Discussion Zen 5 Speculation (EPYC Turin and Strix Point/Granite Ridge - Ryzen 9000)

[DisEnchantment]

Speculate at will

 

[techjunkie123]

Overclocked 9850X3D geekbench 6 run for those interested
View attachment 137592

ASUS System Product Name - Geekbench

Benchmark results for an ASUS System Product Name with an AMD Ryzen 7 9850X3D processor.

browser.geekbench.com

Getting close to M4. Zen 6 could catch up to M5 in ST on desktop. If zen 6 laptop parts can hit ~3800 at least on N2 that would be amazing.

 

[Joe NYC]

Good interview

BTW, Papermaster confirmed ACE for Zen 7.

 

[LightningZ71]

I don’t want an ASUS tuf. What now?

why are they porting RDNA3.5 to N3 then?

That's my question. I suspect that RDNA 5 is a notable XTOR count increase per WGP, and a much smaller RDNA 3.5 iGPU is still more than capable of being effective in poverty markets. The 8CU in KRK does a good job keeping up with the 12CU unit in HWK on actual game performance. We see that with FSR3.x, it's enough for most games to drive a 1080p display well enough to be usable. Even the 4CU unit in KRK2 isn't awful.

 

[poke01]

a much smaller RDNA 3.5 iGPU is still more than capable of being effective in poverty markets.

Why would poverty markets need N3? They could keep on selling N4 RDNA3.5 in those areas.

You know people don’t have to justify whatever AMD does as sane. If Zen5 isn’t good to port over to N3 then RDNA3.5 isn’t either. A waste of time.

 

[LightningZ71]

They aren't going to stop producing KRK and KR2 any time soon. Those are going to have a market they they will sell into for a long time. They still produce HWK in volume and there are still PHX2 products available for sale, though I don't shop in that market so I don't see many.

WCL is not a bad product, and AMD will need something for that and 4Xe PTL going forward. Low end Medusa will handle that well enough.

 

[adroc_thurston]

why are they porting RDNA3.5 to N3 then?

Cheapeast IP they have.

If Zen5 isn’t good to port over to N3 then RDNA3.5 isn’t either. A waste of time.

CPUs matter. GPUs do not.

 

[Joe NYC]

Why would poverty markets need N3? They could keep on selling N4 RDNA3.5 in those areas.

You know people don’t have to justify whatever AMD does as sane. If Zen5 isn’t good to port over to N3 then RDNA3.5 isn’t either. A waste of time.

Since Medusa 1 die is N3P, then AMD is obviously porting RDNA3.5 to N3P.

Interesting question is which process technology AMD will use for Zen 6 desktop IO dies.

If it is N6, we may still be stuck on RDNA2. More likely, it will be N4 (N4C?) and we end up with RDNA3.5 and AMD may add some other things that commercial OEMs ask for in desktop chips (NPU).

Maybe AMD should have been selling (and bundling) M2 sticks with NPU.

 

[Tachyonism]

Since Medusa 1 die is N3P, then AMD is obviously porting RDNA3.5 to N3P.

Is there any advantage to doing this? Why doesnt AMD just keeps RDNA3.5 on 4nm like Intel (18A for CPU and N3E for GPU).

 

[adroc_thurston]

Is there any advantage to doing this? Why doesnt AMD just keeps RDNA3.5 on 4nm like Intel (18A for CPU and N3E for GPU).

Adding packaging overhead to mainstream parts is idiotic.

 

[Det0x]

DDR5-4800 vs. DDR5-6000 Performance With The AMD Ryzen 7 9850X3D In 300+ Benchmarks​

DDR5-4800 vs. DDR5-6000 Performance With The AMD Ryzen 7 9850X3D In 300+ Benchmarks Review - Phoronix

www.phoronix.com

 

[Joe NYC]

DDR5-4800 vs. DDR5-6000 Performance With The AMD Ryzen 7 9850X3D In 300+ Benchmarks​

View attachment 137609

DDR5-4800 vs. DDR5-6000 Performance With The AMD Ryzen 7 9850X3D In 300+ Benchmarks Review - Phoronix

www.phoronix.com

Well, full Zen 5 using DDR5 6000 gets bandwidth of 96 GB/s and single CCD can only use 64 GB/s write and 32 GB/s write.

Reducing memory speed to DDR5 4800 reduces bandwidth to 76.8 GB/s which is still more than 64 GB/s a single CCD can use.

So, the bandwidth is unaffected only memory latency is affected.

Probably other things are going on, such as DDR 6000 being synched to the Infinity Fabric clock and if DDR 4800 also has the clock synched, it would mean lower bandwidth to CCD and if it is not synched, then some other losses take place...

 

[Joe NYC]

So, Zen 5 can hold its on in the CPU core area, even on N4, but the uncore stinks.

Which is the reason AMD should be iterating the uncore annually until they catch up.

https://twitter.com/x/status/2017390286666928298

 

[adroc_thurston]

Which is the reason AMD should be iterating the uncore annually until they catch up.

It's an 8Q cadence with client SoC uncore being the lowest priority (GPGPU > server > GPU > client SoC).

 

[Joe NYC]

Battle of the 128 core CPUs. Granite Rapids vs. Turin full cores. I thought Granite Rapids was closer, in performance:

AMD EPYC 9755 Delivers Decisive Performance Leadership Over Xeon 6 Granite Rapids With Nearly 500 Benchmarks Review - Phoronix

www.phoronix.com

 

[adroc_thurston]

Battle of the 128 core CPUs. Granite Rapids vs. Turin full cores. I thought Granite Rapids was closer, in performance:

AMD EPYC 9755 Delivers Decisive Performance Leadership Over Xeon 6 Granite Rapids With Nearly 500 Benchmarks Review - Phoronix

www.phoronix.com

View attachment 137658

It is closer in SIR2017.
Horrible L3 perf cripples it in real workloads.

 

[adroc_thurston]

Mind you, this is babymode.
Venice is the first Actually Expensive EPYC and it'll show you how winning™ looks like for realz.

 

[branch_suggestion]

Mind you, this is babymode.
Venice is the first Actually Expensive EPYC and it'll show you how winning™ looks like for realz.

Yep, Intel no longer has a high speed mem advantage as that is equalised.
All that leaves them with is AMX, which is neutralised by ACE in Florence.
Gonna hit like a cement mixer.

 

[adroc_thurston]

Yep, Intel no longer has a high speed mem advantage as that is equalised.

Mem speed is whatever (just don't triple stack DDR PHYs and you'll hit relevant ones).
I mean node and packaging and cache ratios and the uncore.
Venice is really $$$ but man does it win some.

 

[Josh128]

So, Zen 5 can hold its on in the CPU core area, even on N4, but the uncore stinks.

Which is the reason AMD should be iterating the uncore annually until they catch up.

https://twitter.com/x/status/2017390286666928298

What is H445 and H350??

 

[Joe NYC]

It's an 8Q cadence with client SoC uncore being the lowest priority (GPGPU > server > GPU > client SoC).

Well, the results show it. AMD can achieve some competitiveness, briefly and then, with no follow up, loses it.

At this pace and with this attitude, AMD will be #5 of the 5 notebook vendors. Behind:
- Intel
- Apple
- Qualcomm
- NVidia

They all invest more and they all take the market more seriously.

Is there any surprise then how the notebook OEMs went exactly as much on the limb for AMD as AMD goes out on the limb about the notebook market?

 

[Geddagod]

So, Zen 5 can hold its on in the CPU core area, even on N4, but the uncore stinks.

Which is the reason AMD should be iterating the uncore annually until they catch up.

https://twitter.com/x/status/2017390286666928298

AMD core power results are questionable.
I'm hoping geekerwan does some idle normalized core power curves. Though IIRC his spec 2017 numbers for AMD (perf, not power) were some straight buns.

What is H445 and H350??

AMD Zen 5 strix point and gorgan point skus.

 

[Joe NYC]

What is H445 and H350??

350 is Kracken and 445 is likely Gorgon

 

[branch_suggestion]

Mem speed is whatever (just don't triple stack DDR PHYs and you'll hit relevant ones).
I mean node and packaging and cache ratios and the uncore.
Venice is really $$$ but man does it win some.

Just saying as there were some benches that liked MRDIMMs.
EPYC with each generation has systematically reduced the number of workloads Intel is viable in each time, Venice makes this nearly absolute.

 

[branch_suggestion]

Well, the results show it. AMD can achieve some competitiveness, briefly and then, with no follow up, loses it.

At this pace and with this attitude, AMD will be #5 of the 5 notebook vendors. Behind:
- Intel
- Apple
- Qualcomm
- NVidia

They all invest more and they all take the market more seriously.

Is there any surprise then how the notebook OEMs went exactly as much on the limb for AMD as AMD goes out on the limb about the notebook market?

What matters is CAGR, make a big enough leap every 2 years and all is well.
New roadmap is alternating CPU and GPU updates, is quite nice really.

 

[Markfw]

Battle of the 128 core CPUs. Granite Rapids vs. Turin full cores. I thought Granite Rapids was closer, in performance:

AMD EPYC 9755 Delivers Decisive Performance Leadership Over Xeon 6 Granite Rapids With Nearly 500 Benchmarks Review - Phoronix

www.phoronix.com

View attachment 137658

Well, I have a pair of 9755's, just not in the same motherboard, and they rock ! And destroy everything else I have, and that's saying something.

edit: so bad in fact that my retail 9654 is powered down, as its not worth it for the first time in years. The 128 9755's do twice the work or more at only 25% more power, if that !