Discussion Future ARM Cortex + Neoverse µArchs Discussion

[soresu]

Good point by @gdansk
View attachment 104744
It's good news for ARM's future.

ARM has an IP release cadence.

As yet they do not actually produce their own SoCs (and likely never will).

So when you say ARM has a slower cadence, what you actually mean is they have a slower IP cadence, and the use of it by their server licensees is pretty up in the air.

Their licensees won't upgrade IP every generation due to SoC design costs, especially if the gains aren't worth the time/effort/cost.

Also unless the outcome of QC/ARM's court case ends up wiping out their custom cores that means a huge fraction of ARM SoC's on the market are no longer ARM Ltd CPU cores so their cadence doesn't matter nearly as QC's does.

QC's CPU core roadmap in the past (Scorpion -> Krait -> Kryo) has not been particularly aggressive, so we will just see how this plays out over the next 5 years.

 

[FlameTail]

Also ARM vendors have another bottle they can uncork:

Increasing frequency.

ARM cores typically ran at much lower clock speeds than x86 rivals. Since they are starting from a lower baseline, it means ARM core designers can increase the frequency much more easily than x86 vendors (who seem to have hit a wall around 6 GHz).

Since ARM cores also have much higher IPC, the same clock increase yields far more raw performance gain.

Apple has deployed this to deadly effect with the M4 (which currently reigns as the supreme king of ST performance). It seems Qualcomm is also going down the same path with Oryon.

ARM with their own Cortex seem to be a bit more conservative. They did raise the Fmax of Cortex X925 to 3.8 GHz, but we won't see that in phone SoCs. Only in a laptop SoC probably.

 

[FlameTail]

QC's CPU core roadmap in the past (Scorpion -> Krait -> Kryo) has not been particularly aggressive

Ooh. Could you give a history lesson on that subject?

 

[soresu]

Ooh. Could you give a history lesson on that subject?

Pretty simple.

Scorpion was their first v7-A custom core in 2008, Krait followed that in 2012, and then Kryo was their short lived and only (consumer market anyway) v8-A custom core in 2015/16.

That being said, ARM Ltd's own release cadence in the same timeframe as was not nearly so aggressive as it has been since A72 onward, with only 4 true high end cores A8, A9, A15 and A57 coming before that (even A57 -> A72 was a 2 year lead time).

Edit: I don't count A12/17 - it's basically trying for A15 v7-A performance at a time when 64 bit ARM was already coming into focus.

 

[soresu]

Qualcomm have to know that if their intent is competing with Apple, then they will have to have a very aggressive roadmap on both their P and E cores.

ARM Ltd's A5xx IMHO has ceased to be ARM's E core in their own PR, replaced by A7xx.

 

[soresu]

I do not think that Kryo perf was the reason QC folded to put A73 in SD 835.

I think ARM showed them their future roadmap, and an outline of their leapfrogging design team strategy.

Combined with Apple A7's performance relative to Kryo they realised that their custom core plans were not nearly aggressive enough in µArch targets or cadence to keep them competitive in the market.

 

[FlameTail]

ARM Ltd's A5xx IMHO has ceased to be ARM's E core in their own PR, replaced by A7xx

No doubt, influenced by Mediatek's radical decision to ditch A5xx in the Dimensity 9300. It worked out well.

Mediatek and ARM seem to have an excellent relationship. After the falling out with Qualcomm, and with Huawei Hisilicon getting nuked by sanctions... Mediatek has risen to become ARM'S premier partner.

The Cortex X925 was supposedly designed with input from Mediatek:

 

[soresu]

No doubt, influenced by Mediatek's radical decision to ditch A5xx in the Dimensity 9300

It's not radical at this point.

I'm not a fan of saying "look what Apple does by comparison..." but in the case of A5x and A5xx they did drop the ball.

A5xx is at least getting a new core every 2 years now, but it's too little, too late with too little improvement over A55 to begin with for A510.

Unfortunately I doubt that we will see a sudden dramatic increase in streamers using A7xx, it's pretty likely that they will remain mostly A5x/A5xx based for quite some time.

Even Google's "brand new" STB like 4K streamer is using an A55 based SoC - just sad when Amazon Fire TV cube 3rd gen is pushing 2 years old with A73 based SoC, and even that didn't change much from the 2nd gen before that.

 

[FlameTail]

Qualcomm have to know that if their intent is competing with Apple, then they will have to have a very aggressive roadmap on both their P and E cores.

I don't think Qualcomm has "real E-cores" in the Snapdragon 8 Gen 4 (ie: It's not a seperate uarch/ground-up design).

My guess is they are doing a ZenC kind of approach to create the E-core. Taking the P-core and modifying it a bit:
• Less domino logic/dark silicon
• Denser/more efficient process libraries
• Hence lower Fmax
• Less L2 cache
• Less SIMD units

They have done this kind of thing the last time they did custom cores with Kryo:

The Snapdragon 820

We'll have to see what the 8 Gen 4 brings.

 

[soresu]

No doubt, influenced by Mediatek's radical decision to ditch A5xx in the Dimensity 9300. It worked out well.

Mediatek and ARM seem to have an excellent relationship. The Cortex X925 was supposedly designed with input from Mediatek:
View attachment 104777

Copy any text and just put it in a quote box:



It takes up less posting space and honors the UI color scheme of the forum the user has set.

I'm on dark mode here and that white background/black text is just:

 

[FlameTail]

Copy any text and just put it in a quote box:

View attachment 104780

It takes up less posting space and honors the UI color scheme of the forum the user has set.

I'm on dark mode here and that white background/black text is just:

I am sorry! I am sorry!

I posted the screenshot so that everyone who sees it knows it's authentic, and it is from DCS.

 

[soresu]

I am sorry! I am sorry!

I posted the screenshot so that everyone who sees it knows it's authentic, and it is from DCS.

If you fear doubts of authenticity just link the post and say:

"You can find the post here for source."

Just highlight the word (like "here" above) you want to make a link and then use this:

 

[Doug S]

Even Google's "brand new" STB like 4K streamer is using an A55 based SoC - just sad when Amazon Fire TV cube 3rd gen is pushing 2 years old with A73 based SoC, and even that didn't change much from the 2nd gen before that.

Wow I didn't realize they skimped so much on the SoC, given the price they're selling it for I assumed they were using a phone SoC like Apple TV. Just sad it is 4xA55 cores for a brand new product in 2024, what a pathetic cheapskate move for something they're asking $100 for!

 

[SpudLobby]

Wow I didn't realize they skimped so much on the SoC, given the price they're selling it for I assumed they were using a phone SoC like Apple TV. Just sad it is 4xA55 cores for a brand new product in 2024, what a pathetic cheapskate move for something they're asking $100 for!

Tbh I think they need more scale, and the $100 probably limited them still. But yeah they should’ve just done a $130 one with 64GB, 6GB of RAM, and a Tensor G2/3.

If you’re already out of the $30-70 ballpark, just go for the full 130-200 like Apple essentially has been doing, and throw it some serious storage and a somewhat decent phone chip.

Really annoying by Google, agreed

 

[FlameTail]

Post in thread 'Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads'

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...akes-discussion-threads.2606448/post-41272560

With both Intel and AMD seemingly losing clarity on the direction they should take in client, I see a golden opportunity for ARM vendors to launch a ferocious assault on the PC market.

Still, I am highly doubtful whether 50% PC marketshare in 5 years, is achievable at all.

https://www.reuters.com/technology/arm-aims-capture-50-pc-market-five-years-ceo-says-2024-06-03/

ARM might hit a brick wall in IPC gains like Apple did post-Firestorm.

Also the spectre of the Qualcomm vs ARM lawsuit looms large. If that goes kaboosh, it will be damaging not only to Qualcomm but also the wider ARM ecosystem.

 

[moinmoin]

Post in thread 'Intel Meteor, Arrow, Lunar & Panther Lakes Discussion Threads'

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...akes-discussion-threads.2606448/post-41272560

I feel you wanted to quote a post there. To do that you press on that "+ Quote" in the bottom right corner of the post and then the "Insert quotes..." button below the reply field to insert previously quoted posts into your reply.

 

[FlameTail]

SIGGRAPH 2024 presentation on mobile graphics:

Moving Mobile Graphics

Moving Mobile Graphics is a half-day course providing a technical introduction to the state of the art in mobile graphics from practitioners at the forefront.

community.arm.com

The above page contains PDFs of the speakers' notes. Very nice read.

 

[FlameTail]

This is good stuff!!

My team within Samsung has worked on handsets using GPUs from every mobile hardware vendor. This year we will ship devices using at least three different GPU architectures, so I have a broad view of the current offerings.

Mali, Adreno and Xclipse!

One final point, the feature set functionality gap between mobile and desktop is actually very small today. For example, HW ray tracing is available from all mobile SoC vendors. We are seeing neural graphics techniques start to appear on mobile. There are a few features that aren’t prevalent yet for architectural reasons, but the gap is now small. The distinction gap between mobile and non-mobile is largely a numbers game i.e. can you fit content into the resource constraints, rather than a question of whether functionality is available at all.

Interesting.

Next up, we are going to spend a few moments talking about two common architectural models for GPUs. If this was 2021, I would have told you that all mobile GPUs are primarily tile-based GPUs, and desktop parts are immediate-mode GPUs. Today there is a bit more nuance to that. Tile-based GPUs still dominate the market, but some Samsung handsets are now based on an IMR.

Oho! So mRDNA does use Immediate Mode Rendering! I have been wondering this for a long time.

 

[soresu]

One final point, the feature set functionality gap between mobile and desktop is actually very small today. For example, HW ray tracing is available from all mobile SoC vendors

Oooof.

Whoever wrote this clearly doesn't want the reader to understand the distinction between a basic feature and the hardware level support for various different micro features that can make the difference between supporting software that relies upon specific API extensions.

 

[FlameTail]

Yes, I tripped on that part too. Sebastian Aaltonen has been highlighting how mobile GPU architectures are trailing behind desktop class architectures in various ways, particularly compute performance.

 

[ikjadoon]

Finally, most of the documents from the Summary Judgment motions from both Qualcomm and Arm have been uploaded by a kind soul.

I'll highlight what I thought were the juicy regions. Unfortunately, the actual juicy bits themselves are all redacted.

Arm's Claims in its Motion for Summary Judgment:

Arm's requests for Judgment for itself on 3 claims: Arm terminated the ALA properly; QC & NUVIA breached the termination provisions; Arm did not breach the termination provisions (a claim by Qualcomm).

1. NUVIA, namely co-founder & CEO Gerard Williams, "sought multiple concessions from Arm" because NUVIA's core was targeted at the "data centre cloud server market".
2. NUVIA & Arm signed an ALA and TLA in Sept 2019.
3. Neither NUVIA nor Qualcomm gave Arm advance notice of the NUVIA acquisition before they broke the news (relevant in that Arm claims it needs to give consent first).
4. Qualcomm then "repudiated the economic terms of the NUVIA ALA".
5. After Arm terminated NUVIA's ALA, Qualcomm admitted "in later correspondence" it was using the code & designs under "a then-terminated ALA" (i.e., NUVIA). Apparently, that admission triggered Arm's lawsuit.
6. Arm expands: "But even after termination of the ALA, Defendants continue to use RTL code developed by NUVIA under the ALA."
7. Arm concludes: "the ALA prohibits this post-termination use of the code developed under the NUVIA ALA. It unambiguously states..."
8. The NUVIA ALA is "governed by California law".
9. One sentence ends at "NUVIA's CEO was right." It's all redacted otherwise.
10. Defendants should have disputed the ALA termination in 2022, not just after the trial started. Notably, Defendants provided a termination certificate.
11. NUVIA requested a few features from Arm, called "CMN". Arm claims it did not use those feature requests in its own internal Arm designs, but "merely ran a 'diff to compare".
12. Defendants had claimed that Arm's adoption of features requested by NUVIA was a breach of the termination agreement; Arm says otherwise.
13. Arm expands that the TLA with NUVIA "permits" [something redacted] with this CMN feature request. Later documents mention a "CMN - Kampos".

//

Arm's expert witnesses:

• On December 20, 2023, Arm served the opening expert reports of its technical experts Dr. Robert Colwell and Dr. Mike Chen, remedies expert Todd Schoettelkotte, licensing expert Guhan Subramanian, and trademark expert Dr. Ravi Dhar.
• Arm seems to be pushing that monetary damages are not adequate (e.g., so Arm wants specific performance).
• Arm claims some QC experts gave their info too late and Arm couldn't respond; if Arm is allowed to respond, it will delay the trial.
• Arm sometimes claims some witnesses should be included b/c they are lawyers; other times, they should be excluded b/c they are not lawyers. ...we need a lawyer here to understand this section LMAO.
• I honestly skimmed this doc, as it got a little too in the weeds re: admissibility of certain expert testimony.

//

Arm's Concise Statement

• 19. Portions of the RTL code Nuvia wrote before March 16, 2021, for use in the Phoenix core are used in Qualcomm cores, including [redacted] and the Snapdragon X Elite and Snapdragon X Plus SOCs.

//

Exhibits

• Exhibit 3 is an AnandTech article! First-page comments were also included in the screenshot, LOL. See page 26.
• Pg 41: part of Nitin Sharma's deposition. Used to work at NUVIA. Now works at SiFive now. Some discussion on "Orion" vs "Oyron" cores, seemingly.
• Pg 66: part of Murali Annavaram's deposition. QC expert witness. Seems to have a bit of a dispute with the lawyer at one point.
• Pg 76: part of Manu Gulati's deposition. A NUVIA founder. Cores with a "Y" seemingly reference "the series of Qualcomm custom CPU cores". There are some derivative cores of Oryon, seemingly, but they kept the name as Oryon cores. Notably is given a light warming, "Just caution you not to reveal any attorney-client communication" near the end.
• Pg 98: part of Paul Williamson deposition. Arm executive. Claims Arm was not aware of the NUVIA acquisition before it hit the media. He was apparently the main letter writer to Qualcomm re: the dispute before the lawsuit.
• Pg 120: part of Gerard Williams deposition. NUVIA founder. They were still working on the server application after the Qualcomm acquisition.
• Pg 159: part of Ziad Asghar's deposition. Qualcomm executive. QC says the QC ALA "takes precedence" and "was a way to make sure that essentially we can close the acquisition quickly."
• Pg 177: part of Jignesh Trivedi's deposition. Seemingly worked at NUVIA, now at Qualcomm. I think. Virutally all redacted.
• Pg. 188: part of Ramakrishna Chunduru's deposition. Believe he works at Qualcomm, according to random Google hits. All redacted.
• Pg. 197: part of Pradeep Kanapathipillai's deposition. Worked at NUVIA, now Qualcomm. Mentions annual February meetings held by ARM called "ARM TAB" (Technical Advisory Board); these are technical discussions.
• Pg 238: part of expert report by Patrick F. Kennedy, Ph.D. Qualcomm's expert and part of the counter-claim by Qualcomm. Managing director at "Stout Risius Ross LLC". Arm inappropriately received some fees & royalties (re: the CNM QC claims Arm stole).
• Pg 249: part of Simon Segar's deposition. Ex-ARM CEO. Humorously, one email is called "YAASS - Yet another Arm server startup". Rest is redacted.
• Pg. 280: part of Guy Larri's deposition. Arm engineer. Mostly redacted.
• Pg 295: deposition of Vedaraman Geetha. NUVIA "technical staff". Humorously, was deposed once before: when? "I was working at Intel for a number of years, and then I think it was troll something patent case."

There's many more documents, but I need a breather, holy fudge. I'll definitely do Qualcomm, but Arm's documents were submitted first in the docket.

EDIT: swapped a "NUVIA" for a "Qualcomm"

 

[FlameTail]

The Arm architecture reference manual is,
you say, approximately 20,000 pages or something along those lines?

That is impressive.

 

[Nothingness]

That is impressive.

It's only 15000 pages long 😀

It's quite verbose, and contains pseudocode that is validated and can be automatically processed by tools which partly explains its length. It can be a pain to find anwers to specific questions, but everything is in there. Just ensure you have a very fast PDF reader.

 

[FlameTail]

SIGGRAPH 2024 presentation on mobile graphics:

Moving Mobile Graphics

Moving Mobile Graphics is a half-day course providing a technical introduction to the state of the art in mobile graphics from practitioners at the forefront.

community.arm.com

The above page contains PDFs of the speakers' notes. Very nice read.

One big takeaway for me from this is that 'Neural Graphics' techniques are coming to mobile. Think Nvidia DLSS style- Frame Generation and Super Resolution, based on Neural Networks that are hardware accelerated.