Discussion ARM Cortex/Neoverse IP + SoCs (no custom cores) Discussion

[Gideon]

Looks like C1-Nano is the only LITTLE core they'll release, so it replaces everything from A520 to A320 and all the A53 being used in smartwatch SoCs.

Wait and all they managed to do is 26% efficiency and 5% performance uplift, using a OoO design?



It will still be absolutely dumpstered by Apple and Qualcomm small cores on perf/w then.

I'm sure it has its uses and I don't mind seeing it in wearables, but I'd rather have those in my phone.

 

[adroc_thurston]

It will still be absolutely dumpstered by Apple and Qualcomm small cores on perf/w then

Cortex-A7x of the Gen is competition for those.

 

[Cardyak]

Wait and all they managed to do is 26% efficiency and 5% performance uplift, using a OoO design?

View attachment 130337

It's still In-Order

 

[DZero]

In order and even ARM is putting the Nano cores on the lowest tier. Seems that targets only wearables.

Still Huawei "small" core might end trashing hard those cores.

 

[511]

Hiding behind SME

https://twitter.com/x/status/1970113376111276034

 

[okoroezenwa]

Hiding behind SME

You’ll get over it.

 

[511]

You’ll get over it.

It just sucks that every tech tuber uses Geekbench and mobile doesn't have benchmark variety for CPUs.

 

[gdansk]

You’ll get over it.

I won't. I complained about AES the entire GB5 life cycle and won in the end.
AMD and Intel will add AMX joke units but ultimately it's another memory bandwidth to in SoC accelerator and it's pretty useless to include in GB7 when it'll be ubiquitous.

 

[511]

I won't. I complained about AES the entire GB5 life cycle and won in the end.

At least AES is more useful than SME and it is relevant for people in everyday life.

AMD and Intel will add AMX joke units but ultimately it's another memory bandwidth to in SoC accelerator and it's pretty useless to include in GB7 when it'll be ubiquitous.

I am waiting for APX addition and how much score it will boost cause it's a real boost as for AMX it is already supported in GB6.

 

[gdansk]

At least AES is more useful than SME and it is relevant for people in everyday life.

I am waiting for APX addition and how much score it will boost cause it's a real boost as for AMX it is already supported in GB6.

AMX is not included in any client CPUs. It's what is comparable to SME.

 

[soresu]

I am waiting for APX addition and how much score it will boost cause it's a real boost as for AMX it is already supported in GB6.

Going by PR from this new "x86 dudes blah blah we're cool don't buy ARM" group I think we might for once see new instruction sets released across both Intel and AMD µArchs within a very similar timeframe.

With ARM now the real market competition (and RISC-V a shadowy threat on the far horizon) Intel has to know that trying to pull a fast one over AMD and implementing brand new instruction sets on their on (and worse botching their implementation) only serves to make x86 options look fragmented and less attractive to prospective customers looking at long term investment.

I don't think that we will see a return to the long bygone days of motherboards accepting either Intel or AMD SoC's, but we may at least see more instruction set parity in comparable hw generations going forward.

This would potentially make compiler support significantly simpler for modern x86 CPUs.

 

[511]

With ARM now the real market competition (and RISC-V a shadowy threat on the far horizon) Intel has to know that trying to pull a fast one over AMD and implementing brand new instruction sets on their on (and worse botching their implementation) only serves to make x86 options look fragmented and less attractive to prospective customers looking at long term investmen

was AVX-2 bad as well cause AVX-512 dropped clocks to the ground and it took some time for them to figure stuff out? adding Instruction is only part you need to put software support as well which AMD lacks.

 

[MS_AT]

I am waiting for APX addition and how much score it will boost cause it's a real boost

If they pick the sub-benches right it will be nice round 0%

This would potentially make compiler support significantly simpler for modern x86 CPUs.

What would make it simpler is if Intel and AMD would provide support in the compilers for their CPUs. Intel has been doing that just fine from my observations. AMD just doesn't care. They no longer have an excuse they are recovering, they just simply don't care about anything else than AOCC. (I don't track their involvment with msvc so don't now about that).

 

[511]

If they pick the sub-benches right it will be nice round 0%

Well I doubt there will exist such a subtest with 0% improvement but only time will tell.

 

[soresu]

What would make it simpler is if Intel and AMD would provide support in the compilers for their CPUs. Intel has been doing that just fine from my observations. AMD just doesn't care. They no longer have an excuse they are recovering, they just simply don't care about anything else than AOCC. (I don't track their involvment with msvc so don't now about that).

Hopefully with AMD claiming that they are shifting to a much more software focused company they might start solving these issues.

For sure HIP RT feels like an also ran project rather than any kind of priority at the moment with the RDNA4 update coming something like 4-5 months after the actual hardware.

 

[adroc_thurston]

For sure HIP RT feels like an also ran project rather than any kind of priority at the moment with the RDNA4 update coming something like 4-5 months after the actual hardware.

Who cares about proviz, they have bigger fish to fry.

 

[soresu]

Who cares about proviz, they have bigger fish to fry.

That's the point tho, why even bother at all if you are going to be 4-5 months late?

 

[poke01]

Hopefully with AMD claiming that they are shifting to a much more software focused company they might start solving these issues.

Sigh, you believe that? AMD is a software company in the sense Apple is a AI company, in dreamland only.
It’s just marketing bull. AMD is a hardware company thru and thru.

And that stupid ROCm 7 update today proved my point.

 

[adroc_thurston]

That's the point tho, why even bother at all if you are going to be 4-5 months late?

Because they still sell proviz stuff but that's not a priority.

AMD is a hardware company thru and thru.

Everyone is.

And that stupid ROCm 7 update today proved my point.

?
MI350 runs real nice there.

 

[poke01]

Because they still sell proviz stuff but that's not a priority.

Everyone is.

?
MI350 runs real nice there.

Why is it not as good on Strix halo?

 

[adroc_thurston]

Why is it not as good on Strix halo?

Because Meta/Oracle/MS/whatever do not run racks of stxH.

 

[511]

Hopefully with AMD claiming that they are shifting to a much more software focused company they might start solving these issues.

I think it was only for AI software.

 

[soresu]

It will still be absolutely dumpstered by Apple and Qualcomm small cores on perf/w then.

Cortex-A7x of the Gen is competition for those.

This ☝️

Apple's "little" cores have been more equivalent to A7x/7xx for many generations now, and from what I can tell Qualcomm small/little cores are just the big ones with stuff taken out to cut on area, sort of like the difference between A78 and X1 working from the same foundation, just with more/wider silicon in the case of X1.

ARM Ltd have been steadily increasing IPC of A7xx/Cx Pro for several generations now, but it seems that perf/watt has been their prime focus above all else as efficiency gains are usually the touted metric.

 

[soresu]

?
MI350 runs real nice there.

Absolute performance for a given hw SKU is not the only measure of software.

From what I have heard ROCm has historically been something of a dog to get working properly on Linux, though I haven't asked since v7 if that is still the case.

I wouldn't say that it doesn't matter if it performs well if SDK usability/maintenance is up a creek, but it certainly doesn't help when competing against an entrenched mature software platform.

 

[adroc_thurston]

and from what I can tell Qualcomm small/little cores are just the big ones with stuff taken out to cut on area,

They're basically halved.

From what I have heard ROCm has historically been something of a dog to get working properly on Linux, though I haven't asked since v7 if that is still the case.

It hasn't been the case for eons, unless you somehow try to RDNA. Don't.

I wouldn't say that it doesn't matter if it performs well if SDK usability/maintenance is up a creek, but it certainly doesn't help when competing against an entrenched mature software platform

Consumer POV.
You're not the person ROCm is for (for now, anyway).