Discussion Future ARM Cortex + Neoverse µArchs Discussion

[ikjadoon]

Well I wasn't expecting that to happen....

Troubled AI Processor Developer Graphcore Finds a Buyer: SoftBank

www.anandtech.com

Not sure exactly what gain this could bring for ARM beyond a more complete silicon IP for AI/ML to augment their Ethos NPUs.

There were a few rumors, but one never know how these things might end up:

SoftBank closer to buying Graphcore, say reports

Talks between SoftBank Group and AI chipmaker Graphcore over a potential acquisition of the latter have moved forward.

www.eenewseurope.com

https://www.eetimes.com/is-graphcore-deal-finally-about-to-close/

Officially, Arm is hinting they're looking at NPUs again more seriously, but it's been pretty light on info. Underline is mine.

PCWorld: You mentioned an NPU before. But you don’t build an NPU, at least not in the CSS architecture. Did I miss something?

Arm CEO: "There isn’t an NPU today on the PC side. We have NPUs today on what I would call the entry-embedded line. But yeah, we haven’t gone public with our NPUs for the high end."

//

PCWorld: And you’re going to have an NPU for these high-end processors. You just haven’t come to market yet.

Arm CEO: "You can extrapolate that."

//

There are too many rumors these days to consider, but this one is tangentially related while quickly veering into wild (nuclear fusion...?):

The UK-based Arm will create a specialized division dedicated to AI chips, with the goal of having a prototype ready by spring 2025. Following the prototype processor in spring 2025, mass production of Arm's processors is anticipated to commence in the fall of 2025 at a contract maker of chips, such as TSMC, Samsung Foundry, or Intel.

SoftBank intends to build datacenters using its own processors across the U.S., Europe, Asia, and the Middle East by 2026. Due to the substantial power requirements of datacenters, the company will reportedly also expand into power generation. Plans include the development of wind and solar power facilities, with an eye on pioneering next-generation fusion technology, Nikkei claims.

//

Phoronix has some Graviton4 (Neoverse V2 / Cortex-X3) benches up:

AWS Graviton4 Benchmarks Prove To Deliver The Best ARM Cloud Server Performance - Phoronix

www.phoronix.com

 

[DrMrLordX]

Phoronix has some Graviton4 (Neoverse V2 / Cortex-X3) benches up:

Just glancing at the Geomean results (yeah yeah not always a good idea), the performance looks solid. Graviton4 seems to be beating competing Xeon instances. But you'll never see that exact hardware config outside of AWS so . . . meh.

 

[Nothingness]

More Graviton4 tests from Phoronix: https://www.phoronix.com/review/graviton4-96-core

 

[KompuKare]

More Graviton4 tests from Phoronix: https://www.phoronix.com/review/graviton4-96-core

Looks like they're now in second place behind AMD but ahead of Intel in most of the benches, which is pretty impressive.

 

[FlameTail]

A demonstration of ARM KleidiAI:

https://twitter.com/x/status/1813965261935989042

 

[ikjadoon]

Arm gave barely more details on X925 at Computex: X925 uses a "grounds-up new microarchitecture" & that with expanded vector, it's +50% TOPS (via the CPU).

//

Re Arm v Qualcomm: both filed motions for summary judgment on 10 July (the date set by the Court for these "case dispositive" motions). A summary judgment motion: the Court rules / makes a Judgment on claims ABC now before trial. So both Arm and Qualcomm are trying to squeeze in a (partial) win before trial.

1. Arm has asked for a partial summary judgment against Qualcomm (#371). Arm claims the Court has enough evidence already (via depositions, expert witnesses, documents, contracts, etc.) to rule that 1) Arm properly terminated NUVIA's ALA, 2) NUVIA and Qualcomm both violated Section 15.1 of the NUVIA ALA, and 3) Arm did not breach Section 15.1 of the NUVIA ALA.
2. Qualcomm is asking for summary judgment against Arm (#389). Unfortunately, their motion is not public currently. But as it's not written as a "partial" summary judgement, I assume it focuses on Arm's core claims.

Sadly, if Arm's summary judgment succeeds, it likely means we, the public, will not have a chance to see the evidence relating to those claims [other claims will still need trial → public evidence]. But, if Arm's motion is dismissed, all the evidence should be publicly accessible during trial.

Now, if Qualcomm's motion succeeds, that may be the end of this litigation, sans appeals by Arm. So that would mean virtually no evidence will ever be made public. But we won't know how the Court rules until at least late August, if not later.

ignore the left column of dates; those are the old dates


//

Unrelatedly, MediaTek is allegedly entering the datacenter CPU market?

Source: https://money.udn.com/money/story/11162/8094902

English commentary: https://xpu.pub/2024/07/16/mediatek-server/

 

[SarahKerrigan]

English commentary: https://xpu.pub/2024/07/16/mediatek-server/

That site has some incredibly obnoxious Javascript on it. No right click, pops up a message that says "You are not allowed to copy content or view source" if you ctrl-c???

 

[soresu]

That site has some incredibly obnoxious Javascript on it. No right click, pops up a message that says "You are not allowed to copy content or view source" if you ctrl-c???

I've seen numerous sites like it.

Some retail sites do it too.

 

[soresu]

Arm gave barely more details on X925 at Computex: X925 uses a "grounds-up new microarchitecture" & that with expanded vector, it's +50% TOPS (via the CPU).

Curious that it is a "ground up new" µArch yet A725 has more advanced µArch features.

 

[SarahKerrigan]

I've seen numerous sites like it.

Some retail sites do it too.

Yeah, and it's obnoxious.

 

[soresu]

Yeah, and it's obnoxious.

And pointless if a technical person really wants it.

 

[ikjadoon]

That site has some incredibly obnoxious Javascript on it. No right click, pops up a message that says "You are not allowed to copy content or view source" if you ctrl-c???

Ha, you're right. Stumbled onto the site a few weeks ago and never had a chance to notice that. It is restriction, especially when I just want to grab a quick quote. As soresu mentioned, content farms are going to copy it anyways, so I'm not sure how well it works.

I've noticed Videocardz.net also doesn't allow you to select any text.

Didn't notice their WCCFTech source link, which works fine.

//

Curious that it is a "ground up new" µArch yet A725 has more advanced µArch features.

I was curious about this and maybe I missed it. What's sparking people's interest in the A725?

The official disclosures appear like normal improvements (though larger % than usual): wider, bigger queues, more cache, smaller node, less L3 traffic, and some vague improvements else, etc.

I noticed an earlier comment re fused ops & zero-latency moves as "exclusive" to A725, but the X925 software manual lists the same chapters, though with fewer items.

What might be special in the A725 that the X925 does not get?

 

[ikjadoon]

Not sure where else to post this (as I think each firm gets its own quarterly earnings report forum post): it's the season again for some minor disclosures and maybe 1-2 interesting bits.

In the next ~14 days are most semi firms, sans NVIDIA later in August:

23 July: Alphabet/Google
30 July: AMD, Microsoft
31 July: Arm, Qualcomm, Mediatek
1 August: Amazon, Apple, Intel
28 August: NVIDIA

 

[soresu]

I was curious about this and maybe I missed it. What's sparking people's interest in the A725?

The official disclosures appear like normal improvements (though larger % than usual): wider, bigger queues, more cache, smaller node, less L3 traffic, and some vague improvements else, etc.

I noticed an earlier comment re fused ops & zero-latency moves as "exclusive" to A725, but the X925 software manual lists the same chapters, though with fewer items.

What might be special in the A725 that the X925 does not get?

View attachment 103485

View attachment 103486

Ah interesting.

I got my information from one of the coders working on FEX emu that also referenced one of the official ARM documents.

It's possible that the X925 documents simply weren't fully cooked when that happened which was weeks ago now, perhaps early June.

 

[soresu]

Ah found it....

Instruction fusion for FCMP+AXFLAG

Coming from a FEX emu dev it makes sense that they would prioritise it from ARM's AXFLAG description:

Convert floating-point condition flags from Arm to external format

This instruction converts the state of the PSTATE.{N,Z,C,V} flags from a form representing the result of an Arm floating-point scalar compare instruction to an alternative representation required by some software.

I think I also misread their words

Oh interesting, X925 doesn't have the same zero latency moves and instruction fusion of A725

as meaning that X925 doesn't have any of those features, rather than just a less extensive list of them.

My bad guys - for an Englishman I don't English so well it seems 🤣😂

Must be my close proximity to Scotland and many years living in Wales 😅

 

[FlameTail]

Curious that it is a "ground up new" µArch yet A725 has more advanced µArch features.

Clean sheet redesign of the core?

The last clean sheet design was the Cortex A76.

 

[soresu]

Clean sheet redesign of the core?

The last clean sheet design was the Cortex A76.

I don't think anything at ARM is truly "clean sheet" as they have a modular design philosophy where they share different µArch features between different design teams.

Even A76 had pieces of it in use 2-3 generations earlier.

 

[SarahKerrigan]

Clean sheet redesign of the core?

The last clean sheet design was the Cortex A76.

Yeah, nah. A76 is related to both the A72 and the A75, and A75 is closely related to A73. A73 itself is a 64-bit-ized and improved A12/A17, which itself as a highly aggressive derivative of the A9.

If it ain't broke, etc.

 

[Nothingness]

Even when a clean sheet design is done many things are reused. For instance there’s no need to redesign datapaths (unless you want to change target frequencies and change pipelining of these paths). That applies to all CPU companies in particular the ones that design close to the metal such as Intel used to.

 

[SarahKerrigan]

Even when a clean sheet design is done many things are reused. For instance there’s no need to redesign datapaths (unless you want to change target frequencies and change pipelining of these paths). That applies to all CPU companies in particular the ones that design close to the metal such as Intel used to.

Right.

Absolutely nobody goes from definition meetings to

entity someNewCore is

end someNewCore;

in a blank directory.

 

[Nothingness]

entity someNewCore is

end someNewCore;

.

I wonder if anyone is still using VHDL

 

[SarahKerrigan]

I wonder if anyone is still using VHDL

Less than they used to, but it lacks a lot of the particular brain-damages of Verilog (but then, so do some dialects of SystemVerilog.)

Defense semiconductor looooves VHDL, though.

 

[Nothingness]

Less than they used to, but it lacks a lot of the particular brain-damages of Verilog (but then, so do some dialects of SystemVerilog.)

Defense semiconductor looooves VHDL, though.

That’s basically C vs Ada all over again

I know some defense companies still use VHDL in Europe; I wonder if some American ones (still?) do.

 

[SarahKerrigan]

I wonder if some American ones (still?) do.

They do!

 

[FlameTail]

Someone fixed the Windows Subsystem for Android because Microsoft refuses to — here's how you can do it

The Surface Pro 11 is finally good as a tablet

www.xda-developers.com

Sinclair was the first to discover that the actual reason WSA won't boot is because parts of it use Armv7 to run, which isn't supported by the Armv8 Snapdragon X series. Now, he's created a way to get Android apps up and running, and you can do it yourself.

Will the same issues pop up when the CPUs get upgraded to ARMv9? I guess not- since ARMv9 and ARMv8 aren't too dissimilar.