ZEN 4: Explaining AMD's Strategy

At Computex 2022 AMD shared first infos about the upcoming Zen 4 CPUs and the new AM5 platform. In this video I will talk about what I think AMD focused on while developing Zen 4, why it will be a major step for AMD but could turn out to be less exciting for the average gamer.

0:00 Intro
0:43 Intel's Market Force
2:34 Zen 4 Performance
3:57 Integrated Graphics
4:50 AVX-512
5:20 Multi-Thread Focus
6:05 I/O & AM5 Platform
6:45 Final Thoughts

[knofi]

You have good arguments about AMD's focus and these points are also very important for AMD's future success. But I think AMD is actually sandbagging this time too. Marketing usually promotes performance gains with 'up to percentage' values. But AMD only says >15% and could probably easily say 'up to 30%, or up to 40%' instead. In addition, they concealed the actual performance increase in the render test and comparison with the Intel 12900K, because the render times of Ryzen 7000 (204 sec) compared to Intel 12900K (297 sec) actually result in a performance gain of 45.6% (297/204). Of course, one could also say that the Ryzen 7000 finished rendering in 69% of the time (204/297) of the Intel 12900K, but normally one extrapolates the percentages from the slower processor (100% performance) to infer the performance gain.

[squelchedotter]

Great Video! I was disappointed too at first, but the more I thought about it, the more it made sense

[OverseerPC]

Yes, that dedicated RDNA part in Zen4 was really interesting to me when Lisa announced it. Your argument is solid regarding OEM requirements and features that AMD needs to tackle. Good one, mate. Subscribed! 👍

[ramr]

I watched your video a second time, very insightful indeed! So much information to digest! Good to see AMD producing a better rounded product.

[TrueThanny]

It's not that AMD is hiding the full performance specs. They don't know them yet. It's too early to know exactly what clock speeds they'll be using across the SKU's, so they don't actually have a full suite of applications tested for IPC and overall uplift.

What they showed was that the 16-core pre-production chip was more than 15% faster than the 5950X in a single-threaded Cinebench run. Cinebench is very much on the low end of IPC uplifts in general, especially those achieved mostly with additional cache (such as the doubled L2). So saying more than 15% is very conservative.

If I had to guess, I'd say the average IPC is probably right around 10%, give or take. And the single-core clock speed is up at least 10%.

The biggest performance change with Zen 4, however, is the vastly higher all-core boost speeds. Over 5GHz on all cores under load. It's too soon to say exactly how high they go with all 32 threads fully loaded, but the fact that they got >40% more multi-threaded performance than the 5950X in Blender shows it's pretty significant.

[theminererz]

Level headed POV as per usual. Exactly why I come here for another take on tech news(besides myself). Always appreciate the input/insight! I couldn't agree more! I was rather disappointed there was no info on the new GPUs, but I can understand considering it's so close to their "6050 refresh" release but still, those who are waiting....are waiting. A little official news would have been good though imo. Would have provided a solid foundation for the inevitable speculation that would have kept expectations based on such speculation from getting out of hand and keep people talking. Oh well, not like I will be expecting them any time soon. Thanks!

[TrueThanny]

It's worth pointing out that DDR5, PCIe 5.0, and integrated graphics aren't features of _Zen 4._ They're features of _Raphael._ AMD could, if they wanted to, release Zen 4 on AM4, with a new I/O die to drive DDR4 and PCIe 4.0 instead (could even have integrated graphics on it). They chose those features for AM5 specifically. They aren't something brought along from the server market, where the I/O die is completely different.

AVX-512, however, is something brought along for the ride.

Oddly enough, AMD looks like it will be the first company to have AVX-512 support across all its consumer processor line (assuming it's preserved in _Phoenix, _ which is something I don't know). While Alder Lake supports AVX-512, it does so only if you disable the E-cores. Despite all the bluster about it not being a problem long before launch, they were completely unable to sort out the scheduling of cores with different instruction availability. Whether it's mostly Intel's fault or Microsoft's fault doesn't much matter - it doesn't work, in the end.

So to the extent that more mainstream software gets AVX-512 support, it's probably going to be attributable to AMD rather than Intel, which is really quite odd.

[opinali]

The feature-set catching up is indeed critical, AMD eventually had to deal with both AVX-512 and standard iGPU capability. And they have a similar problem on the GPU side, where the RDNA2 generation was superior to their competition in some important areas: rasterization, perf/power, mobile, APUs, price; but there was always a "but". But it doesn't have DLSS. But raytracing and video encoding not good enough. But CUDA. But 4K resolution, due to narrow VRAM bus. So now it seems that RDNA3 and FSR2 will make big steps towards most of these areas of deficiency. Overall, I expect the Zen4+RDNA3 combo to be in a much stronger position to grow market share, not because they will win in every benchmark and every fine-grained feature checklist (no single product ever does that), but because they won't be significantly behind, or completely missing, any important feature. Integrators or DIY consumers won't have those red-line reasons to not even consider AMD's offering for particular use-cases e.g. media creators or engineers that use the few apps that really need AVX-512 wouldn't even look at Ryzen no matter the price or other features. Now they will be in every race.

[FredBKevinn]

The first logical explanation about Zen4 that i have heard post the presentation

[imperiousleader]

Thank you for the thoughtful insights! I totally support the strategic rationale you put forward. I tend to think that gaming seems to be mostly a GPU battle - leaving the CPUs to battle on the broader issues of compute (server) vs user experience (home) - and the use case each person/company has will determine everything. As a home user who tinkers with programming I will hugely welcome AVX512 as this will immediately double my current compute application performance - depending on the instruction sets supported, and the iGPU will probably be a significant issue for all those office PC deployments around the world where you need a basic display for MS Office and data entry. For the home gamer I can see that this release may not offer the big ticket uplifts seen by the AMD 5800X3D and Intel Alderlake type processors.

[devlmn]

46% faster is just enough to keep them ahead of Raptor Lake, but Intel will likely release Meteor Lake and Arrow Lake before Zen 5 arrives

[maxwellsmart]

AMD's market share constraints are probably more due to wafer allocation as opposed to adoption resistance. OEMs need volume and Intel delivers that along with marketing money etcetera. As far as Zen 4 performance, I would guess a >8% IPC with a ~16% frequency boost to give it at least a 25% performance increase over Zen 3. As usual it will depend on the application or type of game running at the time. Maybe the rumors of a 3D v-cache part coming sooner than later is to bolster the IPC. We'll all know soon enough.

[jonathonfidiam]

Awesome analysis. On a 3900x here total overkill for gaming
Not gonna switch until Q4 2024 at the soonest

[RobBCactive]

I've just noticed speaking after a few seconds in would be better as YT isn't playing the sound for a few seconds, unless I block ads entirely (but then some videos don't play). Some creators have a little filler, intro, then have a theme section. Tested it with Ublock Origin and I heard you right away but it probably kills Adsense.

[thevaultsup]

Zen 4 was all about wider Zen 3 with memory subsystem overhaul aswell as caches optimization (as seen from patents), it looks like they actually traded wider Zen 3 for much higher frequency which is fine as long as the results are the same, and will help with a bunch of workloads like games, they have tho to significantly improve the FCLK to reach such gains as it is the biggest bottleneck for games with Zen 3, even overclocking the cpu in some cases doesn't matter as the fabric can't keep up.
Multi-core improvements looks solid, 40% is pretty much expected, as for ST i guess 20% is what to expect, seeing how easily it hits 5.5GHZ i would expect higher, remember also the screenshot of a B650 motherboard running an engineering sample at 1.55V, that some crazy stuff, sounds like the biggest bottleneck this gen will be heat management.
Also i expect cross-ccd communication throught RDL bridge, that why the 2 ccds are so close of each others, which will greatly improve latency and why not have a virtual L4 too (Where one CCD can access the others CCD cache, imagine V-cache Zen 4 then... but that pure speculation at this point).

[eugkra]

I want to know how the L2 cache changes will impact things. Some clock for clock comparisons would be interesting when it releases.

[joechang]

the reality of the last decade plus is that frequency only helps workloads that run inside cache, mostly the L1/L2 as L3 latency is rather high and is perhaps really there to reduce memory bandwidth load. Hence IPC gains on "existing binaries" largely impractical, aside from brute force larger cache which has limitations of its own. This does not mean that tremendous are not possible by completely rewriting critical code to not jump all over memory but rather design around what is in L1/L2 and perhaps even using the AVX256/512 registers in imaginative manners.
Everyone could probably benefit from a true properly multi-threaded compression/decompress algorithm, having different code paths and methods for single large files vs. many small files and bypassing compression on already compressed content.
My question is: while there is code that benefits from AVX256/512, is this type of code better done in the CPU or GPU? already, a large part of the core is SIMD, yet many apps do not use it. Hence should we have more smaller general purpose CPU cores with a big SIMD engine as a separate entity?

[Pillokun]

yep, zen4 is an alderfication of zen3, ie getting more l2$ will probably make it jump up in perf to the same levels of alderlake/5800x3d, and if they increase the amount of l3$ by just a bit(the ccd looks big with only l2$ increase) it will naturally be even better perf wise than alderlake/5800x3d in gaming. But I would love zen4 being put on the am4 cpu package for ddr4. So the upcoming price hike in am5 platform would be more acceptable imo. Am5 boards looks super OP feature wise = expensive.

[strongforce]

I think they may be sandbagging a little bit, to confuse intel's raptor lake lineup but also..think about it, they showed us core clock numbers, why not showing actual FPS gains this time arround ? It may be too early but.. I don't know, it's what, 4 months away now max ?

[DavidStanleymusic]

In waiting for a 16 core with a 12 core integrated graphics Dragon range with stacked cache