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

[DisEnchantment]

Speculate at will

 

[adroc_thurston]

I was under the impression that SH5 was physically different (and bigger) then SP5.

No.

So is then Venice SP5, SH5 or something else?

no gonna tell you bro

 

[Joe NYC]

No.

no gonna tell you bro

haha, I am going to guess it will be one of the following four: SP5, SP7, SH5, SH6

 

[adroc_thurston]

SH5, SH6

H is heterogeneous aka APUs only.

 

[Joe NYC]

H is heterogeneous aka APUs only.

Oh, so that's what the H stands for. Good to know.

As far as the heterogeneous, MLID floated an idea that AMD may be adding some AI chiplets to Turin at some point. If that happens, there is going to be a whole lot of heterogenousity (or is it heterogeneousness ?) even in SP5 sockets.

 

[adroc_thurston]

MLID floated an idea that AMD may be adding some AI chiplets to Turin at some point

That's just tip of the iceberg.

 

[Frenetic Pony]

That's just tip of the iceberg.

I'd imagine they're busy trying to maximize AI inference chips in any way they can. The Xilinx stuff is all inference and between self driving cars and the sudden explosion in generative AI inference, especially in the cloud, and just how much it costs to generate each query (especially in the cloud stuff) there's a huge market for high efficiency inference.

AMD offering a kind of inference heavy server chip somewhere on the spectrum of a traditional CPU socket and a MI300 type offering seems a natural direction to go.

 

[A///]

there was rumor once the xilinx purchase was approved by all major countries that zen 6 might get some ai tech and it was hinted at by amd a couple years ago and even last year by lisa during the zen 4 announcement. it's all speculos on our part and anything we discuss is worth nothing as all of that would have been designed into place long before any of our fartings on the topic were summed up in a digital town square speech like in this thread or the zen 6 thread if there is one because I have never seen it myself but it doesn't exist one of you should make it.

 

[igor_kavinski]

zen 6 might get some ai tech

Zen 5 has integrated AI according to this slide:

 

[soresu]

Zen 5 has integrated AI according to this slide:

View attachment 84597

I suspect this means instructions for bf16 and matrix handling.

AMD has typically lagged far behind ARM for CPU µArch design in this arena.

 

[Thibsie]

there was rumor once the xilinx purchase was approved by all major countries that zen 6 might get some ai tech and it was hinted at by amd a couple years ago and even last year by lisa during the zen 4 announcement. it's all speculos on our part and anything we discuss is worth nothing as all of that would have been designed into place long before any of our fartings on the topic were summed up in a digital town square speech like in this thread or the zen 6 thread if there is one because I have never seen it myself but it doesn't exist one of you should make it.

Not a problem, those tech were license by Xilinx to AMD long before the deal.

 

[A///]

I suspect this means instructions for bf16 and matrix handling.

AMD has typically lagged far behind ARM for CPU µArch design in this arena.

The last time z5 ai inferencing and optimisations came up this was what was said. I don't expect any dedicated tile for ai work on consumer ryzen until zen 6 or 7. bloody bizarre to even type that out as I'd laughed at the idea of an amd comeback even on the even of zen's launch.

 

[Fjodor2001]

I suspect this means instructions for bf16 and matrix handling.

AMD has typically lagged far behind ARM for CPU µArch design in this arena.

I wonder how much that will improve the performance for typical machine learning tasks. Will it be a viable option compared to using GPUs?

 

[A///]

Not a problem, those tech were license by Xilinx to AMD long before the deal.

of course, amd was very open with their working together long before the deal was announced. xilinx holds onto a lot more ip than amd has ever used. interesting enough in their disclosure reports mid acquisition xilinx was approached by another paarty seeking to buy them out. never was said who it was and we can only wildly guess.

 

[A///]

I wonder how much that will improve the performance for typical machine learning tasks. Will it be a viable option compared to using GPUs?

gpu will always be more efficient for the power used than a cpu. unless x86 can do the work of a gpu under an apple m processor 100% load power draw.

 

[Fjodor2001]

gpu will always be more efficient for the power used than a cpu. unless x86 can do the work of a gpu under an apple m processor 100% load power draw.

So I guess the question then is why add this to the CPU at all, if everyone doing serious machine learning tasks will use GPU anyway?

 

[soresu]

I wonder how much that will improve the performance for typical machine learning tasks. Will it be a viable option compared to using GPUs?

As with all things I suppose it depends on the ML frameworks used and their optimisations.

There's no reason that both can't be used in tandem - in fact that is ARM's "Total Compute" focus for the last couple of years, and Qualcomm have similarly touted a combined AI compute spec in recent years including the CPU, Adreno and Hexagon DSP together.

 

[A///]

So I guess the question then is why add this to the CPU at all, if everyone doing serious machine learning tasks will use GPU anyway?

The assumption you have is everyone will have a top of the line consumer gpu or quadro. that isn't the case. the "engines" on processor can be flexed to be resourceful for speeding up or improving general software. it's a chicken and egg problem if you ask me.

 

[soresu]

gpu will always be more efficient for the power used than a cpu

Not always.

GPUs are great for certain tasks, and terrible at others.

A many core CPU will always be better than a GPU general compute at video encoding, especially as complexity and interframe dependencies increase with each new codec generation.

I believe data heavy processing favor GPUs and task/decision heavy (uber branching) processing favors CPUs.

 

[A///]

Not always.

GPUs are great for certain tasks, and terrible at others.

A many core CPU will always be better than a GPU general compute at video encoding, especially as complexity and interframe dependencies increase with each new codec generation.

You're mixing up software encoding with hardware encoding here. if you use the igpu on intel or amd to encode video it's hardware encoding and the quality will not be there. We're at the point were brute forcing software encoding is very fast, at leas tcompared to a mere 5 years ago.

 

[Fjodor2001]

The assumption you have is everyone will have a top of the line consumer gpu or quadro. that isn't the case. the "engines" on processor can be flexed to be resourceful for speeding up or improving general software. it's a chicken and egg problem if you ask me.

Well, I'm thinking there's not that many desktop PC users doing machine learning training tasks as a "hobby project". And those that do it professionally will use GPUs anyway.

But regardless it will be interesting to see how an AMD 8950X Zen 5 CPU with these AI extensions will improve ML task performance compared to 7950X. Will it e.g. be 2x or 5x faster for such operations?

And how will 8950X perform against e.g. an RTX4080/4090 GPU. Will the GPU be e.g. 2x, 5x, or 10x faster?

 

[CakeMonster]

I fear people might be disappointed, whatever 'AI' instructions are coming in Z5 was probably planned years ago before the current hype and is probably some really simple stuff. I'm sure it can be useful, but I doubt its revolutionary, that will probably come in later generations.

 

[soresu]

You're mixing up software encoding with hardware encoding here

No I'm not.

I specifically said general compute - as in using the shaders instead of an ASIC.

An ASIC can be more efficient for power consumption, but it's utility is minimal as a fixed function hardware block.

ASIC encoding quality is also generally not up to snuff with software encoding in terms of quality tuning for at least the first 5+ years of codec availability, and often beyond.

 

[A///]

Well, I'm thinking there's not that many desktop PC users doing machine learning training tasks as a "hobby project". And those that do it professionally will use GPUs anyway.

But regardless it will be interesting to see how an AMD 8950X Zen 5 CPU with these AI extensions will improve ML task performance compared to 7950X. Will it e.g. be 2x or 5x faster for such operations?

And how will 8950X perform against e.g. an RTX4080/4090 GPU. Will the GPU be e.g. 2x, 5x, or 10x faster?

Quadro isn't cheap.

your other questions... why not ask me when humans will grow wings. easier to answer.

 

[A///]

No I'm not.

I specifically said general compute - as in using the shaders instead of an ASIC.

An ASIC can be more efficient for power consumption, but it's utility is minimal as a fixed function hardware block.

ASIC encoding quality is also generally not up to snuff with software encoding in terms of quality tuning for at least the first 5+ years of codec availability, and often beyond.

If you want the most power efficient video encoder you buy that AMD/Xilix hardware. 1500 bones gets you something more efficient than the two. I'll send you my xmas gift list and you lot can take match sticks and see who ends up buying that for me.

 

[soresu]

We're at the point were brute forcing software encoding is very fast, at leas tcompared to a mere 5 years ago.

Sure, if you can afford a TR 3990X or 5995WX system.