Question Zen 6 Speculation Thread

[poke01]

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.

You know what’s worse than AMX and they still include it in client, an NPU.

IMO I rather they add the smallest NPU and give more space to AMX.

 

[adroc_thurston]

IMO I rather they add the smallest NPU and give more space to AMX

GPU man. GPU is king.
Radeon guys won the client matmul war at AMD, I just dunno when it impacts the roadmap. Probably zen8 gen (2030).

 

[Win2012R2]

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

AMX is far more niche than even AVX512 and that took off only when Zen4 got released.

I can see why they'd want to keep AMX separate - it bombs clocks on Intel when used, like original AVX512, but share unit will have crap perf and only good for legacy code, but who codes in AMX? Only some server side people, super limited, it's basically dead - that kind of processing is meant for GPUs, which in server environment no brainer choice.

Low latency inference probably only use case here, some limited financial sector stuff, that's just way too niche.

You know what’s worse than AMX and they still include it in client, an NPU.

It's sensible to include iGPU in client, even if it's relatively small but enough to pump out NPU level stuff, which currently isn't even used because Microsoft CoPilot totally bombed and it's getting worse all the time.

 

[adroc_thurston]

AMX is far more niche than even AVX512 and that took off only when Zen4 got released.

again, you're projecting Intel skill issues on vendors that actually ship stuff.
Apple has AMX/SME for eons and it works great.

 

[Joe NYC]

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

It's going to be tough time for Intel in server market. It already is.

Here is some "homework" I did on another forum, based on 2 datapoints (coming from an analyst):
- Server CPU market grew 16% YoY (according to one analyst, Rikers, 15%-17% estimate)
- Intel server CPU revenue grew 8% YoY (same analyst, but Intel report shows 7%)

My "homework":
-----------------------------

With whole market growing 16%: 100 * 1.16 = 116
A gain of 16
Intel, being 60% of the market, growing 8%: 60 * 1.08 = 64.8
Gain of 4.8
Gain left for AMD: 16 - 4.8 = 11.2
If new AMD is 40 + 11.2 = 51.2
then, for AMD, it is a gain of 51.2 / 40 = 1.28 or 28% YoY
New AMD market share would be 51.2 / 116 = 44.13% up from 40%

 

[adroc_thurston]

It's going to be tough time for Intel in server market

underestimation of the decade man.
we're entering the pain zone, summon the cenobites.

 

[Win2012R2]

Apple has AMX/SME for eons and it works great.

There is no Apple2 to compare it against - where as for AMX there is a vendor who invented it and got a dedicated AMD unit per core, so it will run X times faster with all things being equal, and when X is high number like say 12 in Zen6 CCDs then for those who actually care to use AMX (very few) this will be unacceptable.

So the question is then why even bother bringing something that got basically no software for it anyway.

Having said all that I will certainly be glad if AMD brings it as shared unit though, I don't want them to waste so much silicon on marginal crap like that.

 

[adroc_thurston]

There is no Apple2 to compare it against - where as for AMX there is a vendor who invented it and got a dedicated AMD unit per core, so it will run X times faster with all things being equal, and when X is high number like say 12 in Zen6 CCDs then for those who actually care to use AMX (very few) this will be unacceptable.

My point is that a shared matmul unit is a bit hacky but works well and gives client nice matmul rates per cluster while giving the programmer a very CPU-ish model to work with.

AMD will ship ACE, it'll be everywhere and it'll be the standard.

So the question is then why even bother bringing something that got basically no software for it anyway.

ship h/w first and let people build for it.

 

[Joe NYC]

AMX is far more niche than even AVX512 and that took off only when Zen4 got released.

There are going to be joint standards now. Something like AMD64 / x86-64 Version 2. It will take off faster.

 

[poke01]

There is no Apple2 to compare it against - where as for AMX there is a vendor who invented it and got a dedicated AMD unit per core, so it will run X times faster with all things being equal, and when X is high number like say 12 in Zen6 CCDs then for those who actually care to use AMX (very few) this will be unacceptable.

So the question is then why even bother bringing something that got basically no software for it anyway.

Having said all that I will certainly be glad if AMD brings it as shared unit though, I don't want them to waste so much silicon on marginal crap like that.

think about why all ARM vendors who also have SME (Qualcomm/mediatek) went per cluster and not per core. Its not something your average user is gonna use every day so its a waste to do it per core.

There are going to be joint standards now. Something like AMD64 / x86-64 Version 2. It will take off faster.

Yeah whoever ships it first in all areas will make it happen.

 

[Win2012R2]

There are going to be joint standards now.

If they agree, it's clearly in AMD's interest not to about AMX/ACE or at least delay it for a long time because shared implementation obviously will have performance deficiencies vs dedicated, if it's not used then why bother, and if it is then they will be at disadvantage, so why help competitor? Might as well not agree to standard and sell GPUs.

think about why all ARM vendors who also have SME (Qualcomm/mediatek) went per cluster and not per core. Its not something your average user is gonna use every day so its a waste to do it per core.

Yes, obviously, so why AMD should waste per cluster AMX _and_ be easy target for Intel showing how much faster their server AMX CPUs are, which is the only area where AMX code lives right now. Marginalising AMX is in the long term interest for AMD who would prefer to sell GPUs.

 

[adroc_thurston]

Yeah whoever ships it first in all areas will make it happen.

you already know the answer.

 

[Joe NYC]

think about why all ARM vendors who also have SME (Qualcomm/mediatek) went per cluster and not per core. Its not something your average user is gonna use every day so its a waste to do it per core.


Yeah whoever ships it first in all areas will make it happen.

Maybe Razer Lake on Intel side? NVL is supposed to have APX, AVX10, but not ACE. Razer Lake should have all 3. And Zen 7 should also have all of the new standards implemented.

 

[Joe NYC]

If they agree, it's clearly in AMD's interest not to about AMX/ACE or at least delay it for a long time because shared implementation obviously will have performance deficiencies vs dedicated, if it's not used then why bother, and if it is then they will be at disadvantage, so why help competitor? Might as well not agree to standard and sell GPUs.

Intel has AMX only in server and it is not very well supported.

ACE will be across both client and server. I don't see why AMD would not want to implement it in the nearest possible core release.

It is still going to be poor man's implementation, for limited tasks. It is not going to take business from gigantic GPUs like Rubin and Mi400.

Yes, obviously, so why AMD should waste per cluster AMX _and_ be easy target for Intel showing how much faster their server AMX CPUs are, which is the only area where AMX code lives right now. Marginalising AMX is in the long term interest for AMD who would prefer to sell GPUs.

AMD is going after Apple, not just Intel (with Zen 7). There is no way to catch up to Apple in certain applications without ACE.

 

[DrMrLordX]

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

Which is unfortunate that after all this time and after so many opportunities for libraries to be introduced to make GPUs better code targets, we're still not there.

 

[Kepler_L2]

If they agree, it's clearly in AMD's interest not to about AMX/ACE or at least delay it for a long time because shared implementation obviously will have performance deficiencies vs dedicated, if it's not used then why bother, and if it is then they will be at disadvantage, so why help competitor? Might as well not agree to standard and sell GPUs.

Yes, obviously, so why AMD should waste per cluster AMX _and_ be easy target for Intel showing how much faster their server AMX CPUs are, which is the only area where AMX code lives right now. Marginalising AMX is in the long term interest for AMD who would prefer to sell GPUs.

Both Intel and AMD are doing 1 ACE unit per 2 cores

 

[poke01]

Both Intel and AMD are doing 1 ACE unit per 2 cores

so in a 12 core CPU there are 6 ACE units. Hopefully those units are small

 

[Tachyonism]

I hope that ACE/AMX/whatever those matrix multipliers are called will replace the useless NPU on the current CPUs in the future.

 

[poke01]

I hope that ACE/AMX/whatever those matrix multipliers are called will replace the useless NPU on the current CPUs in the future.

depends, NPUs don't belong in desktop CPUs but have a place in SoCs. Don't know why Intel is so keen on adding NPUs to desktops its just extra cost for no reason

AMD isn't adding NPUs on desktop

 

[511]

depends, NPUs don't belong in desktop CPUs but have a place in SoCs. Don't know why Intel is so keen on adding NPUs to desktops its just extra cost for no reason

Microsoft thingy probably i don't want silicon wasted on NPU

 

[poke01]

Microsoft thingy probably i don't want silicon wasted on NPU

is MS forcing NPUs on desktop? pretty sure thats just Intel

 

[511]

is MS forcing NPUs on desktop? pretty sure thats just Intel

CoPilot+

 

[poke01]

CoPilot+

nobody cares about that, local AI is dead on Windows

 

[Claudiovict]

Radeon guys won the client matmul war at AMD

This means AMD/Radeon will *actually* try to compete with nvidia?

 

[Fjodor2001]

AMD isn't adding NPUs on desktop

Are you sure about that? If so I guess it means they will not be Copilot+ compliant.