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!

 

[DrMrLordX]

Or maybe they are only doing a 7800X3D, and not a regular 7800X.

That would be a wise assumption.

CB is everything but a good indicator for average ST performance.

It seems to be a worst-case scenario for Zen4's ST performance uplift vs. Zen3. CBR23 anyway. AMD has indicated "15%+", where 15% is CBR23 ST.

 

[inf64]

It seems to be a worst-case scenario for Zen4's ST performance uplift vs. Zen3. CBR23 anyway. AMD has indicated "15%+", where 15% is CBR23 ST.

Even that is questionably low. They stated 8-10% IPC jump in specific apps (amongst which is CBR23). If the top chips clock north of 5.5Ghz, then the uplift is at minimum ~21%.
It only makes sense if the claim covers the whole Zen4 range where 7600X has the lowest ST boost of approx. 5.2Ghz => 5.2/4.9 x 1.08 ~= 1.15

For top SKU, if it can reach the 5.8Ghz for ST as some rumors indicated, then we have : 5.8/4.9 x 1.08 ~= 1.28 or 28% higher ST CBR23 score versus 5950X.
If the IPC in Cinebench is closer to the upper cited band (~10%), the uplift is ~30%.

 

[DrMrLordX]

It only makes sense if the claim covers the whole Zen4 range where 7600X has the lowest ST boost of approx. 5.2Ghz => 5.2/4.9 x 1.08 ~= 1.15

That's probably the case.

 

[DisEnchantment]

Even that is questionably low. They stated 8-10% IPC jump in specific apps (amongst which is CBR23).

Zen 4 will not have 8% IPC in CB R23. AMD added SPEC and GB besides CB in the footnotes for the Zen 4 IPC estimation because using only CB would make the IPC gains look bad.
Zen 4 does not introduce additional FP (or even int ports, from leaked manual) execution ports. Similarly, compiler configuration for execution cost remains unchanged and therefore we can surmise that no change in throughput as well. For FP execute bound loads like CB where the cache sub-system and FE is mildly engaged, low single digit IPC gains at best.

 

[inf64]

Zen 4 will not have 8% IPC in CB R23. AMD added SPEC and GB besides CB in the footnotes for the Zen 4 IPC estimation because using only CB would make the IPC gains look bad.
Zen 4 does not introduce additional FP (or even int ports, from leaked manual) execution ports. Similarly, compiler configuration for execution cost remains unchanged and therefore we can surmise that no change in throughput as well. For FP execute bound loads like CB where the cache sub-system and FE is mildly engaged, low single digit IPC gains at best.

If that's true, than SPEC and GB would need to provide much larger ST uplifts so the average is close to 10%. I don't see that happening, as GB gains are usually very close to Cinebench gains, so that would leave only SPEC to pull up the average by a lot (think >18%)..

 

[inf64]

I just checked the latest leak on twitter (Genoa ES), and Zen4 ES @ 3.5Ghz scored 1227 pts in R23 ST test. On the other hand Zen3 (both non-Vcache and Vcache) @ 3.52Ghz scored 1151/1143 pts. This shows ~7% IPC increase based on the ES results for 1T.

EDIT: Zen3 server ST scores in R23 are in line with 5950X's results (almost perfectly) : 1151 x 4.9/3.52=1600 pts

CPU 2021 Benchmarks - Compare Products on AnandTech

CPU 2021 benchmarks: Compare two products side-by-side or see a cascading list of product ratings along with our annotations.

www.anandtech.com

 

[DisEnchantment]

GB gains are usually very close to Cinebench gains,

GB scales with many things besides FP... cryptographic instructions, bandwidth, int instructions, cache latency, etc. CB is mostly FP execution bound.

I just checked the latest leak on twitter (Genoa ES), and Zen4 ES @ 3.5Ghz scored 1227 pts in R23 ST test. On the other hand Zen3 (both non-Vcache and Vcache) @ 3.52Ghz scored 1151/1143 pts. This shows ~7% IPC increase based on the ES results for 1T.

Based on what data you are sure that it is running fixed at that clock?

 

[Abwx]

I just checked the latest leak on twitter (Genoa ES), and Zen4 ES @ 3.5Ghz scored 1227 pts in R23 ST test. On the other hand Zen3 (both non-Vcache and Vcache) @ 3.52Ghz scored 1151/1143 pts. This shows ~7% IPC increase based on the ES results for 1T.

EDIT: Zen3 server ST scores in R23 are in line with 5950X's results (almost perfectly) : 1151 x 4.9/3.52=1600 pts

CPU 2021 Benchmarks - Compare Products on AnandTech

CPU 2021 benchmarks: Compare two products side-by-side or see a cascading list of product ratings along with our annotations.

www.anandtech.com

I pointed that those results are dubbious, not only there s no IPC gain in R23 but there s also large regressions in R20 and R15.

FTR 5950X score 1657 in R23, so that s actually a regression as well.

Intel Core i9-12900K, i7-12700K & i5-12600K im Test: Benchmarks in Anwendungen

Intel Alder Lake im Test: Benchmarks in Anwendungen / Ein rekordverdächtiger Testumfang / Windows 10 vs. Windows 11

www.computerbase.de

 

[inf64]

GB scales with many things besides FP... cryptographic instructions, bandwidth, int instructions, cache latency, etc. CB is mostly FP execution bound.

Based on what data you are sure that it is running fixed at that clock?

The guy who did the tests said so? You can see the specs of the Genoa ES in the chart, the max clock is ~3.5Ghz. There is also a screenshot while it's running ST tests, the clock is ~3.46Ghz...

I pointed that those results are dubbious, not only there s no IPC gain in R23 but there s also large regressions in R20 and R15.

FTR 5950X score 1657 in R23, so that s actually a regression as well.

Intel Core i9-12900K, i7-12700K & i5-12600K im Test: Benchmarks in Anwendungen

Intel Alder Lake im Test: Benchmarks in Anwendungen / Ein rekordverdächtiger Testumfang / Windows 10 vs. Windows 11

www.computerbase.de

The R23 ST score for Zen3 server parts is ~3% off from the desktop part, which is normal due to slower DDR4 memory. The test is not perfect, the guy said it's all done on ES HW and pre-production BIOS/firmware. It still gives us a ballpark values for R23, at least.

 

[Thibsie]

Yes and AFAIR it was also stated that bios was unstable. Can we really derive any conclusion at this point ?

 

[inf64]

Yes and AFAIR it was also stated that bios was unstable. Can we really derive any conclusion at this point ?

Yes, the R23 ST was still 7% faster at the same clock vs Zen 3 , even with all the bugs. It's valuable information.

 

[Abwx]

The R23 ST score for Zen3 server parts is ~3% off from the desktop part, which is normal due to slower DDR4 memory. The test is not perfect, the guy said it's all done on ES HW and pre-production BIOS/firmware. It still gives us a ballpark values for R23, at least.

You mean Zen 4.?.
Genoa use DDR5 and even then RAM has no influence in ST scores.

If the score is 1151@3.46GHz then it would score 1663@5GHz, that s the same score as the 5950X@5GHz since Computerbase sample run at this frequency for CB ST and wich is measured at 1657.

Edit : Only way the leaked score would be relevant is that frequency hoover somewhere between base frequency and 3.46GHz

 

[inf64]

You mean Zen 4.?.
Genoa use DDR5 and even then RAM has no influence in ST scores.

If the score is 1151@3.46GHz then it would score 1663@5GHz, that s the same score as the 5950X@5GHz since Computerbase sample run at this frequency for CB ST and wich is measured at 1657.

No, I meant Zen 3. Zen 3 parts scored 1151pts in R23, check the chart again. That puts it in line with 5950X, adjusted for clocks (~3% slower as it's server vs desktop).


Genoa ES part scored 1227pts @ 3.5Ghz(~3.46Ghz), which is 7% faster than Zen 3 parts at around the same clock.


This means that 7950X @ 5.8Ghz would score 5.8/4.9 x 1.07 x 1650 = 2090 pts. Knowing that for Zen3 the server parts scored ~3% lower due to memory being slower (my guess), the best case scenario for 7950X is 2090 x 1.03~= 2152 pts

 

[Abwx]

No, I meant Zen 3. Zen 3 parts scored 1151pts in R23, check the chart again. That puts it in line with 5950X, adjusted for clocks (~3% slower as it's server vs desktop).


Genoa ES part scored 1227pts @ 3.5Ghz(~3.46Ghz), which is 7% faster.
View attachment 65096

In that case there s your 7% IPC improvement, and eventually a tiny more since the 5950X doesnt really hold the 5GHz mark in ST due to short frequency dips, that s why AMD stated 4.9 officially.

AMD Ryzen 5950X, 5900X, 5800X & 5600X im Test: Analysen zu IPC, Takt und Latenzen

AMD Ryzen 5000 im Test: Analysen zu IPC, Takt und Latenzen / AMD vergleicht die IPC jetzt anders / IPC-Zuwachs im Testparcours

www.computerbase.de

If Zen 4 can sustain solidly 5.5GHz in ST that s good for 18% ST perf improvement in R23.

 

[inf64]

In that case there s your 7% IPC improvement, and eventually a tiny more since the 5950X doesnt really hold the 5GHz mark in ST due to short frequency dips, that s why AMD stated 4.9 officially.

AMD Ryzen 5950X, 5900X, 5800X & 5600X im Test: Analysen zu IPC, Takt und Latenzen

AMD Ryzen 5000 im Test: Analysen zu IPC, Takt und Latenzen / AMD vergleicht die IPC jetzt anders / IPC-Zuwachs im Testparcours

www.computerbase.de

I'm not sure if you are now agreeing or disagreeing with me . I thought I showed clearly how R23 ST results for Zen 4 are 7% better at the same clock Vs Zen 3, at least in the leaked Genoa ES benchmarks. There are issues in this leak such as low R15/R20 ST scores, but that could be due to low ST clock for Genoa parts (for whatever reason).

 

[Abwx]

I'm not sure if you are now agreeing or disagreeing with me . I thought I showed clearly how R23 ST results for Zen 4 are 7% better at the same clock Vs Zen 3, at least in the leaked Genoa ES benchmarks. There are issues in this leak such as low R15/R20 ST scores, but that could be due to low ST clock for Genoa parts (for whatever reason).

I agree but i wouldnt count on 5.8GHz, so far leaks point to 5.52GHz in ST, that will be enough to roughly match the competition in R23 ST, be in agreement with the stated 8-10% IPC improvement and largely exceed 15% ST perf improvement.

 

[inf64]

I agree but i wouldnt count on 5.8GHz, so far leaks point to 5.52GHz in ST, that will be enough to roughly match the competition in R23 ST, be in agreement with the stated 8-10% IPC improvement and largely exceed 15% ST perf improvement.

Regarding the maximum ST clocks for Raphael, I was referring to this : https://www.angstronomics.com/p/computex-beyond-the-coverage?s=w

"Regarding frequency targets, the game demo showing 5.55GHz maximum frequencies was also not with the final version. While Angstronomics is aware of an Ordering Part Number (OPN) that is fused for a 5.85 GHz Fmax, we will have to wait and see what the retail stepping fuses will be set at. "

~5.5Ghz was shown to be an easy clock for Raphael ES even in a gaming workload on 1+ threads, so higher than that is a reasonable assumption.

 

[FangBLade]

Wait for reviews d'not assume

When it comes to Intel you hype every leak, but when it comes to AMD it is better to wait for reviews, you are same on every forum.

 

[deasd]

Anyone tell me how reliable Angstronomics is? I remembered he said AMD aimed 7% IPC uplift for Zen4 few days after Computex, and more few days later AMD official said 8-10%. I wonder if '7% IPC' and '5.85 GHz Fmax ' are conservative number that AMD intentionally told journalists.


And Cinebench is a strange application to say the least. IIRC, some review sites data I collected, result vary from version to version for some newish CPUs like Zen3 and Alderlake.

just for IPC comparison. IIRC

R10, Alderlake won Zen3 with double digit %
R11.5, both tied, ADL won slightly but closed to 0%
R15, both tied, no difference
R20, ADL dominating just like R10
R23, Intel enlarge the lead (~2-3%) compared to R20, despite the fact that rendering scene is same


OTOH, when comes to Genoa leak, R23 has 7% uplift while R20, R15 has regression. Despite the ES immaturity, I guess CB result varying might due to Cinebench itself but not ES Genoa's fault.

 

[Abwx]

Anyone tell me how reliable Angstronomics is? I remembered he said AMD aimed 7% IPC uplift for Zen4 few days after Computex, and more few days later AMD official said 8-10%. I wonder if '7% IPC' and '5.85 GHz Fmax ' are conservative number that AMD intentionally told journalists.


And Cinebench is a strange application to say the least. IIRC, some review sites data I collected, result vary from version to version for some newish CPUs like Zen3 and Alderlake.

just for IPC comparison. IIRC

R10, Alderlake won Zen3 with double digit %
R11.5, both tied, ADL won slightly but closed to 0%
R15, both tied, no difference
R20, ADL dominating just like R10
R23, Intel enlarge the lead (~2-3%) compared to R20, despite the fact that rendering scene is same


OTOH, when comes to Genoa leak, R23 has 7% uplift while R20, R15 has regression. Despite the ES immaturity, I guess CB result varying might due to Cinebench itself but not ES Genoa's fault.

Cinebench is a best case scenario for Intel for obvious reasons starting with R11.5...

R10 likely use X87 FP instructions.

R11.5 use SSE2 and was designed using Intel s software suite like Thread Profiler and compiled with ICC.

R15 extended instructions up to SSE4.2 as well as using Embree, wich is an Intel devellopped renderer.

R20 extend instructions up to AVX.

Dunno what has been done to R23 but likely that it was recompiled to take advantage of the soon to be relesed ADL uarch as well as being directly ported for Apple s uarch.

 

[inf64]

Cinebench became popular after Zen 1 launched. I remember intel telling press that it's not real world test, but now that they have advantage they are all over it
According to all these leaks, it's obvious that AMD will reach near parity with Alder/Raptor lake in R23 ST test, maybe trail by a single digit %. It will be interesting to see how much can AMD push all threads on top 32T part, that will determine if they get the overall performance crown Vs Raptor Lake.

I think the top parts (no Vcache on AMD side) will be neck and neck, in almost everything - AMD having the perf./watt advantage of course. Once they launch the Vcache parts, AMD should win in gaming (and by a considerable margin, my guess is ~15+%).

 

[eek2121]

Regarding the maximum ST clocks for Raphael, I was referring to this : https://www.angstronomics.com/p/computex-beyond-the-coverage?s=w

"Regarding frequency targets, the game demo showing 5.55GHz maximum frequencies was also not with the final version. While Angstronomics is aware of an Ordering Part Number (OPN) that is fused for a 5.85 GHz Fmax, we will have to wait and see what the retail stepping fuses will be set at. "

~5.5Ghz was shown to be an easy clock for Raphael ES even in a gaming workload on 1+ threads, so higher than that is a reasonable assumption.

5.85mhz fmax means a 5.7-5.75ghz max boost clock fyi, the chip would never hit 5.85ghz. Likely the 7950x has a 5.7ghz boost and a base clock in the mid 4ghz range.

 

[eek2121]

Cinebench became popular after Zen 1 launched. I remember intel telling press that it's not real world test, but now that they have advantage they are all over it
According to all these leaks, it's obvious that AMD will reach near parity with Alder/Raptor lake in R23 ST test, maybe trail by a single digit %. It will be interesting to see how much can AMD push all threads on top 32T part, that will determine if they get the overall performance crown Vs Raptor Lake.

I think the top parts (no Vcache on AMD side) will be neck and neck, in almost everything - AMD having the perf./watt advantage of course. Once they launch the Vcache parts, AMD should win in gaming (and by a considerable margin, my guess is ~15+%).

If rumors are true, we should have the v-cache parts soon. Hopefully they won’t have to reduce clocks this time around.

 

[inf64]

5.85mhz fmax means a 5.7-5.75ghz max boost clock fyi, the chip would never hit 5.85ghz. Likely the 7950x has a 5.7ghz boost and a base clock in the mid 4ghz range.

Yeah, I'm aware of that - it's the same thing as fmax for 5950X. What I was trying to say is that 5.5Ghz for ST is a low number IMO, especially if they demonstrated multiple threads running at 5.5Ghz in a gaming workload. Whether they reach 5.7 or 5.8Ghz is anyone's guess, but the difference is really small (~1.5%).

 

[eek2121]

Yeah, I'm aware of that - it's the same thing as fmax for 5950X. What I was trying to say is that 5.5Ghz for ST is a low number IMO, especially if they demonstrated multiple threads running at 5.5Ghz in a gaming workload. Whether they reach 5.7 or 5.8Ghz is anyone's guess, but the difference is really small (~1.5%).

It is possible that the number may have even gone up by now. Non of the chips AMD had floating around were final.