Question Zen 6 Speculation Thread

[Covfefe]

Thanks.

From your linked document:


From your link:

As much as I respect this source, this is speculation ..... and I seriously doubt it.

I believe that future e-cores will support AVX10 with a 128 bit path and that P cores will likely support a full 512 bit path. This is my speculation.; however, the SPEC linked by @511 clearly shows that it supports 3 different bit paths .... not ONLY 512bit.

Also, his statement is incomplete. "AVX10 had intended to allow 256 or 512 bit modes depending upon processor capabilities". He left off 128bit further raising doubt to the speculation statement.

And I don't know how you interpret adroc's statement:


... any other way than adroc doesn't believe AVX10 supports 128b or 256b (which it clearly does). Of course he COULD clarify his difficult to justify statement to explain himself.

128-bit-only mode was already killed a couple years ago.

FCLC (@fclc@mast.hpc.social)

LETS GOOOOO #Intel listened! No more ambiguity, no more questioning, 128 bit only implementations are no longer valid for #AVX10! A massive thanks to the Intel ISA folks! And a major thanks to @chipsandcheese@techhub.social for giving me a platform to express my concerns on the spec! For...

mast.hpc.social

 

[511]

Thanks.

From your linked document:


From your link:

As much as I respect this source, this is speculation ..... and I seriously doubt it.

I believe that future e-cores will support AVX10 with a 128 bit path and that P cores will likely support a full 512 bit path. This is my speculation.; however, the SPEC linked by @511 clearly shows that it supports 3 different bit paths .... not ONLY 512bit.

That's not true so initially there were three implementation of vector length 128/256/512b 128 got killed and so did 256bit vector linked in phoronix article now we only got 512 vector length . I mean the size supported by HW.

And than there is the size of data you can have 128/256/512b vectors with respective instructions which is different from hardware Maz supported vector length.

 

[OneEng2]

It's my understanding that AVX10 is designed to provide a single code base the ability to run on a variety of cores having various levels of bit widths in the execution path.

AMD are certainly not going to need to support anything other than 512b.

Intel p cores ? Likely 512b.

Intel e cores? I'm guessing 128b and maybe 256b.

What is the point of AVX10 over AVX512 if not for the variable bit widths?

 

[adroc_thurston]

It's my understanding that AVX10 is designed to provide a single code base the ability to run on a variety of cores having various levels of bit widths in the execution path.

No. It's 512b only. The underlying FMA hardware can be smaller, but regs/shuffle etc no luck.

What is the point of AVX10 over AVX512 if not for the variable bit widths?

It was the point.
Then AMD showed the rightful way.

 

[MS_AT]

It's my understanding that AVX10 is designed to provide a single code base the ability to run on a variety of cores having various levels of bit widths in the execution path.

AVX10 does exactly the same register bit width support as AVX512 with extensions.

ISA however does not specify how it should be implemented in hardware.

What is the point of AVX10 over AVX512 if not for the variable bit widths?

That it fixes the extension mess of AVX512. By unifying all popular and de facto standard extensions under one name.

Intel also intended to make 512b registers optional so they could make their life easier on desktop and keep other nice features while keeping 512b for server.

Fortunately they were made to understand that it was stupid approach because it would fragment the ISA they wanted to unify. If they want to save space they can do it via hw implementation (like AMD does).

This matters because if the same instructions are supported everywhere then you can write and test your software on cheap CPU (that does 256b datapaths in hw) and then deploy it on faster more expensive CPU (that does 512b datapaths in hw) and get the expected speed up without lifting a finger, at least in theory as there are always corner cases😉

If they kept AVX10/256 that would not be the case making life harder for everyone.

 

[ETI4711]

No. It's 512b only. The underlying FMA hardware can be smaller, but regs/shuffle etc no luck.

It was the point.
Then AMD showed the rightful way.

Robert Hormuth describes it as follows:
Regular x86 Ecosystem Advisory Group working sessions have already surfaced “aha” moments that only an open forum can uncover. One highlight to note, cloud providers made it clear that memory tagging, once considered a debug only feature, is mission critical for protecting production workloads. Another important moment to highlight is that maintaining identical ISA support across all form factors dramatically streamlines software delivery; letting developers write once and deploy anywhere.

 

[DrMrLordX]

From practical point of view you can do that with AVX512, as only Xeon Phi does not support 128b/256b AVX512 extension afaik. Yes you need to check one more flag for dynamic dispatch compared to AVX10 but I would say that's hardly a problem.

Are there any instances of developers using AVX-512 to handle 256b length vectors?

No, that is desd.
AVX10 is 512b only.

Hmm. Learn something new every day. I guess it makes sense.

 

[511]

Are there any instances of developers using AVX-512 to handle 256b length vectors?

there are 256b instructions as part of AVX-512 that can do what AVX-512 does but with 256b vectors

Intel® Intrinsics Guide

Intel® Intrinsics Guide includes C-style functions that provide access to other instructions without writing assembly code.

www.intel.com

 

[DrMrLordX]

there are 256b instructions as part of AVX-512 that can do what AVX-512 does but with 256b vectors

Intel® Intrinsics Guide

Intel® Intrinsics Guide includes C-style functions that provide access to other instructions without writing assembly code.

www.intel.com

Yes, I'm aware (now) but how many developers have actually utilized that in their code?

 

[MS_AT]

Yes, I'm aware (now) but how many developers have actually utilized that in their code?

I know of at least one (myself 😉 ).

Are there any instances of developers using AVX-512 to handle 256b length vectors?

Well compiler can use it for you like here https://godbolt.org/z/fT1WYfKvz (it's trivial example, but here we are summing 72 floats, there are 16 floats in 512b reg, so it compiles down to 4 512b ops and one 256b op (ymm are 256b regs, zmm are 512b ones).

So knowing compiler can use them for you and/or you can explicitly use them (for example for tail handling, small transpose or gemm kernels etc) it's really hard to give you more detailed answer than: "For sure".

 

[RTX]

https://twitter.com/x/status/2039738699899187377

Some new changes vs Zen5

 

[LightningZ71]

Robert Hormuth describes it as follows:
Regular x86 Ecosystem Advisory Group working sessions have already surfaced “aha” moments that only an open forum can uncover. One highlight to note, cloud providers made it clear that memory tagging, once considered a debug only feature, is mission critical for protecting production workloads. Another important moment to highlight is that maintaining identical ISA support across all form factors dramatically streamlines software delivery; letting developers write once and deploy anywhere.

It took a whole a** working committee to come up with: if we use a standard ISA across our whole stack, it makes development easier?!?!

 

[511]

It took a whole a** working committee to come up with: if we use a standard ISA across our whole stack, it makes development easier?!?!

The committee was there just for the name honestly it's Intel's fault of screwing with the ISA and causing fragmentation in the first place even though they never wanted it but 10nm made it.

 

[Tachyonism]

Seems like ARM was considered for SD2. Did Zen6 changed/improved or was the Crapdragon chip so bad they gave up?

 

[gdansk]

In that case the CPU is of low importance and the GPU and its software is more important.

 

[LightningZ71]

The GPU is BY FAR the limiting factor on these handhelds in 99% of the cases.

 

[adroc_thurston]

Did Zen6 changed/improved or was the Crapdragon chip so bad they gave up?

Valve was begging for a semicustom Deck 2 part and got blueballed by AMD.

In that case the CPU is of low importance and the GPU and its software is more important.

AMD is just cruel.
You need bajillion lifetime units to warrant a SC slot.

 

[LightningZ71]

Valve needs to remember, the chip that they CURRENTLY use wasn't even made custom for them. It was someone else's project that they bought into.

this stuff isn't cheap AT ALL...

 

[511]

AMD is just cruel.
You need bajillion lifetime units to warrant a SC slot.

Only Sony fits the bill

 

[adroc_thurston]

Valve needs to remember, the chip that they CURRENTLY use wasn't even made custom for them. It was someone else's project that they bought into.

this stuff isn't cheap AT ALL...

Money's not the problem (Valve's rich). It's the units total that gatekeep stuff.

Only Sony fits the bill

Xbox gets one too. But not Valve.

 

[Joe NYC]

Valve was begging for a semicustom Deck 2 part and got blueballed by AMD.

AMD is just cruel.
You need bajillion lifetime units to warrant a SC slot.

I don't know how the licensing would work out, but can't Steam just buy either the upcoming XBox handheld or PlayStation handheld CPU?

 

[Tachyonism]

Valve was begging for a semicustom Deck 2 part and got blueballed by AMD.

Yes, that's obvious. But why did they give up on Snapdragon?

I know that their GPU driver on Linux is ass, even worse than Nvidia.

The GPU is BY FAR the limiting factor on these handhelds in 99% of the cases.

Most handheld APUs are just repurposed laptop chips, so it's understandable that the CPU part is too strong but the iGPU is too weak. I'd say a weak CPU like LNL from Intel would fit better in a handheld.

If AMD want to capture the handheld market, they need a custom line for these kinds of products. But it's understandable that it's a too risky bet for the company.

 

[Tachyonism]

I don't know how the licensing would work out, but can't Steam just buy either the upcoming XBox handheld or PlayStation handheld CPU?

I think both parties have exclusive deal with AMD that prevents 3rd-party from buying those products. Furthermore, PS6 CPU probably doesnt have enough functionalities of a normal desktop CPU, and Gabe wont like that.

Imo, the best case scenario for Valve is if they can harvest the defects from those semi-custom SoC like the BC-250.

 

[gdansk]

If AMD want to capture the handheld market, they need a custom line for these kinds of products

But they captured it accidentally without designing a single SoC for handhelds. A few special SKUs here and there and the consumers buy it up.
I'm still not convinced they have to do special part (for handhelds) unless paid to do so.

 

[adroc_thurston]

I don't know how the licensing would work out, but can't Steam just buy either the upcoming XBox handheld or PlayStation handheld CPU?

No.

If AMD want to capture the handheld market, they need a custom line for these kinds of products. But it's understandable that it's a too risky bet for the company

They care a negative amount about client.