slow from that, and Windows was not really usable so I went back to DOS until the replacement came. I had to get by on playing Master of Magic, which was the only program I had that worked fast enough lol!

[@techpupil-au]

Really great experiment mate, good stuff

[@MrAtthedrivein925]

Love the testing Steve! Seriously what a great idea to test. Very informative and we appreciate all you do!!

[@PowellCat745]

3D V Cache helps Zen 4 enormously because Zen 4 is primarily memory starved. The default 2000MHz Infinity Fabric for Zen 4 offers a per-CCD bandwidth equivalent to DDR5-4000. Intel running DDR5-7200 is equivalent to a Zen 4 CPU with a 3600MHz Infinity Fabric. This is my theory why L3 cache helps Zen 4 enormously while it barely helps on Raptor Lake. Edit: People don’t realize how much of a memory bottleneck there is on Zen 4. Steve tested the same thing for the 10th gen Intel CPUs, but the memory used was DDR4-3200, equivalent to a 1600MHz Infinity Fabric speed on Zen 4, which is merely 20% slower than an actual Zen 4 CPU. Steve’s Raptor Lake test bench uses DDR5-7200, equivalent to a 3600MHz Infinity Fabric speed. This is 225% faster than the 10th gen and 180% faster than Zen 4. That’s why memory bandwidth is rarely an issue on RPL and more L3 cache didn’t help much.

[@Hardwareunboxed]

In the past it has also helped Intel CPUs a lot.

[@PowellCat745]

@@HardwareunboxedYes, but the 10th gen only supports DDR4, and your 10900K test setup used DDR4-3200 CL14, which is equivalent to a 1600MHz infinity fabric, much closer to Zen 4’s memory bandwidth than that of Raptor Lake.

[@PowellCat745]

@@Hardwareunboxed maybe as a follow up, you can retest with ddr5-4000 😁

[@Aggrofool]

Referring to Intel Broadwell-S, L4 looked good because Skylake was stuck with early-gen DDR4.

[@aberkae]

This will happen again with AM5/Zen5 3d where Intel is heading towards 10 ghz ram kits.

[@Nunkuruji]

If Intel had a part that dropped e-cores, un-hampering the ring bus, and replaced the die space with cache, they could have probably consistently edged out the X3D parts in gaming. Oh well, things will get real interesting once re-architected parts start being produced with backside power delivery.

[@OGPatriot03]

Yea but now everyone is running Intel stock baseline, which brings all the numbers down a good bit.

[@jurgengalke8127]

Why weren't games that are known benefit from Vcache tested i.e. COD, Assetta Corsa, Forza etc?

[@PowellCat745]

Good point.

[@pedroferrr1412]

Even more, like FS, RF2 and in VR! I really do not understand why hardware sites, never benchmark CPU´s in VR... it´s the future, even if most do not think so, just try once, and you will never return to an pancake. I have just 1 eye, lost the other 1, two years ago, duo to cancer. But i still only play SIM games in VR, started many years ago, with 3 monitors, but never returned. I have 3 monitors connected to the PC, but just for work.

[@LupusAries]

Or DCS, Arma3 and if you want to eliminate the GPU bottlenecking, Falcon BMS 4.37.

[@HowlingMoai]

​@@LupusAriesthose games aren't used for benchmarking because they are much more variable and not as good for performance comparison because of how they are designed. A benchmark in Arma 3 can average 75 fps on one run but then get 65 fps on the next run (even though they should be identical).

[@LmgWarThunder]

You have a very good point. I think the biggest reason is that if you look at blanket benchmarks of the 7800 x3d versus the 7800x, you would find that it performs better in almost all games, not just some and these games would be included in that.

[@HanSolo__]

I'm reading comments - people seem to not understand the idea of this test and its results. Intel processors do not work the way AMD CPUs do. Those are different types of constructions.

[@2drealms196]

AMD's 3D cache makes use of a chip stacking/packaging technology (Chip-on-Wafer) that TSMC invented. So Intel would to make similar advances in its chip stacking capabilties before they'd be able to implement a similar stacked cache.

[@OGPatriot03]

The difference in cache is so small that I don't think you can draw those sorts of conclusions, for example 3D cache for AMD isn't a tiny increase in chache it's rather massive which makes a lot more room for potential gains. Some games don't use much of the cache and in those cases neither would see much or any benefit, however plenty of games do make use of it and I would imagine where those games gained, AMD would have similar and proportionately larger gains due to the increased cache.

[@K31TH3R]

Would probably be helpful to explain why it is the case instead of saying "they just work differently." AMD uses a "chiplet" based approach which relies on something called the "Infinity Fabric" (IF) for the communication between the memory controller (IMC), cores (CCDs), and SoC (which primarily handles PCI-E communication). Games are often very sensitive to latency, and the frequency at which the IF runs has a large impact on the overall performance of AMD CPUs since it directly impacts overall latency. But unlike the name would suggest, the IF can't actually run at infinite speeds, and because of this, it will always introduce additional latency between the CCD(s), IMC, and DRAM. The 3D cache significantly reduces how often the cores need to cross the IF to access DRAM, which means it will significantly reduce overall latency as long as the working set mostly fits within the 3D cache. Intel CPUs do not use chiplets (for now), they use a monolithic design with an integrated IMC, so accessing the DRAM has an overall much lower latency penalty than AMD's design. That being said, Intel CPUs will also benefit from additional cache, but since the DRAM subsystem already has much lower latency than AMD's design, AMD chips will tend to benefit more.

[@bretthake7713]

"some dorks will complain about 5ghz so we also ran 5.7" how did HU get so goated

[@limeisgaming]

what would be more interesting is a comparison between the i5 14500 @5Ghz which uses the old 1.25MiB L2 vs the i5 14600@5Ghz which uses 2 MiB L2 per P-Core just like the Server Cpus of 4th and 5th Gen Intel Xeon Scalable. i5 14500 6 P-Cores 6x1.25MiB L2 + 24MiB L3 vs i5 14600 6 P-Cores 6x2MiB L2 + 24MiB L3 and compare that to the relative results between a @5Ghz capped ryzen 7 7700 and a Ryzen 7 7800X3D (its also capped at 5Ghz) PS.: there does intel with "3DV-Cache" exist its those high Cache Emerald Rapids SKU's, with Emerald Rapids also using Golden Cove for its Cores it should offer some comparison. Intel Xeon Gold 6548N 32 Cores 32x2MiB L2 60MiB L3 (1.875MiB L3/Core) Intel Xeon Gold 6530 32 Cores 32x2MiB L2 + 160MiB L3 (5MiB L3/Core) Server(Xeon 4th and 5th) Desktop/Mobile(Alderlake, Raptorlake, Raptorlake Refresh) are all based (more or less) on "Golden Cove" Server has 1.875MiB L3 as its standard with 5MiB L3/Core for high Cache Variants all of which have 2MiB L2/Core Desktop had until "Raptorlake Refresh" (1.25MiB L2 + 3MiB L3)/Core, Raptorlake-Refresh has (2MiB L2 + 3MiB L3)/Core

[@Roll_the_Bones]

From what I understand from the original TechPowerUp tests, the 14600K can be OC'd up to 14700 level on a decent air-cooler, but the 14700 brews up if it's even mildly OC'd? And the 14700/14900's need to have substantial expensive cooling to gain any slight advantage, at least for gaming? So, I'm thinking for the 14th series, the 14600K might be the value/performance sweet-spot for gamers?

[@makobe584]

Would've liked to have seen some cache-limited games like Factorio, Assetto Corsa, and flight simulator.

[@themalcore_]

Something that would be worth mentioning is the higher the core clock speed (relative the memory access latency) the more exaggerated the penalty for instruction window full stalls will be, and therefore the more valuable large caches are.

[@tremelai]

Unlike the 10th generation, (Comet Lake) a refresh++ of Skylake which had a uArch that was initially released and optimized for throughput server workloads, the 12th generation Alder Lake and by extension, 14th Raptor Lake both using the Golden Cove and Raptor Cove P-Core respectively was initially released for desktop workloads and released later for server workloads in Intel's sapphire rapids. Intel has been down scaling CPUs uArchs to desktop/laptop workloads since the 686 Pentium Pro with the 12th generation being the first to go the other direction since the 586 Pentium. AMD followed Intel in the practice of designing and releasing uArchs for workstation/server workloads first with the K8 Athlon64 FX/Opteron release, followed on by the downscaling to socket 939 desktop variants. This trend continues today with the Zen uArch. Server/Workstation uArchs traditionally have more bandwidth, both IO and memory, than their desktop/laptop counterparts, as such, the integer and FPU/streaming execution units are optimized for the larger number of memory channels afforded to the server platform. An Epyc Zen3 sports 8-channels of DDR4-3200 or 204.8 GB/s or theoretical bandwidth. Compare that to the 57.6 GB/s of theoretical bandwidth that is afforded dual channel DDR4-3600. X3d was developed for Zen3 to afford more bandwidth to high core count Epyc CPUs and is not generally available on Epyc CPUs below 16-cores per socket. Intel going the other direction, of scaling up a desktop/laptop optimized P-core to server workloads, took the dual channel DDR5 optimized scheduler/execution unit core and scaled up the memory/PCI bandwidth to meet higher than 8 p-cores per socket. AMD, and Intel with 11 back to Pentium pro, would trim core count to find the upper limits of desktop parts while ramping frequency. With Zen3 and Zen4, going from 8-channel / 12-channel memory channels respectfully, allowed for a limited amount of trimming core count before you impact the ability to feed individual core's integer execution units with actionable work to do. Adding X3d to the desktop parts helped close that architectural bottleneck. Sizing core count can close the memory bottleneck only so much when your branch predictor is tuned for 4x more memory bandwidth. Same was seen on the 10th generation Intel where the conditions were much the same. (See Intel's Broadwell i7-5775C with it's 128MB of L4 cache, dubbed eDRAM.)

[@Skidzo19]

5:37 "Ratchet and Clank like Assassin's Creed mirage" Wow they're so in sync.

[@clifflenoir4323]

Interesting that 8 cores can occasionally give better performance than 6 cores, even with identical clock speeds and cache (Ratchet & Clank data @ 10:56)

[@saricubra2867]

More threads always means smoother 1% lows and frametimes in the long run. That is the main reason why i went for a 12700K instead of a 7700X or 7600X, the IPC is higher as well, i really don't care about clockspeeds and average framerates. A friend owns a 5900X and the game experience on any game is smooth AF without microstuttering or stutters, i watched the 5800X3D on RUST and it was a stuttering mess. The 7800X3D gets way higher averages than both and for now does have the 1% low lead, but once it runs out of threads, the 5900X will simply pull ahead (Zen 4 vs Zen 3 is just a mediocre IPC upgrade).

[@Redmage913]

One problem I see right now is that they’re on such a (relatively) large node that to modify current designs to include more cache could blow out the size of the processor to an unreasonable size and therefore price of each CPU to produce, and then have to sell. We can hope Intel’s smaller nodes work out, and perhaps the amount of L3 cache can be solved that way. I’d love Intel to be more competitive in the energy efficiency department for their high-end parts. I still buy their products - my dad recently got an Intel i7-1260P (I think) laptop that’s amazing. But its temps are all over the place on a whim and needs a lot of fan to cool in sustained work. Good chips, bad efficiency for work :/

[@Pedone_Rosso]

Science videos are the best videos. (Except when I'm actually interested in buying hardware shortly after watching the video) Thanks for your tests and videos!

[@thestrykernet]

Intel massively boosted the cache available with Emerald Rapids while keeping the same/higher clockspeeds and power consumption as the Sapphire Rapids parts they replaced. That's not a perfect comparison given the 2 vs 4 tile design, but it does at least indicate they wouldn't have to sacrifice peak clocks like AMD does. I think Intel's lack of extra cache at the consumer level has more to do with it being expensive for them to implement versus the volume of sales available to such a CPU. I'd be curious to see a 12100 added to the 4c results since it only has 12MB L3, but this would require one of the motherboards with an external clock multiplier. Really appreciate this sort of content as it's always good, unique and interesting.

[@drival060590]

I would find it super interesting to see how much power the 14700K/14900K(S) consumes with disabled e-cores in games at 5.0 GHz (8c/16t). A Ryzen 7 7800X3D clocks similarly and it would be interesting to see what Intel's efficiency looks like when the processors are undervolted or clocked significantly lower at 5.0 GHz compared to completely over the top clock frequencies.

[@michaelmcconnell7302]

what an extremely useful and informative video. thanks, steve.

[@willfancher9775]

Great work. The only issue I have with this testing is that 3D vcache adds a lot more cache than you tested here, and the performance effects might not be linear. It could be that there is some threshold of cache that, once surpassed, begins having more significant results. Of course there's no way to test this, so there's not much use speculating. It would be really cool for someone more knowledgeable in chip design to drill down and explain the "why" of these results, though.

[@naamadossantossilva4736]

Not good when Intel is already facing silicon degradation issues with the 250W PL2.

[@rodrigoferreiramaciel4815]

I wish you guys made some big comparison between generations of one single family of chips, like taking the last 5 or 10 i5's and looking at how much improvement each gen brings. That would be great for anyone that is looking to upgrade an old pc and is considering buying used pieces of hardware. I'd love to see how it compares between intel and amd

[@xxovereyexx5019]

In general yes all cpu & gpu can get a benefit from stacked cache, but each cpu & gpu has a different case or problem. Stacked cache is one of the end game in gaming, smart move by amd

[@kosmosyche]

Common sense tells me it's a great way to decouple CPU performance from memory and memory controller speed and bandwidth restraints. After all this is the main reason cache was invented - to minimize memory addressing as much as possible, and especially when it slows things down quite a bit and holds CPU performance. I wonder if the benefit from bigger L3 cache would be higher if tested with lower clocked RAM or even with DDR4 RAM. Basically my theory is there needs to be more of a memory bottleneck for a bigger L3 cache to really start to shine and make a difference.

[@jamegumb7298]

@@kosmosyche L2/L3 will always be faster than ram.

[@PowellCat745]

@@jamegumb7298Adding more cache can hurt latency and therefore performance. Bigger doesn’t always mean better, but since Zen 4 is severely bottlenecked by the infinity fabric, it benefits a lot from more L3 cache.

[@jamegumb7298]

@@PowellCat745 It _can_ add latency but does not have to. On Zen it does because chiplets. If one core needs from the cache on the other chiplet, latency penalty. I have a 7900X3D since a few weeks and measured it once using a bench tool. Latency to the big cache is hit but if both chiplets had the cache it would be far less an issue with a good scheduler. IF is another factor but I do not think it is _that_ bottlenecked by it, at least not in every scneario. The large X3D cache makes _some_ applications (mostly games) faster but for other stuff I'd have been better off using the 7900X.

[@PowellCat745]

@@jamegumb7298 more cache always results in higher latency. Do you even know how cache line works? AMD tried to stack two 64MB cache blocks on **one** CCD and performance went backwards. It has nothing to do with 2 CCDs.

[@tldrinfographics5769]

You know what’s really amazing? Intel being able to keep its speed crown for this long without TSMC advantage. They did it all without chiplets or 3nm… Imagine if 14th gen was on 3nm

[@depth386]

This is fascinating, in the context of how different it is from that 10th gen. I remember watching that. Thank you so much for making this video Steve.

[@nurofenusa1]

@ Steve - I am always impressed by your deep dives, research and project management that must be undertaken in your efforts to produce content that really digs deep into the hardware and related information that no other channel is prepared to undertake. The information you provide graphically is easily digestible by the layman, and gives me insights on bettering the communication to others I may be working with on projects. You and Tim put in the hours where its needed and have a great synergy together on screen. Oh and I'm a Pom so take that priase for what it is :)

[@sublime2craig]

I have asked myself this same question many of times inside my own head, so happy you guys are here to give me an answer!!! Love this kind of testing, please keep em coming!

[@REgamesplayer]

This is why I love Intel chips. They focus on what is truly important for a CPU. A single threaded performance. 7800x3D doesn't do that and it runs terribly single thread apps like MSI Afterburner. It also tend to bottleneck at its main core. Whatever we like it or not, most apps are not excellently optimized. There will always be a higher dependence on a single core and CPU manufacturers should move towards prime core concept. Making a primary core a lot bigger and powerful than the others. Then they should add performance cores and then efficiency cores which they exclusively use for background applications. Such a design is a lot better suited for high performance gaming.

[@J_..._]

I know Intel and AMD have different methods of evicting lines in the L2/L3 cache. Explaining the differences between them might make for an interesting short~ish video.

[@CakeCh.]

12:27 So I'm interested to see a test results with some Xeon CPUs.

[@yeahitsme4427]

It would be good to join 7800X3D to the charts as a comparison, because its an 8core part with 5GHz boost clock. The same clock you limited in the Intel series.

[@ETophales]

There's a major difference between AMD and Intel with regards to the L3 cache in that AMD's L3 cache is linked to all cores, and remains the same size when fewer cores are available, while Intel's L3 cache has 3MB per P core or four E cores, so if a CPU has fewer cores, the L3 is reduced. This leads to quite a few questions, the first of them being: when you disable cores, do the other cores still have access to their cache? This would affect how valid the results here are. I think it would be good to verify this (by using some software which reports memory access performance for different data sizes) just to make sure. (Edit: There does seem to be a change in results based on the amount of cache, which might indicate that this isn't a problem, but it doesn't always show. After testing that, it would be interesting to test whether there's higher latency accessing the cache of another core, though that probably won't be trivial to test.) (Edit 2: Raptor Lake has 2MB L2 cache per core, as opposed to 1MB for Zen 4 and 512KB for Zen 3, which could be why L3 cache is less important on the Intel side.)

[@ramiad90]

Good review, but note that the difference in cash between the i7 and i9 here is just 3mb With a larger cash pool for the cores to use freely things might be different So to wrap it up we woulnd never know for the 14th gen if more cash would help in gaming

[@OGPatriot03]

Obviously more cache is better if it doesn't sacrifice speed/latency.

[@jlgroovetek]

In some of the tests with 8 cores, 5% performance increase from just 36mb vs 33mb cache. If there was another 30mb increase, there could be 50% increase based on that, which becomes very significant, no?

[@Petch85]

Interesting video. But what is actually stored in the L3 V-cache when gaming. Course 36 MB to 96MB is quiet a lot more than from 33-36 MB. Maybe 96 MB is enough for some games to reduce the number of calls to memory needed for each frame, but 36MB requires several calls. That might not be a linear scaling. But with a linear scaling The last of Us part 1 gained 5% by adding 3 MB cache (9%). Thus adding 60 MB (266%) would give you 100% (133% if you used the %). This is clearly not true, course L3 cache and performance do not scale linear. And this is only for one game. But we already know that X3D chips only benefits in some games. But I think your data shows that Intel could benefit form having more cache. If we try to imagine the perfect memory system for a CPU, that would be a memory system that always have the data needed available for computation. Memory faster than that would not provide any benefit. Thus if the RAM was as fast as the L3 cache that would be the same as the L3 cache having the same size as your memory. The question is, how many clock cycles is an Intel CPU wasting while gaming. I don't know if that is something any tool could tell, it might only be intel who cooled even hope to know this. But if 50% of all cycles are wasted infinit memory speed could only give os a 100% performance improvement. But I have no idea how many cycles are wasted in a cpu while running a game. I tried looking for a Intel Xeon CPU with more cache, but you end up with minimum something like a W5-3435X (16 cores 45 MB Smart Cache) or the w7-3455 (24 cores 67.5 MB Smart Cache) (or w9-3495X, 56 cores 105 MB Smart Cache). They are all expensive or super expensive, they might be ok at video editing if you want to give it a go 😂. But it is going to be tuff, the motherboards and BIOS on this type of systems are often very limited and non of them can run at 5 GHz (unless you find a way to OC them), so you have to rerun the benchmarks at 4.5 GHz or something, to do the testing.... I would defiantly watch the video, but I think it would be hard to get a good return on investment unless you finde something else you also could test. Maybe number of memory channels, gaming with 1TB of memory, (google chrome with 4 TB memory 😂), lov clocked P cores vs E cores (course the W5, 7 and 9 only have performance cores, thus they have to clock down when using 16, 24 or 56 of them) , cooling test (I think they limit them to 270W all core). Maybe others can come up with better idears but those videos sound more like a LTT video than a HU video to me.

[@danielt91]

It would be interesting to see the exact same test with the Ryzen 7000 lineup. Again - same clock speed, same core count, just differences in l3 caches (including the x3ds) At a later time it could even be used to compare the IPC gains of Ryzen 9000 🙂

[@jimmyjiang3413]

I mean going forward needs a 3D V-cache variant for workstation CPUs as well cuz maybe bigger L3 cache shall also help pro visualization CAD rendering workloads in addition to using pro GPUs. This is where there maybe a necessity for creating Ryzen Pro variant for both workstation OEMs and some of SKUs available to general public (through retail for select higher end SKUs). There, there shall be need for dual 3D V-cache equipped CCDs for workstation rendering loads, and similar for Ryzen Threadripper Pro with 3D V-cache for even more workstations

[@WilliamJasonSherwood]

Not disagreeing with your results, but it is possible that there is a threshold where between the 36 MB L-3 Cache of the Intel Parts and the nearly triple 96 MB of the X3D parts. Cache would likely function similarly to ram/v-ram, where if you either have enough to contain all the things you frequently need, or you don't and going from say 8GB to 10GB won't make a significant difference, but going to 16GB suddenly increases performance by 60% (okay, maybe 30%) or something ridiculous. While an absolutely ridiculous use, one potential way to test for this could be using something like the Intel w9-3495X with its 105 MB of cache and disabling the appropriate number of cores, and testing different cache configurations. But they only clock to like 4.8 Ghz.

[@HDJess]

If they would just slap 3dv-cache on current intel cpus, nothing much would happen. The architecture needs some tuning as well to make use of that extra cache. The software, including games, need to be optimized too to make good use of that cache. It is likely that a game's engine running on an intel CPU with 36MB of L3 cache would not even know what to do with an extra 100MB of L3 cache, if the developers haven't tested and built the game's engine for it.

[@felipeavlopes]

I wonder if the performance reduction when dropping to 4 cores isn't caused mosly by the Operating System itself rather than anything else? Modern Windows are resource gluttons that love to use more resources than what would really be necessary for computers to work.

[@TrueThanny]

I don't think this testing really shows what would happen by adding another 64MB or so of L3 cache. The cache differences with 10th gen were more significant, and starting from a lower baseline. Raptor Lake is starting from a higher baseline, and the cache differences are smaller. Basically tripling the L3 cache would very likely have a more significant difference, even if the method of adding it required lowering the clocks.

[@AtomSymbol]

If the same method to extrapolate usefulness of 96MB L3 cache on Intel CPUs as used in the video was used in a scientific article, the scientific article would be rejected by reviewers.

[@dominicsilvestre9412]

Remember i7-5775C with 128mb L4 cache? good times.

[@chromedonut2441]

Maybe Intel should consider having someone else manufacture their chips for them since they are till having trouble shrinking their process node. How would the performance differ for 14th Gen if say it were done on the same process node as Zen 4?

[@yuvanraj2271]

What are you talking about? They are already in the process of finalizing the 15th gen arrow lake chips. They have intel's smaller node as well as tsmc's small node.

[@mingyi456]

It seems the problem with intel delays now is their design teams, not their foundry teams. Their nodes seem to be ready long ago, just that there is nothing to use them. Ironically, intel also seems to have a cost issue with their nodes, because using tsmc now seems to save them money over using their own nodes, even after letting tsmc get their share of profits.

[@moist_ointment]

RPL was DTCO'd too heavily with Intel 7. Design and Fabs were too tightly coupled and only newer gens separated that. Porting RPL to another node would've been too expensive

[@TTM1895]

In 2016, Opera was acquired by an investment group led by a Chinese consortium

[@thejumper9303]

You expect that every test is just frequency limited, even though you never tested 2 different cach amounts vs frequency. It could also be engine hard frequency caps like many Ubisoft titles/ports as well as scheduling problems, engine only utilizing 4/6 cores,.... But for you it can only be frequency dependent. Which might be a symptom but most currently not the main cause

[@최치헌-x6x]

It would have been better that tested with heavily Cache bound games. Such as COD, Assetta Corsa, Forza, Starcraft and Factorio etc. Most games in this review, (except Cyberpunk) usually are not very well to test it because they are mainly more GPU-bound AAA titles.

[@Magmafire]

Today's Intel 13th and 14th gen cpus are like the Prescott Pentium 4s from the past.

[@StudyWithNika]

I hope intel can catch up with amd in terms of gaming performance at adequate power consumption with their next gen cpus. Hopefully they will make a good use of the latest asml machines that they recently installed

[@lietuviss1]

It would nice to see the architecture differences, different cache coherence protocols between gens and VS AMD maybe? Faster DDR5 ram could reduce the bottleneck of fetching from ram on cache misses.

[@TheZoenGaming]

My first thought is maybe, but only a little since the V-cache is more heat sensitive and Intel's chips seem to be less thermally efficient than AMD's.

[@cracklingice]

Not that it matters much, but it'd be interesting to have this data side by side with some ryzen data for the same selection of titles.

[@ShermSpinner]

Very interesting results, especially with the 1ßth gen testing you meantioned in mind. I'll say it would be interesing to see how or if that scaling behavior changes with slower memory, given that on AMD it really helped with memory bandwidth limits. Though of course i understand that you cannot possibly test everything and there is definitely an argument to be made that even *if* running something like base JDEC DDR5 would flip things around and make cache a massive deal on Intel it wouldnt really change the conclusion at all, given that intel can run these higher speeds just fine.

[@GeneralLee131]

With NVME SSD bandwidth approaching that of RAM memory, and the discovery that cache size is an apparent bottleneck in all systems, i can envision a near future where L2 cache is up to a gigabyte, L3 cache could grow to several gigabytes, RAM is entirely deleted from the system, and the remainder of left up to the swap file on the disk.

[@vladislavkaras491]

Oh, so games starts to utilize more than 4 cores! What a great news for Xeon owners! :D Thanks for the benchmarks!

[@Frozoken]

lower core counts probably increases per p core usage meaning the individual l2 cache runs out more and the l3 as a result has to get used more aa a result

[@pino_de_vogel]

i have a feeling the architecture has bin stretched so much the past 4-6 years that there are other limiting factors within the architecture that just cant keep up with the cache or clocks.

[@AgentSmith911]

I wonder what AMD will change for their 9800X3D, maybe add more cache, or increase speed and reduce latency, or even all of that combined.

[@XBnPC]

Intel need TSMC, 3D V-CACHE, Santa Cluas & Jesus to potentially compete again 💯

[@conza1989]

Always good to see these videos, just another factor to consider in purchasing, do I go for the better part, not because of more cores but more cache. Ryzen 8000/9000 (probably 9 isn't it), might be interesting to see if there's differences... If they sell enough parts for there to be cache differences, aside 3D Vcache of course.

[@dabombinablemi6188]

I'll always be tempted to buy an i7 5775C, just because of the edram/L4 cache. From what I remember the performance in games placed it second to the 4790K.

[@stephenkennedy6358]

I wonder if we would see a bigger difference between 8 and 6 cores if hyperthreading was disabled

[@chrischen6664]

Nice video Steve, but how about another video on 3D V-cache on the Phenom x6 1055t?

[@cooldudep]

What?

[@tqrules01]

No. they need to first create the interconnect fabric that could support fast DDR5. This is the main issue between monolithic and chiplets. Intel needs a good interconnect and learn how to seperate logical cores and interlaying different Layers of cache so that there is minimal latency.

[@ctjmaughs]

Worked for the Pentium Pro back in the day

[@sgdude1337]

As a 5775C owner I’d love to replace my chip with another Intel big cache chip. Maybe call it a 15775C

[@IPowder4105]

With all that heat they can't use 3d v-cache

[@maynardburger]

Has nothing to do with it. The CPU's do not inherently put out a ton of heat.

[@Avantime]

Microsoft Flight Simulator has massive improvements, like outlier-type of performance with 3D V-cache.

[@andre_ss6]

This makes sense, as Ryzen processors were always very sensitive to memory bandwidth and latency, while Intel were not, indicating already that they would not benefit much from larger caches.

[@p_mouse8676]

Thanks for the grwat work as always 👍 Maybe it would be interesting to make a video about why AMD does scale up well with cache but Intel CPUs don't? It's now stated as a conclusion coming from all the testing, but it leaves the why question open.

[@marufulislam4311]

Here is a Question For Q&A Series? Given the if a GPU (4070 super) can run a at 120+ fps at 1080p, it will surely run that game 60+ Fps at 1440p My question is which way would you guys prefer to game at? would you prioritize visuals & play at 1440p 60 FPS? Or would you make no comprise with frame rates & play 1080p at 120 FPS?

[@ramonzaions7522]

Whit this Voltage And Tempture 3D Cache will burn!

[@ThePurplePassage]

And here I was thinking that Adamantine would enable Intel to retake the performance crown

[@Bluelagoonstudios]

Because they lacked, in the previous years, innovations and kept "upgrading" old CPUs to another "generation". And now they are on their limits. With a company, you cannot sit back and cash. It's a struggle every day.

[@Gnarfendorf]

Given that the 7800x3d also only has 8 cores and doesnt really clock higher than 5ghz iirc, maybe the issue is that even with 36mb l3 its simply not enough to have a noticable impact?

[@SlyNine]

Probably because putting a blanket over that oven is too scary.

[@pvdgucht]

More cache seems to make a huge difference in flight simulators. More than in any other “game”. Idk why, the only reason I can think of is that maybe the aerodynamic calculation fits inside the cache and because the sim calculates that 2 to 4 times per frame that might make a difference idk 🤷🏼‍♂️

[@saricubra2867]

Cache increases memory bandwidth, Intel isn't affected by cache as much as AMD because of the monolithic ring bus design, the memory controller will simply be fast and it feeds the cores quickly. Pair both AMD or Intel CPU with base JEDEC DDR5 4800 and Intel will always be faster. AMD is bottlenecked by the Infinity Fabric, always has been.

[@pvdgucht]

@@saricubra2867 Makes sense thanks for explaining 👌!

[@saricubra2867]

@@pvdgucht Simulators always have been bottlenecked by memory. It's the main reason why my 12700K with DDR4 is slower than a 5800X3D (gets a bandwidth increase with the extra cache), but if i use DDR5, it's faster (specially when i have the hybrid design with 8 very powerful cores and physically big on the sillicon die).

[@Doshi-lp4tn]

The takeaway should be, demand games be made on universal optimized game engines instead of those game engines that benefit from x, y or z resulting in gamers needing to overspend on PC hardware toys to play a 60$ game.

[@Vladek16]

very interesting. So intel is boosting frequencies because that's their only lever to increase performance on this architecture

[@hflx]

That's why you need your own Power Plant to run a Intel CPU... it needs POWER and cores, would be nice to understand why cache benefits AMD more , probably Intel need to rethink their entire architecture ...

[@hotscott6619]

know do same test with thread ripper how do we know all cashe used

[@Hardwareunboxed]

TR is completely different and the L3 cache is per CCD.

[@xsvrrx]

so the cache is also usesless for cores when they are turned off?

[@korcommander]

If only Intel can get that heat under control

[@SoficalAspects]

Aren't the AMD 3D V-cache methodology covered by a bunch of patents? I suppose Intel could come up with something new and equivalent, but given their last decade of development, I assume it's going to be a huge challenge?

[@Hardsky5123]

Intel did that extra cache thing in the past so they are first xD

[@PowellCat745]

@@Hardsky5123 eDRAM is not close to being the same thing as V cache.

[@wewillrockyou1986]

Yes, sort of, and not really. Though, with Intel's die configuration it might actually be better to have a base die to handle IO and cache rather than having a simple "cache block extension" that AMD uses.

[@Aggrofool]

It's actually TSMC's technology, not AMD's. Intel has Foveros so there is no patent issue.

[@VoldoronGaming]

@@Aggrofool Incorrect. The technology is AMD's. What is TSMC is the packaging.

[@ivonakis]

The cache impact is a problem which can be simulated so I am sure Intel and AMD have optimized the cost benefits. As for why no benefit in this test - I think the jump in cache is small and It would require a scenario where 33MB was not enough but 36 is. My question is why no x3d chips in mobile - the combination is set to dominate mobile gaming - low clocks and power consumption

[@joshhernandez8680]

This sounds like R&D testing for Intel, are we watching you work

[@Indyfficient]

I’m curious though, are these results due to hardware limitations or due to the manner in which amd vs Intel have their scheduling programmed for instance?

[@eugkra33]

Intel did add extra cache. That's why the 14th gen is faster than the 12th gen, even at the same clock speeds.

[@ylstorage7085]

attaching a question here: why isn't L2 cache size included, I thought intel had far larger L2 cache than AMD.

[@eugkra33]

@@ylstorage7085 probably because there is no way to manually tune it. Increase it. If you disable cores the L3 stays the same and is used by every core. It's shared. But L2 does get disabled when disabling cores. So if you take a 14900k and disable cores you don't get a 14600k. You get a 14600k with more L3, showing you gains of extra L3. But you do get the L2 structure of a 14600k. Meaning you can't see what happens if you increase the L2 with these CPUs. You could probably throw in a 12600k, though, to compare to 14600k. But all that gets too complicated.