Discussion Google Stadia CPU - what could it possibly be?

[tamz_msc]

2.7 GHz, multi-core, hyperthreaded , AVX2, 9.5 MB L2+L3.

People are saying it's Zen2, but I think it's Intel.

 

[Hitman928]

The google rep said it was a custom x86 CPU (probably not very custom but not off the shelf) and only announced AMD as a partner and not intel which leads me to think it's Zen 2. It could still be intel, but given the above, I'm thinking they stuck with AMD for both GPU and CPU.

 

[Dayman1225]

Lisa’s Twitter specifically says “GPU partner”

https://twitter.com/x/status/1108072456755908608

So does their Investor Relation post
http://ir.amd.com/news-releases/new...pus-and-developer-tools-tapped-new-generation

No mention of AMD supplying the CPU...

 

[tamz_msc]

The L2+L3 configuration is unlike anything that Zen has.

 

[Dayman1225]

The L2+L3 configuration is unlike anything that Zen has.

4/8 Skylake SP CPU with 1MB L2 and 1.375 L3 per core = 9.5MB

 

[amd6502]

I sure hope it's not hyperthreating. If I were interested in a cloud gaming box I'd feel much safer with SMT.

Fwiw, I don't think this Stadia is a good idea. But google will probably come with some interesting add ons that are big improvements over their androde VR/3d tech.

 

[SarahKerrigan]

I sure hope it's not hyperthreating. If I were interested in a cloud gaming box I'd feel much safer with SMT.

Fwiw, I don't think this Stadia is a good idea. But google will probably come with some interesting add ons that are big improvements over their androde VR/3d tech.

HT is SMT. (Except where it's CGMT/SoEMT.)

 

[krumme]

This is from the AT article

(When Google says custom CPU and custom GPU - this could be early hardware of AMD's upcoming generations of technology, put into a custom core configuration / TDP. We're likely looking at a Zen 2 based CPU, based on AVX2 support listed, and a Radeon Instinct based GPU with tweaked settings specifically for Google.)

??

 

[jpiniero]

Knowing Google, it'll be cancelled in 6 months.

 

[coercitiv]

Knowing Google, it'll be cancelled in 6 months.

I doubt they have the latency to cancel in 6 months.

 

[HurleyBird]

Reading between the lines, it looks to me like a Vega 56 with 8 GB of HBM2 + a Skylake-X derivative with 8 GB of DDR4.

 

[Asterox]

Reading between the lines, it looks to me like a Vega 56 with 8 GB of HBM2 + a Skylake-X derivative with 8 GB of DDR4.

If Google using "defult Skylake Xeon server CPU", as you see there is no defoult Skylake Xeon with 2.7ghz stock CPU speed.There is one 4/8 model with 2.6ghz CPU, but what would you get with 100mhz speed bump?

https://en.wikipedia.org/wiki/List_of_Intel_Xeon_microprocessors#Skylake-based_Xeons

But there is Coffe lake based 6/12 Xeon with 2.7ghz cpu.

https://en.wikipedia.org/wiki/List_of_Intel_Xeon_microprocessors#Coffee_Lake-based_Xeons

Also Epyc Embedded series is available, or 3151 model 4/8 with 2.7 ghz cpu.

https://www.amd.com/en/products/specifications/embedded/8161

Maybe CPU L2+L3 configuration is wrong or incorrectly presented.

 

[jpiniero]

If it's Intel, I would think it would be a 4S or 8S Skylake-SP. The XP 8168 has a base clock of 2.7 with 24 cores and Google would disable AVX-512. Stick 6 GPUs in there per processor.

 

[HurleyBird]

The cache (and "Hyperthreading") really suggests an Intel processor, and the fact that all press stuff so far lists AMD as only a "graphics" partner pretty much cements it. That there is no commercial offering with the same clocks is irrelevant since the CPU is listed as a custom solution.

The lack of PR from Intel is interesting though. If I had to hazard a wild guess, it's that Google already intends to switch over to something Zen 2 based in the future, and sooner than later, so that Intel and Google making a big hoopla about what is destined to be an extremely short lived partnership is likely to end up somewhat awkward for the former in the near future, hence the lack of said hoopla. That might not be it, but it would make sense.

 

[jpiniero]

Or Google is repurposing the Skylake-SP chips they have for Stadia because of Smeltdown.

 

[piesquared]

Lisa’s Twitter specifically says “GPU partner”

https://twitter.com/x/status/1108072456755908608

So does their Investor Relation post
http://ir.amd.com/news-releases/new...pus-and-developer-tools-tapped-new-generation

No mention of AMD supplying the CPU...

Maybe that's because they aren't ready to release specifications on the Rome architecture. The EPYC architecture and EPYC platform is a no brainer, it makes no sense to adopt an inferior platform from intel. The only other option than EPYC IMO is in house ARM design.

 

[tamz_msc]

4/8 Skylake SP CPU with 1MB L2 and 1.375 L3 per core = 9.5MB

Seems plausible. I wonder if we'll get to know exactly what it is by going into the game's settings menu and running a benchmark, like the way it was discovered that GeForce Now uses Tesla GPUs.

 

[Dayman1225]

Maybe that's because they aren't ready to release specifications on the Rome architecture. The EPYC architecture and EPYC platform is a no brainer, it makes no sense to adopt an inferior platform from intel. The only other option than EPYC IMO is in house ARM design.

They specifically said Custom X86 so ARM is 0 chance, and this is definitely not Rome. It’s either Naples or Skylake SP and my bets are on Skylake

 

[HurleyBird]

Maybe that's because they aren't ready to release specifications on the Rome architecture.

Announcing AMD as the processor partner hardly requires releasing Rome specifications.

 

[Hans de Vries]

Stadia clearly looks to be a custom single die AMD GPU with the x86 CPU on chip to run the games autonomously on the GPU.


STADIA: ONE SINGLE DIE
GPU: 56 out of 64 CU used.
CPU: 7(?) out of 8 cores used.
MEM: 16GB of HBM2 (two stacks) at 484 GB/s

 

[amd6502]

Stadia clearly looks to be a custom single die AMD GPU with the x86 CPU on chip to run the games autonomously on the GPU.


STADIA: ONE SINGLE DIE
GPU: 56 out of 64 CU used.
CPU: 7(?) out of 8 cores used.
MEM: 16GB of HBM2 (two stacks) at 484 GB/s

If it's monolithic then it's a very good guess.

I think it's a little more likely to be MCM because the 64CU (minus 8CU )GPU is so massive and on 12nm would be almost 400mm2. So they use 7nm GPU chiplet which is not custom (used in other product lines) to lower otherwise very high development cost.

I agree with folks it could either be intel CPU solution or maybe just Pinnacle-L, a one CCX version of Pinnacles, with half a MB of L3 or L2 unused maybe for use as a buffer that both GPU and CPU can quickly share data with.

 

[JasonLD]

Stadia clearly looks to be a custom single die AMD GPU with the x86 CPU on chip to run the games autonomously on the GPU.


STADIA: ONE SINGLE DIE
GPU: 56 out of 64 CU used.
CPU: 7(?) out of 8 cores used.
MEM: 16GB of HBM2 (two stacks) at 484 GB/s

"Lisa and her team designed a chip to bring you more than 10 teraflops of power which was coupled with custom CPU to make up a single Stadia instance"

That is not the indication that the GPU+CPU is on single die.

 

[Hans de Vries]

If it's monolithic then it's a very good guess.

I think it's a little more likely to be MCM because the 64CU (minus 8CU )GPU is so massive and on 12nm would be almost 400mm2. So they use 7nm GPU chiplet which is not custom (used in other product lines) to lower otherwise very high development cost.

I agree with folks it could either be intel CPU solution or maybe just Pinnacle-L, a one CCX version of Pinnacles, with half a MB of L3 or L2 unused maybe for use as a buffer that both GPU and CPU can quickly share data with.

How would an Intel CPU use the HBM2 memory attached to the GPU as the main memory?
~600 mm2 on 14nm would not be to big (although an MCM is not necessary impossible)
The CPU clock frequency of 2.7 GHz and the smaller L3 cache (6MB?) point to an integrated CPU (Zen1)

 

[JasonLD]

How would an Intel CPU use the HBM2 memory attached to the GPU as the main memory?
~600 mm2 on 14nm would not be to big (although an MCM is not necessary impossible)
The CPU clock frequency of 2.7 GHz and the smaller L3 cache (6MB?) point to an integrated CPU (Zen1)

16GB of "Total Memory"
"up to" 484GB/s transfer speed

It could be that it has 8GB of main memory and 8GB of HBM2.

 

[HurleyBird]

How would an Intel CPU use the HBM2 memory attached to the GPU as the main memory?

I don't think I've seen that anywhere. I've seen that the GPU uses HBM2, and that there's a total of 16 GB of RAM, and that each node has up to 484GB/s transfer speed. All these things can be true even without a unified pool of HBM 2.

So we can have 8GB of HBM2, and 8GB of DDR4. And there is up to 484GB/s if you're looking at the HBM2 block and not the DDR4 block. The usage of "up to" is actually pretty odd when listing bandwidth, but makes sense when you look at each node as having two separate memory pools. Not to mention, why list HBM2 under GPU header when there's a separate Memory section? That seems pretty telling. The only thing I can think of is because that HBM2 pool is only available to the GPU. It looks like a bit of marketing speak to make the thing look a bit more impressive than it really is.