Question Zen 6 Speculation Thread

[511]

On SPEC anyone remember when Zenver3 was slower than Zenver5 🤣🤣

 

[Josh128]

On SPEC anyone remember when Zenver3 was slower than Zenver5 🤣🤣

Shoot, I remember when Zenver1 was slower than Zenver3!

Re: SPECintRate 2017 as being a good benchmark for the average users desktop PC, what an ridiculous notion. Lets see, I can download and run 5 versions of Cinebench, who knows how many versions of Geekbench, Blender, VRay, and multiple other benchmarks before I can even get a working executable of sir2017, not to mention I have to pay for it. Even ChipsNCheese says its challenging to run and literallly takes hours to run after you've compiled it, and results vary depending on how you compile it, lol.

 

[Nothingness]

Even ChipsNCheese says its challenging to run and literallly takes hours to run after you've compiled it, and results vary depending on how you compile it, lol.

Yeah SPEC is definitely not only a CPU benchmark: it's both a compiler and CPU benchmark (and to a lesser extent, an OS/lib benchmark).

The CPU design companies I know run SPECrate with 1 CPU; it's much faster than SPECspeed (and less prone to compiler tricks to autoparellize). And even beyond that, CPU companies tend to stick to the same compiler and same set of flags for long: all they're interested in is how much faster their new uarch is compared to the previous ones, and you can only do that for binaries that don't change for years.

SPEC is definitely a benchmark for pros, not for the average user.

 

[511]

SPEC is definitely a benchmark for pros, not for the average user.

i thought we all were pros on the forum certainly average user won't come on the forum and discuss 300 pages on one topic

 

[Abwx]

Shoot, I remember when Zenver1 was slower than Zenver3!

Re: SPECintRate 2017 as being a good benchmark for the average users desktop PC, what an ridiculous notion. Lets see, I can download and run 5 versions of Cinebench, who knows how many versions of Geekbench, Blender, VRay, and multiple other benchmarks before I can even get a working executable of sir2017, not to mention I have to pay for it. Even ChipsNCheese says its challenging to run and literallly takes hours to run after you've compiled it, and results vary depending on how you compile it, lol.

Cinebench is a FP bench, there s nothing further from usual consumers apps than CB, 7Zip is a much better metric since it s INT.

Beside CB R2X and 2024 are the only rendering tests where Intel has a ST advantage, in Blender, Corona, Povray and VRay AMD has better IPC be it ST or MT, so even for FP it s a mediocre bench to compare AMD to Intel since it s at odd with all other such tests.

 

[Markfw]

Cinebench is a FP bench, there s nothing further from usual consumers apps than CB, 7Zip is a much better metric since it s INT.

Beside CB R2X and 2024 are the only rendering tests where Intel has a ST advantage, in Blender, Corona, Povray and VRay AMD has better IPC be it ST or MT, so even for FP it s a mediocre bench to compare AMD to Intel since it s at odd with all other such tests.

I just usage phoronix tests, sometimes chips-and-cheese. They usage a wide variety of tests and are fair, and good summaries.

 

[Nothingness]

i thought we all were pros on the forum certainly average user won't come on the forum and discuss 300 pages on one topic

I didn't want to dismiss people on that forum, but I doubt anyone here is ready to shell $1000 and spend days fine tuning a benchmark

As all rules, there is at least one exception I know of here.

 

[DrMrLordX]

I doubt anyone here is ready to shell $1000

That is a dealbreaker for most.

 

[OneEng2]

Guys, the cadence is always 4Q’s.

I think the cadence between generations will match the cadence of process improvements.

A14 without BSPDN would likely be a cadence for Zen 6-Zen7 of about 2 years. If AMD decides they need BSPDN (and they might well decide this for DC), then it might be more like 2.5-3 years.

Zen 5 (mobile) was released in July 2024. I suspect we will not see Zen 6 until late 2026 ... so likely > 2 years.

 

[luro]

I think the cadence between generations will match the cadence of process improvements.

A14 without BSPDN would likely be a cadence for Zen 6-Zen7 of about 2 years. If AMD decides they need BSPDN (and they might well decide this for DC), then it might be more like 2.5-3 years.

Zen 5 (mobile) was released in July 2024. I suspect we will not see Zen 6 until late 2026 ... so likely > 2 years.

Zen 6 mobile is CES afaik

 

[OneEng2]

Zen 6 mobile is CES afaik

From what source? Best I can see it's 2027

 

[SmokSmog]

33% more cores = 70% better MT performance means that 27.5% contains clock gain and IPC.

My guess:
10% IPC 16% clock bump or 12% IPC and 14% clock bump

Desktop Zen6:
1.5*1.1*1.16= 1.914

So >90% gain in multi-core over 9950X.

 

[yuri69]

33% more cores = 70% better MT performance means that 27.5% contains clock gain and IPC.

My guess:
10% IPC 16% clock bump or 12% IPC and 14% clock bump

Desktop Zen6:
1.5*1.1*1.16= 1.914

So >90% gain in multi-core over 9950X.

We learned a lesson with Zen 5 - do not estimate any other segments based on a server performance figure.

Compared to desktop:
* Server got a completely different IOD (different characteristics)
* Server is gonna get more memory channels
* Server workloads behave differently (memory b/w, TDP limitation)
* Server is gonna get the TDP headroom expanded (we don't know whether desktop gets that too)

 

[adroc_thurston]

We learned a lesson with Zen 5 - do not estimate any other segments based on a server performance figure

You should also learn a lesson of a single vs double shrinks.

 

[reaperrr3]

We learned a lesson with Zen 5 - do not estimate any other segments based on a server performance figure.

Compared to desktop:
* Server got a completely different IOD (different characteristics)
* Server is gonna get more memory channels
* Server workloads behave differently (memory b/w, TDP limitation)
* Server is gonna get the TDP headroom expanded (we don't know whether desktop gets that too)

Maybe, but:

• Zen5 used the same IOD as Zen4, Zen6 gets a new one (better IF and better mem controller are safe bets)
• Zen5 used only a slightly improved node vs. Zen4 (Zen6 uses a much faster node)
• Zen5 had no L2/L3/CCX improvements worth mentioning (Zen2 doubled L3/CCX, Zen3 doubled CCX, Zen4 doubled L2), while desktop Zen6 increases L3 per CCD/CCX.
  Last time the CCX/max. amount of accessible L3 per core/thread grew in desktop (Zen3), games got an above-average IPC benefit out of it.
• Zen5 invested a lot of the added transistors into full-rate AVX512 with doubled FP_PRF, while most desktop workloads are rather Integer-bound and INT_PRF only grew by a measly 16 entries from 224 to 240, still smaller than Sunny Cove's 280 (Ice Lake from 2019) and one major bottleneck now, according to something adroc mentioned before.
  So a pretty low-hanging fruit to pick with Zen6 (anything less than growing IntPRF to at least 288 would be disappointing; best-case would be 336, matching int PRF/ALU ratio of Zen4).

Anyway, all the biggest deals of Zen6 - faster memory support, bigger CCD/CCX, (presumably) stronger INT focus, and of course much higher turbo clocks - are good for gaming perf, which is what desktop is largely about.
For desktop, I expect Zen6 to be a bigger absolute upgrade over Zen5 than Zen4 was over Zen3, because of the reasons above.

 

[BorisTheBlade82]

Maybe, but:

• Zen5 used the same IOD as Zen4, Zen6 gets a new one (better IF and better mem controller are safe bets)
• Zen5 used only a slightly improved node vs. Zen4 (Zen6 uses a much faster node)
• Zen5 had no L2/L3/CCX improvements worth mentioning (Zen2 doubled L3/CCX, Zen3 doubled CCX, Zen4 doubled L2), while desktop Zen6 increases L3 per CCD/CCX.
  Last time the CCX/max. amount of accessible L3 per core/thread grew in desktop (Zen3), games got an above-average IPC benefit out of it.
• Zen5 invested a lot of the added transistors into full-rate AVX512 with doubled FP_PRF, while most desktop workloads are rather Integer-bound and INT_PRF only grew by a measly 16 entries from 224 to 240, still smaller than Sunny Cove's 280 (Ice Lake from 2019) and one major bottleneck now, according to something adroc mentioned before.
  So a pretty low-hanging fruit to pick with Zen6 (anything less than growing IntPRF to at least 288 would be disappointing; best-case would be 336, matching int PRF/ALU ratio of Zen4).

Anyway, all the biggest deals of Zen6 - faster memory support, bigger CCD/CCX, (presumably) stronger INT focus, and of course much higher turbo clocks - are good for gaming perf, which is what desktop is largely about.
For desktop, I expect Zen6 to be a bigger absolute upgrade over Zen5 than Zen4 was over Zen3, because of the reasons above.

While I agree with your points, a bigger increase than Zen3 to 4 would be a very positive surprise IMHO. While back at the time most people were a bit disappointed of the IPC gains, together with the significant clock gains it was no slouch at all - more the biggest increase in the whole Zen era until now wrt to ST performance.

 

[StefanR5R]

literallly takes hours to run

Yes, it's 10 individual benchmark programs in the Integer suite, and 13 more in Floating Point after all.

Benchmark durations at this order are exactly what one wants if one is interested in steady-state performance of an entire machine. But if one is looking for transient burst performance, SPEC CPU is obviously not the right tool. Or if one happens to benchmark an algorithm + dataset for which one knows (or can monitor) when the relevant computer subsystems reach steady state in this particular workload, then of course it is possible to save on time spent on benchmark runs.

 

[Thunder 57]

Maybe, but:

• Zen5 used the same IOD as Zen4, Zen6 gets a new one (better IF and better mem controller are safe bets)
• Zen5 used only a slightly improved node vs. Zen4 (Zen6 uses a much faster node)
• Zen5 had no L2/L3/CCX improvements worth mentioning (Zen2 doubled L3/CCX, Zen3 doubled CCX, Zen4 doubled L2), while desktop Zen6 increases L3 per CCD/CCX.
  Last time the CCX/max. amount of accessible L3 per core/thread grew in desktop (Zen3), games got an above-average IPC benefit out of it.
• Zen5 invested a lot of the added transistors into full-rate AVX512 with doubled FP_PRF, while most desktop workloads are rather Integer-bound and INT_PRF only grew by a measly 16 entries from 224 to 240, still smaller than Sunny Cove's 280 (Ice Lake from 2019) and one major bottleneck now, according to something adroc mentioned before.
  So a pretty low-hanging fruit to pick with Zen6 (anything less than growing IntPRF to at least 288 would be disappointing; best-case would be 336, matching int PRF/ALU ratio of Zen4).

Anyway, all the biggest deals of Zen6 - faster memory support, bigger CCD/CCX, (presumably) stronger INT focus, and of course much higher turbo clocks - are good for gaming perf, which is what desktop is largely about.
For desktop, I expect Zen6 to be a bigger absolute upgrade over Zen5 than Zen4 was over Zen3, because of the reasons above.

What's your guess as to "much higher turbo clocks" in terms of numbers?

 

[adroc_thurston]

What's your guess as to "much higher turbo clocks" in terms of numbers?

Two shrinks for 15% speed is normal.

 

[reaperrr3]

What's your guess as to "much higher turbo clocks" in terms of numbers?

~800 MHz.
I don't believe in 7 Ghz yet, but 6.5 for 1-2 cores on the top bin SKUs sounds realistic enough with N2P vs. N4P plus some expectable clockability optimizations of Zen6 over Zen5.
Maybe another 100-200 MHz if early yield is good enough.

I could imagine them maybe going to 7 GHz if the CCDs were still 8 cores.
But with 12 cores per CCD, I don't know how much all-core clock AMD is willing to sacrifice just to clock 1 or 2 cores slightly higher, just to do slightly better in some synthetic benchmarks or ancient single-threaded games.
And the ST clocks achievable by the 24C top model within TDP are kinda the ceiling of the line-up for marketing reasons, so...

 

[Hulk]

33% more cores = 70% better MT performance means that 27.5% contains clock gain and IPC.

My guess:
10% IPC 16% clock bump or 12% IPC and 14% clock bump

Desktop Zen6:
1.5*1.1*1.16= 1.914

So >90% gain in multi-core over 9950X.

Actually it's 50% more cores.
9950X has 16 cores, 50% of that is 8. 18+8+24

33.3% would be if cutting core count from 24 to 16. Then it would be 33.3% less cores.

I do that calc wrong sometimes as well when I'm not thinking am I looking at increase or decrease!

 

[Josh128]

Maybe, but:

• Zen5 used the same IOD as Zen4, Zen6 gets a new one (better IF and better mem controller are safe bets)
• Zen5 used only a slightly improved node vs. Zen4 (Zen6 uses a much faster node)
• Zen5 had no L2/L3/CCX improvements worth mentioning (Zen2 doubled L3/CCX, Zen3 doubled CCX, Zen4 doubled L2), while desktop Zen6 increases L3 per CCD/CCX.
  Last time the CCX/max. amount of accessible L3 per core/thread grew in desktop (Zen3), games got an above-average IPC benefit out of it.
• Zen5 invested a lot of the added transistors into full-rate AVX512 with doubled FP_PRF, while most desktop workloads are rather Integer-bound and INT_PRF only grew by a measly 16 entries from 224 to 240, still smaller than Sunny Cove's 280 (Ice Lake from 2019) and one major bottleneck now, according to something adroc mentioned before.
  So a pretty low-hanging fruit to pick with Zen6 (anything less than growing IntPRF to at least 288 would be disappointing; best-case would be 336, matching int PRF/ALU ratio of Zen4).

Anyway, all the biggest deals of Zen6 - faster memory support, bigger CCD/CCX, (presumably) stronger INT focus, and of course much higher turbo clocks - are good for gaming perf, which is what desktop is largely about.
For desktop, I expect Zen6 to be a bigger absolute upgrade over Zen5 than Zen4 was over Zen3, because of the reasons above.

Zen 5 had a 50% increased L1D cache size and double the available bandwidth to L1 and FPU, which is not nothing.

My previous prediction is that in nT, this should be even better (due to core count increase) to the ~50% increase Zen 4 got over Zen 3. Power limits and heat dissipation would be the only things that could potentially be a limiting factor there.

In 1T, Im afraid the only way we'll come close to the +29% that Zen 4 got over Zen 3 is if we get the rumored exotic boost frequency increases. If we get an average of +10% IPC like the leaked slides say, that plus +14% boost clock increase (6.5 vs 5.7) would put it at +24% 1T overall uplift, which is 5% short of what Zen 4 got over Zen 3 on average.

 

[adroc_thurston]

rumored exotic boost frequency increases

It's 15%.
Spread over 2 shrinks. Far less exotic than Glymur&N3X friends.

 

[511]

Two shrinks for 15% speed is normal.

15% fmax is too much for two shrinks that are not Fmax Focused or should I say 1.5 Shrinks.o would say 10-12% is reasonable to

 

[adroc_thurston]

15% fmax is too much for two shrinks that are not Fmax Focused

TSM nodes are inherently fmax-focused, especially in the finflex era.

would say 10-12% is reasonable to

Maybe for Intel.
Other vendors have no such skill issues.
QCOM got 19% fmax off N3X plus some Vmax cranking.