Question Zen 6 Speculation Thread

[adroc_thurston]

It’s pretty good at FP64 too

yeah le DGEMM is the only argument I have for GEMM blobs in client outside of .gb6 pumping but who the hell runs DGEMM in client?

 

[Win2012R2]

we don't care about Intel because Intel sucks.

They are certainly benchmark for AMX execution and will be compared against AMD.

So the question stands: if Intel uses dedicated per core AMX units then why the **** perf won't be terrible if single AMX unit gets shared by multiple cores, potentially high number like 8 or more.

 

[adroc_thurston]

They are certainly benchmark for AMX execution and will be compared against AMD.

no, the benchmark is Apple since they ship it everywhere.

So the question stands: if Intel uses dedicated per core AMX units then why the **** perf won't be terrible if single AMX unit gets shared by multiple cores, potentially high number like 8 or more.

a) you know you can build a bigger matmul unit for 8 cores. Apple one is p chungus and has a ton of juice. very nice for DGEMM.
b) no, seriously, ARM SME works. read the docs and look the benchmarks.

 

[adroc_thurston]

Man, I wish talking to some people here wasn't the equivalent of trying to potty train a dumb toddler. embarassing.

 

[poke01]

yeah le DGEMM is the only argument I have for GEMM blobs in client outside of .gb6 pumping but who the hell runs DGEMM in client?

no one. But openBLAS did add support for SGEMM for SME targets which is good to see.

 

[Win2012R2]

you know you can build a bigger matmul unit for 8 cores

So that's equivalent of multiple AMX units then, reduces shared contention, but it's still a shared resource that run slower for when all cores need it.

no, the benchmark is Apple since they ship it everywhere.

Not in x86 market.

 

[adroc_thurston]

no one. But openBLAS did add support for SGEMM for SME targets which is good to see.

yup.
Apple Pro/Max/Ultra sticks were kind of funny wrt ad-hoc DGEMM crunch.

So that's equivalent of multiple AMX units then

NO.

reduces shared contention

NO.
The unit is heavily multithreaded and has its own ld/st pipeline in SME case.

 

[Win2012R2]

The unit is heavily multithreaded and has its own ld/st pipeline in SME case.

Will classic Zen cores have it per cluster (12 cores?) too?

 

[adroc_thurston]

Will classic Zen cores have it per cluster (12 cores?) too?

yes? Again, feature support is fully flat across the uarch family.

 

[Joe NYC]

They are certainly benchmark for AMX execution and will be compared against AMD.

So the question stands: if Intel uses dedicated per core AMX units then why the **** perf won't be terrible if single AMX unit gets shared by multiple cores, potentially high number like 8 or more.

I bet that when Intel implements ACE, AMX will be depreciated into legacy.

Then the question is if ACE makes it to Diamond Rapids or Coral Rapids.

On AMD side, it will be in Zen 7, which should launch between Diamond and Coral.

 

[adroc_thurston]

Then the question is if ACE makes it to Diamond Rapids or Coral Rapids.

Coral. DMR is AMX.

On AMD side, it will be in Zen 7, which should launch between Diamond and Coral.

Coral is trying to intercept Florence.

 

[Win2012R2]

yes? Again, feature support is fully flat across the uarch family.

Ok, shared unit too, so what will run faster at the same frequency -
1) 12 Intel cores with AMX as is now
or
2) 12 AMD cores with per cluster AMX

 

[Joe NYC]

Coral is trying to intercept Florence.

We will see if it succeeds.

It should also be a competition, by proxy, between TSMC A14 and Intel 14A, in time to HVM.

 

[adroc_thurston]

Ok, shared unit too, so what will run faster at the same frequency -
1) 12 Intel cores with AMX as is now
or
2) 12 AMD cores with per cluster AMX

The former.
But why would you want that? If you wanna do Actual Real Matmul at AMD, GPU is your friend.

It should also be a competition, by proxy, between TSMC A14 and Intel 14A, in time to HVM.

man that's an optimistic view of 14A lmao

 

[Joe NYC]

man that's an optimistic view of 14A lmao

Then, if TSMC A14 beats 14A to HVM, it would imply that Florence beats Coral to market, assuming Florence is one of the first products on A14

 

[poke01]

will Intel implement amx per core in client?

 

[adroc_thurston]

Then, if TSMC A14 beats 14A to HVM, it would imply that Florence beats Coral to market, assuming Florence is one of the first products on A14

well Venice beats DMR to market and clowns all over it so...

will Intel implement amx per core in client?

no lmao, their core area is bad enough as is.

 

[Win2012R2]

The former.

Ok, good, dedicated per core units will run faster all other things equal.

Now how much faster would Intel be with 12 cores scenario in this case vs AMD - like 2 times faster, or perhaps 12?

But why would you want that? If you wanna do Actual Real Matmul at AMD, GPU is your friend.

Indeed, why implement this in CPU in the first place, Intel had to do it because of lack of GPU, having integrated GPU seems far better way, maybe with higher latency than CPU.

 

[adroc_thurston]

dedicated per core units will run faster all other things equal.

It's not 'faster', you just dedicate more area to matmul.

Now how much faster would Intel be with 12 cores scenario in this case vs AMD - like 2 times faster, or perhaps 12?

idk.
How much faster is a single gfx1311 CU is over an AMX core in GEMM?

 

[Win2012R2]

How much faster is a single gfx1311 CU is over an AMX core in GEMM?

No idea, we are talking about CPUs here, you just don't want to answer obvious question that a shared single unit for 12 cores under heavy usage will be slower like more than 12 times than in scenario of 1 dedicated unit per core: ie dogsh*t slow like I said, which is obvious to anybody who actually profiles stuff and runs into contentions like this would be.

Solution is obviously running this stuff on GPUs and drop AMX crap.

It's not 'faster', you just dedicate more area to matmul.

Oh wow, 600 sq mm GPU is not faster than 200 sq mm, it's just got more area dedicated.

Or an 8 core CPU isn't faster than 1 core, it's just got more area dedicated to cores.

 

[adroc_thurston]

No idea, we are talking about CPUs here, you just don't want to answer obvious question that a shared single unit for 12 cores under heavy usage will be slower like more than 12 times than in scenario of 1 dedicated unit per core: ie dogsh*t slow like I said, which is obvious to anybody who actually profiles stuff and runs into contentions like this would be.

shared units give you nice matmul rates are relatively minimal area investment.
Again, see Apple SGEMM/DGEMM benchmarks on M4-era stuff.

Solution is obviously running this stuff on GPUs and drop AMX crap.

No, cuz shared units are cheap and nice nuff.

 

[adroc_thurston]

Oh wow, 600 sq mm GPU is not faster than 200 sq mm, it's just got more area dedicated.

not my point.
they're both 8c CPUs here.
Obviously you know that, but you gotta be obtuse.

 

[Win2012R2]

shared units give you nice matmul rates are relatively minimal area investment.

Yes sure, nice for you to acknowledge it is shared unit, thus leading to contention = reduction in speed.

How relevant will it be for AMX that almost nobody is using, well, I reckon THOSE few people who actually use it will want proper perf from all cores, shared contention will result in non-deterministic latencies and lower throughput, so it seems dumb to support this very niche thing in the first place, especially in client and especially when you sell GPUs.

 

[poke01]

why implement this in CPU in the first place, Intel had to do it because of lack of GPU, having integrated GPU seems far better way, maybe with higher latency than CPU.

Because the cpu is standardised unlike GPU or NPU. It’s easy for devs too

 

[adroc_thurston]

thus leading to contention = reduction in speed.

It's not 'reduction in speed', you're just doing less matmul ops a clock.
That's like saying half-size WMMA in gfx13 is doing 'contetion' vs gfx1250.

Because the cpu is standardised unlike GPU or NPU. It’s easy for devs too

bingo.