Question Zen 6 Speculation Thread

[Kryohi]

Zen 5's 139@5.7GHz translates into 174 if you give it +10% IPC in this application and 6.5GHz, which I think is already pushing realistic expectations as to what Zen 6 can deliver.

I don't think a +13-15% IPC improvement instead of 10% is unrealistic though, that's actually more likely than 6.5GHz fmax imho.

 

[Det0x]

Zen5 is not 139 points @ 5.7ghz anyway (atleast not my systems)

 

[511]

Zen5 is not 139 points @ 5.7ghz anyway (atleast not my systems)

that's what available online

 

[Josh128]

Zen5 is not 139 points @ 5.7ghz anyway (atleast not my systems)

Its close enough. I think my 9900X does 137 IIRC. Its not cracking 140 without some help.

 

[Meteor Late]

Cinebench 2024:
M4 Max 177 1t
285K 145 1t
9950X 139 1t

Apple is currently 27% ahead in CB 2024 ST vs Zen 5. Gap will probably close quite a bit with Zen 6 vs M5, as Zen 6 will have full node advantage, but revert back to the ~25% or so advantage for M6 if that uses N2 or A16.

Well the cadence of Apple and AMD is quite a bit different, Apple is around 1 year cadence, at least with the base M and A chips, while AMD is on a 2-year or close to 2 year cadence. That's why I compare M6 with Zen 6, because of that cadence.

 

[Meteor Late]

I prefer to compare laptop chips, though, it's where one gets the closest match with memory speeds and timings, with LPDDR comparisons, as DDR is just much faster in terms of timings and latency. M4 Pro vs Strix Halo for example.

 

[Det0x]

Zen5 is not 139 points @ 5.7ghz anyway (atleast not my systems)

Its close enough. I think my 9900X does 137 IIRC. Its not cracking 140 without some help.

I just tested on my system, first ran on v-cache CCD then on regular CCD @ 5.7ghz static

X3D Zen5 core @ 5.7ghz = 147 points in R24 ST


Vanilla Zen5 core @ 5.7ghz = 144 points in R24 ST

 

[poke01]

I just tested on my system, first ran on v-cache CCD then on regular CCD @ 5.7ghz static

X3D Zen5 core @ 5.7ghz = 147 points in R24 ST
View attachment 132355

Vanilla Zen5 core @ 5.7ghz = 144 points in R24 ST
View attachment 132356

Looks like Zen6 will do more than fine. It just needs to clock above 6GHz

 

[Josh128]

I just tested on my system, first ran on v-cache CCD then on regular CCD @ 5.7ghz static

X3D Zen5 core @ 5.7ghz = 147 points in R24 ST
View attachment 132355

Vanilla Zen5 core @ 5.7ghz = 144 points in R24 ST
View attachment 132356

Thats not stock though. Stock chips dont hit 5.7 on the X3D die and dont even always pin max boost on vanilla die on heavy 1t loads. And you have wildly OC'ed memory, lol. The 139 is what you generally can expect with normal setup and EXPO 6000 CL30'ish memory. But I digress, its still just a 3.6% difference.

 

[eek2121]

Thats not stock though. Stock chips dont hit 5.7 on the X3D die and dont even always pin max boost on vanilla die on heavy 1t loads. And you have wildly OC'ed memory, lol. The 139 is what you generally can expect with normal setup and EXPO 6000 CL30'ish memory. But I digress, its still just a 3.6% difference.

Depends on the chip. Some are binned better than others.

 

[Det0x]

Thats not stock though. Stock chips dont hit 5.7 on the X3D die and dont even always pin max boost on vanilla die on heavy 1t loads. And you have wildly OC'ed memory, lol. The 139 is what you generally can expect with normal setup and EXPO 6000 CL30'ish memory. But I digress, its still just a 3.6% difference.

Yes i agree, my memory setup is heavily overclocked.
But i also expect Zen6 with its silicon briges to mop the floor with my SerDES connection, even if i'm running at 8800/2200
(Cinebench R24 actually scales with memory performance, atleast in MT)

So for a hypertical Zen6 X3D core @ 6.5ghz with +10% higher IPC, you would be able to reach ~184points in R24 ST
"Thats only 22% faster" than my maxed out X3D core. (~152 points)

Allthought i think it's more realistic to expect something like ~175 points for the "average user"

 

[Josh128]

Yes i agree, my memory setup is heavily overclocked.
But i also expect Zen6 with its silicon briges to mop the floor with my SerDES connection, even if i'm running at 8800/2200
(Cinebench R24 actually scales with memory performance, atleast in MT)

So for a hypertical Zen6 X3D core @ 6.5ghz with +10% higher IPC, you would be able to reach ~184points in R24 ST
"Thats only 22% faster" than my maxed out X3D core. (~152 points)

Allthought i think it's more realistic to expect something like ~175 points for the "average user"

Im all about the average user. I extrapolated ~174, so we are thinking along the same lines.

I have no doubt you'll be putting up some impressive scores with Zen 6, I look forward to the day when I can sit down, sip on my coffee and ignorantly try to extrapolate other workload performance from them, lol.

 

[CouncilorIrissa]

Yes i agree, my memory setup is heavily overclocked.
But i also expect Zen6 with its silicon briges to mop the floor with my SerDES connection, even if i'm running at 8800/2200
(Cinebench R24 actually scales with memory performance, atleast in MT)

So for a hypertical Zen6 X3D core @ 6.5ghz with +10% higher IPC, you would be able to reach ~184points in R24 ST
"Thats only 22% faster" than my maxed out X3D core. (~152 points)

Allthought i think it's more realistic to expect something like ~175 points for the "average user"

That's assuming AMD will give you more than 32B/clk read with Zen 6. They didn't with STXH. And I don't think latency improvement is on the cards either.

 

[Josh128]

That's assuming AMD will give you more than 32B/clk read with Zen 6.

Its not assuming anything different than Zen 5 perf/GHz *6.5 *1.1.

What are you talking about?

 

[adroc_thurston]

That's assuming AMD will give you more than 32B/clk read with Zen 6

well you gotta feed SP8 MRDIMMs somehow.

They didn't with STXH.

They wanted to keep general CCD layout intact there.

And I don't think latency improvement is on the cards either.

Yup, you'll get moar b/w at a lot lower power.

 

[inquiss]

Yup, you'll get moar b/w at a lot lower power.

Which means more of the power budget for the cores themselves...

 

[adroc_thurston]

Which means more of the power budget for the cores themselves...

D'oh.
That also means moar b/w per CCD, since, well, we're back to 8 CCD everywhere.

 

[CouncilorIrissa]

well you gotta feed SP8 MRDIMMs somehow.

Oh they absolutely will give server all the bandwidth in the world. It's the client interconnect that I'm doubtful about, I wouldn't put it past AMD to cuck client in a funny way.

Its not assuming anything different than Zen 5 perf/GHz *6.5 *1.1.

What are you talking about?

IOD <-> CCD read throughput.
CBR24 is a memory bandwidth-heavy benchmark unlike its predecessors. You can gain like 10% in SPEC CPU integer rate-1, but much less in CBR24 because you didn't change anything in the memory bandwidth department.

 

[fastandfurious6]

Oh they absolutely will give server all the bandwidth in the world. It's the client interconnect that I'm doubtful about, I wouldn't put it past AMD to cuck client in a funny way.

IOD <-> CCD read throughput.
CBR24 is a memory bandwidth-heavy benchmark unlike its predecessors. You can gain like 10% in SPEC CPU integer rate-1, but much less in CBR24 because you didn't change anything in the memory bandwidth department.

AMD can't mess around anymore, market gonna get even more competitive

even if Intel is wailing they'll keep spending extra $$$$$ to get gains

 

[OneEng2]

Apple's "tablet chips" (either the base M in iPad Pro or an iPhone SoC in the rest) aren't all that large so the yield benefit would be marginal. Thanks to their prepayments to TSMC Apple is not capacity constrained so there's little reason to change their monolithic ways on those smaller chips.

How large are they? Zen 5 is 70.6mm2. AI says M4 is 166mm2.

As for the CB24, I wonder how much of this benchmark is simply a bandwidth test? ARL does well in the benchmark despite its notoriously high latency.

Comparing real world MT work, ARL gets beaten badly by Zen 5.

So how representative is CB24 really?

 

[CouncilorIrissa]

How large are they? Zen 5 is 70.6mm2. AI says M4 is 166mm2.

Can't believe this needs to be stated, but why on earth are you comparing CCD area (meaning, only cores + IFOP) and the whole SoC, that has the CPU cores, the NPU, memory controllers and a relatively big iGPU? This is not an apples-to-apples comparison.

 

[okoroezenwa]

Can't believe this needs to be stated, but why on earth are you comparing CCD area (meaning, only cores + IFOP) and the whole SoC, that has the CPU cores, the NPU, memory controllers and a relatively big iGPU? This is not an apples-to-apples comparison.

Oh OneEng2 definitely knows better.