Discussion Speculation: Zen 4 (EPYC 4 "Genoa", Ryzen 7000, etc.)

[Vattila]

Except for the details about the improvements in the microarchitecture, we now know pretty well what to expect with Zen 3.

The leaked presentation by AMD Senior Manager Martin Hilgeman shows that EPYC 3 "Milan" will, as promised and expected, reuse the current platform (SP3), and the system architecture and packaging looks to be the same, with the same 9-die chiplet design and the same maximum core and thread-count (no SMT-4, contrary to rumour). The biggest change revealed so far is the enlargement of the compute complex from 4 cores to 8 cores, all sharing a larger L3 cache ("32+ MB", likely to double to 64 MB, I think).

Hilgeman's slides did also show that EPYC 4 "Genoa" is in the definition phase (or was at the time of the presentation in September, at least), and will come with a new platform (SP5), with new memory support (likely DDR5).

What else do you think we will see with Zen 4? PCI-Express 5 support? Increased core-count? 4-way SMT? New packaging (interposer, 2.5D, 3D)? Integrated memory on package (HBM)?

Vote in the poll and share your thoughts!

 

[uzzi38]

If I see another speed demon design I'm gonna blow a fuse.

Hope you got a replacement ready then.

 

[coercitiv]

Hope you got a replacement ready then.

Got an entire pack, been blowing fuses since 2019: power, clocks, power, clocks, power, clocks.

 

[Qwertilot]

Have to slightly agree with that sentiment. Especially when Apple are showing how incredibly well you can do going the other way.

 

[uzzi38]

Have to slightly agree with that sentiment. Especially when Apple are showing how incredibly well you can do going the other way.

Eh before you get too disappointed about all clocks and little IPC maybe wait and see what the power curve is like?

What matters at the end of the day even more than IPC itself is perf/W improvement, because that has the biggest impact on both server and mobile.

 

[Abwx]

Some really awful napkin math here, but just trying to see what kind of IPC gains are required to get 31% more performance in Blender than the 12900K.

If we assume that Gracemont has an IPC value of 1 or roughly similar to Skylake, Golden Cove should be 1.4 or so. If we assume that the clocks for the 12900K are 5 GHz for GC and 4 GHz for GM, that's a "score" of: 8x5x1.4 + 8x4x1 = 88.

Meanwhile, if we assume that Zen 4 has roughly the same IPC as Golden Cove and runs at 5 GHz all core, the "score" is: 16x5x1.4 = 112.

The ratio is: 112/88 = 1.27.

Again, super rough math and some liberal assumptions, but it's clear that the IPC uplift has to be in the double digits. The >15% ST claim is just that. Greater than 15%.

5950X is 16% faster than 12900K in Blender quick benchmark, i guess that it s the basis of the comparison.

AMD Ryzen 7 5800X3D im Test: Benchmarks in Anwendungen und Spielen

AMD Ryzen 7 5800X3D im Test: Benchmarks in Anwendungen und Spielen / Der Einfluss des Testparcours / Das Leistungsrating lässt sich anpassen

www.computerbase.de

So Gigabyte leak was spot on. 1MB of L2 cache (with same 8-ways sadly), minor updates to core ( increased L2 TLB ) and AVX512 support.

I've always wondered where some guys here are pulling those outrageous claims of "30% IPC increase and 5.5Ghz clocks". The end result is what was predicted by the leaks: Zen3 with 1LMB of L2, same 32MB of L3 on release, DDR5 and 5-10% IPC increase.

Last number floating around was 18%, but without any precision about the frequency/IPC breakdown it was supposed to be all IPC...

 

[Saylick]

5950X is 16% faster than 12900K in Blender quick benchmark, i guess that it s the basis of the comparison.

AMD Ryzen 7 5800X3D im Test: Benchmarks in Anwendungen und Spielen

AMD Ryzen 7 5800X3D im Test: Benchmarks in Anwendungen und Spielen / Der Einfluss des Testparcours / Das Leistungsrating lässt sich anpassen

www.computerbase.de

It appears that the 12900K keeps up and marginally beats the 5950X when the Blender render is small. If it's a long render, the 12900K loses clocks and falls behind. Not sure what duration workload was used to create that 31% better Blender claim.

 

[biostud]

well, i thought that 5800x3d was sort of a test run for the v-cache technology and it would become standard by 7000 generation, at least on say ryzen 9 parts.

I think they'll do it as part of the refresh down the line, depending on how competitive Raptor lake is and probably only on the 7800X, as it doesn't seem to benefit professional workloads running on the 12c and 16c parts.

 

[coercitiv]

What matters at the end of the day even more than IPC itself is perf/W improvement, because that has the biggest impact on both server and mobile.

Is that why they're following Intel with higher max TDP on AM5, because their perf/w is great over 5Ghz? On the server side they're increasing core count, it's a different game.

 

[insertcarehere]

It's either Cinebench ST is not the goal that AMD aimed for, or the SMT performance with Zen4 is larger than expected which would be enough to generate 31% advantage against 12900k in Blender.

View attachment 61942

My guess is that in MT workloads they are clocking Zen 4 significantly harder than Zen 3. A stock 5950x is at ~4ghz when fully loaded with Blender et al, my guess is that's closer to 5ghz with the Zen 4 equivalent, likely due to increased TDPs and power limits.

 

[AtenRa]

Is that why they're following Intel with higher max TDP on AM5, because their perf/w is great over 5Ghz? On the server side they're increasing core count, it's a different game.

I believe the 170W is for the platform not the SKUs.

 

[JoeRambo]

Last number floating around was 18%, but without any precision about the frequency/IPC breakdown it was supposed to be all IPC...

Hey, let's not walk back on outrageous claims of some forum members around here:

Zen4 will have massive IPC advantage even without the V cache. Also there will surely be Zen4 parts with Vcache which will widen the gap even more. AMD is positioned really well to have top to bottom performance and performance/watt crown.

And so on.

But i think an important point is that Zen4 will still be very awesome CPU, if they can achieve >5Ghz all core 6950x i will be very happy buyer.

 

[Abwx]

It appears that the 12900K keeps up and marginally beats the 5950X when the Blender render is small. If it's a long render, the 12900K loses clocks and falls behind. Not sure what duration workload was used to create that 31% better Blender claim.

View attachment 61943

204 seconds for the 7000 and 297 for the 12900K according to the footnote.

 

[Saylick]

204 seconds for the 7000 and 297 for the 12900K according to the footnote.

Got it, so somewhere in between the two tests that LTT ran.

 

[biostud]

Hey, let's not walk back on outrageous claims of some forum members around here:



And so on.

But i think an important point is that Zen4 will still be very awesome CPU, if they can achieve >5Ghz all core 6950x i will be very happy buyer.

I believe the 6xxx parts are zen3+ while zen4 will be 7xxx

 

[Abwx]

Got it, so somewhere in between the two tests that LTT ran.

You ll notice that the 7000 is 45% faster...

 

[AtenRa]

You ll notice that the 7000 is 45% faster...

ehm it is 31% less time

 

[Vattila]

Like many, I found the presentation somewhat underwhelming, although I was impressed by 5.5 GHz gaming speeds and 30% advantage over "Alder Lake" in Blender. With that, I guess "Raphael" should do well against Intel's "Raptor Lake". At least, AMD will have clear performance per watt leadership, if not absolute performance leadership.

The lack of core count increase this generation will disappoint some, though. As I posted previously, I hoped to see AMD somehow take a step to 24-core, to stay in the lead and fend off Intel's 24-core (8+16) "Raptor Lake", but it seems we'll have to wait for "Zen 5" or beyond for more cores.

I get the feeling that the focus of "Zen 4" and its platform design was to catch up with Intel's feature set, i.e. clock frequency, AVX-512/AI acceleration, iGPU, DDR5 and PCI Express 5. In terms of core design, I guess it is mostly a frequency optimised shrink of "Zen 3", with AVX-512/AI and enlarged L2.

Unless there is more going on than meets the eye, it didn't look that they have done much in terms of advancing packaging either. It looks like it is chiplets on organic substrate in pretty much the same arrangement and topology, but likely highly optimised. The eventual V-Cache models will be interesting.

I had hoped for product line-up and launch schedule to be announced. With the lack thereof, I guess they are not quite through bring-up and qualification yet.

 

[Saylick]

You ll notice that the 7000 is 45% faster...

Yeah, that's odd. Maybe they ran the math as: 1 - 204/297 = 0.31 ---> 31%.

Edit: AtenRah beat me to it lol.

 

[coercitiv]

I believe the 170W is for the platform not the SKUs.

ofc it's platform related, that's why I wrote AM5 instead of 6950X...

 

[AtenRa]

ofc it's platform related, that's why I wrote AM5 instead of 6950X...

yes you are right you said AM5.
ps. I believe you mean 7950X.

 

[Abwx]

ehm it is 31% less time

31% less time is 45% faster, isnt it...

So the perf is quite good although it s not the uplift some hoped for in respect of Zen 3.

 

[JoeRambo]

The lack of core count increase this generation will disappoint some, though. As I posted previously, I hoped to see AMD somehow take a step to 24-core, to stay in the lead and fend off Intel's 24-core (8+16) "Raptor Lake", but it seems we'll have to wait for "Zen 5" or beyond for more cores.

16C are PLENTY for tasks CPUs excel at and given AMD's focus on clocks, 5+Ghz MT performance will be awesome.
I fully expect it to beat 13900K in Blender and probably run around equally in Cinebench ( cause Blender is heavier on FP and loves stronger cores, while Intel's small cores are good in CB23 and less strong in previuos CBs that were FP heavier).

The real question is, why would anyone run rendering workload on CPU in year 2022?




Yes, 3090 undervolted is 27 times faster than eight 5Ghz Alder Lake P cores without HT. Like 230W vs 130W too... I am sure 4090 will increase the gap to 40-45 times this summer.

 

[uzzi38]

Is that why they're following Intel with higher max TDP on AM5, because their perf/w is great over 5Ghz? On the server side they're increasing core count, it's a different game.

You know perfectly well that power consumption isn't linear with clocks, especially not when trying to push >5GHz across a full 16 cores.

 

[yuri69]

So did it really take them 23 months (assuming a Sep launch) to design a Zen 3 shrink with widened ISA support?

The $$$ shouldn't have been such an issue compared to dire Zen 1/Zen 2 times. Is frequency optimization such a big deal since, heck, Bulldozer derivates are no that rusty, right?

Platform and core development should not affect themselves.

questions, questions, questions!

 

[biostud]

So did it really take them 23 months (assuming a Sep launch) to design a Zen 3 shrink with widened ISA support?

The $$$ shouldn't have been such an issue compared to dire Zen 1/Zen 2 times. Is frequency optimization such a big deal since, heck, Bulldozer derivates are no that rusty, right?

Platform and core development should not affect themselves.

questions, questions, questions!

Maybe they poured the money into zen5 as they knew zen4 would fare well enough against intel?

For me I still think I will stick to my upgrade plan of going AM5 for these reasons:

-No matter what everything will be a lot faster than what I have now
-Seems like they plan for several generations of CPU's
-I'll get a zen4 7800X now, and the fastest gaming CPU @ AM5 EOL, maximizing system lifetime
-Hopefully be able to get a PCIe5 nvme drive when from start
-Wait until RDNA3/Lovelace cards are lanched