Question Zen 6 Speculation Thread

[fastandfurious6]

diminishing returns, overheating, limited support by games

 

[Tuna-Fish]

To be clear, it will still run 512-bit code, just at half rate. And using 512-bit registers is still a win, because it means frontend can do less work. Just like Zen4.

 

[Markfw]

To be clear, it will still run 512-bit code, just at half rate. And using 512-bit registers is still a win, because it means frontend can do less work. Just like Zen4.

I am not convinced they will go back, after the big success of Zen 5 in this area (yes, most people don't see it, but some of us do), especially in servers.

 

[igor_kavinski]

Yeah. They should mark the FP area as "Don't touch! Changes Forbidden". That's the cost of a future with ubiquitous 512-bit calculations.

 

[MS_AT]

"Don't touch! Changes Forbidden".

Improvements are also changes. Why would you like to deny us these?

That's the cost of a future with ubiquitous 512-bit calculations.

x64 needs ubiquitous AVX512 more than 512-bit calculations.

 

[igor_kavinski]

Improvements are also changes. Why would you like to deny us these?

Only for a few generations so that the developers get bored from lack of advances in FP performance and start optimizing their code to squeeze more performance out of the FP units

 

[Markfw]

Improvements are also changes. Why would you like to deny us these?

x64 needs ubiquitous AVX512 more than 512-bit calculations.

I have many Zen 4 computers, and servers, and also a few Zen5, including a Turin. The avx-512 performance is like 40% faster. So I would not want to lose that is what I am saying. I don't know the technical side, just how they perform.

 

[Kolifloro]

So the next 12 months ... we will have 'GORGON' ...

Then ... starting in the summer of 2026 ... 'Medusa Desktop'...

And 'Medusa Halo' will be among us from summer '27 ...

If you prefer 'Medusa Halo' plus X3D ... then probably by 2028 [you'll be able to get married, incept a baby (actually you'll have time to incept a couple of babies ... maybe even three babies) , become a father ... and finally get divorced ... ] ...

 

[basix]

I have many Zen 4 computers, and servers, and also a few Zen5, including a Turin. The avx-512 performance is like 40% faster. So I would not want to lose that is what I am saying. I don't know the technical side, just how they perform.

Halfed AVX512 is for power and cost limited scenarions. I do not see too much benefit of a full AVX512 implementation on consoles, because for many things you could just use the GPU (vector and matrix operations). Zen 4 alike AVX512 will still be a huge uplift against the stripped down FPUs of the PS5. In most gaming scenarios I hardly see, that halfed AVX512 (or lets say 1.4x less throughput) will limit gaming performance. The important thing here is: It runs AVX512 code, which helps regarding adoption rate.

For DT it is a different story. Here I would like to see full AVX512 width. Server is a no-brainer, there you will have full width anyways. Mobile could get full width as well, as long they are using the DT CCD. The rest will probably use a halfed AVX512 implementation (the smaller, monolithic mobile chips)

 

[Kaffeekenan]

Yeah.
Medusa DT is h2'26, everything else is CES'27.
Obviously, server is earlier than that.

Server is earlier than what? Earlier than CES or earlier than desktop?

 

[StefanR5R]

5-digit nomenclature, was sort unprecedented, when it happened with Intel CPUs.

5-digit names were quite common in the 1980s and 1990s.

 

[soresu]

5-digit names were quite common in the 1980s and 1990s.

80386 my beloved....

 

[soresu]

incept a baby

Incept?

Never heard incept used in place of conceive before, but it tees me up beautifully for dirty gif jokes with the provided context 😂😆

 

[jpiniero]

Halfed AVX512 is for power and cost limited scenarions. I do not see too much benefit of a full AVX512 implementation on consoles, because for many things you could just use the GPU (vector and matrix operations).

This is true but I would think Sony will use Zen 6c cores. Assuming they are still smallish I'm not sure it's worth the engineering effort to gut the FPU.

 

[Tuna-Fish]

This is true but I would think Sony will use Zen 6c cores. Assuming they are still smallish I'm not sure it's worth the engineering effort to gut the FPU.

There is no engineering effort to gut the FPU, AMD has already designed it to be configurable. You just pick whether you want 256-bit or 512-bit ALUs, and I don't see a world where the die area would be better spent on doubling the SIMD execution width, and not in the GPU.

And again, 256-bit execution width doesn't mean that the FPU is crippled. Zen4 and mobile Zen5 with half-width FPU are still really good! If the data path between L2 and L3 doubles in Zen6, I would bet that half-width Zen6 beats full-width Zen5 on a lot of real FPU-heavy loads.

 

[jpiniero]

There is no engineering effort to gut the FPU, AMD has already designed it to be configurable. You just pick whether you want 256-bit or 512-bit ALUs, and I don't see a world where the die area would be better spent on doubling the SIMD execution width, and not in the GPU.

The assumption is that they will continue to use 8 (c) cores, so the space savings is probably 2-3 mm2 at best.

Edit: For the standard PS6, I am thinking more like 8 Zen 6c cores plus 60-80 UDNA CUs on N3P. While Sony will try to market otherwise, IMO it's more like what the PS5 Pro should have been rather than an generational upgrade.

 

[basix]

I would assume 4+8 or so cores for the PS6. Those 8 "c" cores are fine for all PS5 Games and a few faster cores provide some extra Oomph for critical workload paths.

Some speculated silicon areas assuming TSMC N3P:
- 4 big cores ~3.5...4mm2 per core
- 8 "c" cores ~2.5mm2 per core
- All 12 cores ~35mm2
- Halfed FPU for all cores (256bit data path)
- Halfed L2$ (512kByte) ~0.5mm2 per Core (already included above)
- 12...24 MByte L3$ (24MByte would be ~10mm2 in N3P)

So 12 cores + 24 MByte cache could be roughly 45mm2 big. Maybe even a little smaller. This area is very similar to PS5's CPU. I think some additional big cores are worth it.

 

[adroc_thurston]

Halfed L2$ (512kByte) ~0.5mm2 per Core (already included above)

L2 is not configurable.

 

[basix]

Why should it not for a semi-custom chip?

But anyways, you could trade doubled L2$ for halfed L3$ and stay at roughly the same area.

 

[adroc_thurston]

Why should it not for a semi-custom chip?

Because L2 is physically not configurable.

 

[soresu]

Assuming they are still smallish I'm not sure it's worth the engineering effort to gut the FPU.

Given rumours put Zen6c server CCDs at 32C I'd say that they might be smoler than ever.

 

[soresu]

For the standard PS6, I am thinking more like 8 Zen 6c cores plus 60-80 UDNA CUs on N3P.

PS5 Pro already has 60 CUs.

UDNA would have to pull off some kind of magic to extract that much more perf from the same CU count in just a handful of years.

PS4 Pro and PS5 had same CU count, but PS5 had more than 2x the clock speed.

PS5 Pro -> PS6 won't have anywhere near that kind of clock scaling headroom.

My assumption would be more like 80-96 CUs for PS6.

If AMD can make SED chiplet gaming GPUs viable with minimal/zero software problems then they can maybe make one die that can work for their halo APU segment and Sony can use 2x for PS6.

 

[adroc_thurston]

UDNA would have to pull off some kind of magic to extract that much more perf from the same CU count in just a handful of years.

Yes the magic of gfx13. Next.

If AMD can make SED chiplet gaming GPUs viable with minimal/zero software problems then they can maybe make one die that can work for their halo APU segment and Sony can use 2x for PS6

Ah yes SoIC for consoles. Genius.

 

[jpiniero]

PS5 Pro already has 60 CUs.

UDNA would have to pull off some kind of magic to extract that much more perf from the same CU count in just a handful of years.

PS4 Pro and PS5 had same CU count, but PS5 had more than 2x the clock speed.

PS5 Pro -> PS6 won't have anywhere near that kind of clock scaling headroom.

Correct. That's why I was saying it would be more "What the PS5 Pro should have been" than a generational upgrade.

I am assuming the console's price will be decently more than $449 as it is. I don't think they want to go too crazy on pricing and they aren't doing that big of subsidies anymore.

A Pro console at launch is possible.

 

[Ghostsonplanets]

PS6 will be modestly priced. They won't go with over the top specs.

A Pro might boast these kind of specs. But the base PS6? No. Focus will be on more memory, faster RT and better CPU.