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

[FlameTail]

Come on, I did the math just above.

You took the baseline as Cortex X3?

 

[Nothingness]

You took the baseline as Cortex X3?

I think the process I used is clearly described: I picked the 2023 top phone, Galaxy S23 Ultra (so X3) then picked S24 Ultra to derive the generational uplift between X4 and X925.

 

[SarahKerrigan]

I think the process I used is clearly described: I picked the 2023 top phone, Galaxy S23 Ultra (so X3) then picked S24 Ultra to derive the generational uplift between X4 and X925.

Yeah, I think that's reasonable, and it does line up (ish) with the apparent 15% iso-clock gain over X4, I think.

 

[Nothingness]

Yeah, I think that's reasonable, and it does line up (ish) with the apparent 15% iso-clock gain over X4, I think.

That's still guessing, but that looks more or less what is expected.
The only X925 result on Geekbench DB is useless: https://browser.geekbench.com/v6/cpu/5827749

 

[FlameTail]

Another interview with ARM CEO Rene Haas;

Client Challenge

This man loves getting interviewed.

 

[coercitiv]

This man loves getting interviewed.

Sounds to me like ARM is really hungry.

 

[FlameTail]

Sounds to me like ARM is really hungry.

(Meme)

Spoiler

 

[ikjadoon]

According to https://browser.geekbench.com/android-benchmarks
S24 Ultra vs S23 Ultra is +14%
So if the comparison really is against S23 Ultra, that'd be ~19.6% above S24 Ultra.

And GB ST score would be ~2600 (iPhone 14 level).

Arm's GB6.2 1T IPC chart → the Cortex-X4 is likely Arm's baseline.

Step 1 - what is the X925 vs X4 GB6.2 uplift iso-frequency?

Using the handy Plot Digitizer website I've stumbled on, the GB6 IPC uplift is approximately +14.65% from the X4 → X925. This is the only time Arm states IPC vs X4.



Step 2 - so what are the actual scores?

Cortex X4 @ 3.390 GHz = 2287 Pts
Cortex X925 @ 3.390 GHz = 2622 Pts (+14.65% IPC claim via step 1)
Cortex-X925 @ 3.800 GHz = 2939 Pts (+12.10% claimed freq. target)

Step 3 - plug the data into the 2nd chart




We only "know" Bar 3 = Cortex-X925 @ 3.8 GHz = 2939 from Step 2. Thus, Plot Digitizer and a little arithmetic:

Bar 1 | 2023 Premium Android = 2286.87 (Bar 3 / 1.328166)
Bar 2 | 2023 Best-in-Class 3.8 GHz = 2890.91 (Bar 1 * 1.264413)
Bar 3 | Cortex-X925 @ 3.8 GHz = 2939
Bar 4 | Cortex-X925 @ 3.8 GHz + tweaks = 3110.14 (Bar 1 * 1.36)

Step 4 - which CPUs align with those scores?

Bar 1 estimate via Step 3 = 2286.87 (100%)
NBC | Cortex-X4 @ 3.390 GHz = 2287 (100.00%) ✅
NBC | Cortex-X3 @ 3.360 Hz = 2107 (92.13%)

It's the X4. Thus, Bar 1 | "2023 Premium Android" baseline at 1x is the Cortex-X4.

//

Bar 2 estimate via Step 3 = 2890.91 (100%)
NBC | Apple A17 Pro @ 3.780 GHz = 2930 (101.35%) ✅
NBC | Apple A16 @ 3.460 GHz = 2644 (91.46%)

By frequency alone, it's the A17 Pro. But the scores also nearly precisely match. Thus, Bar 2 | "Best-in-Class 3.8 GHz" is the A17 Pro.

So I believe @SarahKerrigan had it right from the get-go and the math lines up:

2023 Premium Android → TCS23 | Cortex-X4; Cortex-X4 phones launched in late 2023.

2023 Best in Class → A17 Pro launched in late 2023. Apple is the well-known "Best in Class" Arm uArch.

Conclusion

With slightly more accurate math, Bar 3 & 4 put X925 perf in GB6.2 perf quite high:

2,939 pts @ 3.8 GHz = 773 Pts / GHz for the Cortex-X925
3,110 pts @ 3.8 GHz = 818 Pts / GHz for the Cortex-X925 tweaked

From my previous chart,
M3 IPC = 774 Pts / GHz
M4 IPC = 844 Pts / GHz

 

[Nothingness]

Conclusion

With slightly more accurate math, Bar 3 & 4 put X925 perf in GB6.2 perf quite high:

2,939 pts @ 3.8 GHz = 773 Pts / GHz for the Cortex-X925
3,110 pts @ 3.8 GHz = 818 Pts / GHz for the Cortex-X925 tweaked

From my previous chart,
M3 IPC = 774 Pts / GHz
M4 IPC = 844 Pts / GHz

Thanks a lot!

Though this kind of computations is prone to errors, this seems to show Arm has caught up in "IPC" vs Apple. Now the question is: how high will Arm customers go with frequency?

 

[DrMrLordX]

this seems to show Arm has caught up in "IPC" vs Apple. Now the question is: how high will Arm customers go with frequency?

Doesn't say anything about power either.

 

[Nothingness]

Doesn't say anything about power either.

Yes, that was implied by my frequency question. I think freq will be limited by power in devices such as phones.

 

[FlameTail]

Thanks a lot!

Though this kind of computations is prone to errors, this seems to show Arm has caught up in "IPC" vs Apple. Now the question is: how high will Arm customers go with frequency?

Well, they have. X925 would have the same IPC as A17 Pro. And a stone's throw away from M4 (if you exclude SME).

Arm's GB6.2 1T IPC chart → the Cortex-X4 is likely Arm's baseline.

Step 1 - what is the X925 vs X4 GB6.2 uplift iso-frequency?

Using the handy Plot Digitizer website I've stumbled on, the GB6 IPC uplift is approximately +14.65% from the X4 → X925. This is the only time Arm states IPC vs X4.

View attachment 102206

Step 2 - so what are the actual scores?

Cortex X4 @ 3.390 GHz = 2287 Pts
Cortex X925 @ 3.390 GHz = 2622 Pts (+14.65% IPC claim via step 1)
Cortex-X925 @ 3.800 GHz = 2939 Pts (+12.10% claimed freq. target)

Step 3 - plug the data into the 2nd chart

View attachment 102204


We only "know" Bar 3 = Cortex-X925 @ 3.8 GHz = 2939 from Step 2. Thus, Plot Digitizer and a little arithmetic:

Bar 1 | 2023 Premium Android = 2286.87 (Bar 3 / 1.328166)
Bar 2 | 2023 Best-in-Class 3.8 GHz = 2890.91 (Bar 1 * 1.264413)
Bar 3 | Cortex-X925 @ 3.8 GHz = 2939
Bar 4 | Cortex-X925 @ 3.8 GHz + tweaks = 3110.14 (Bar 1 * 1.36)

Step 4 - which CPUs align with those scores?

Bar 1 estimate via Step 3 = 2286.87 (100%)
NBC | Cortex-X4 @ 3.390 GHz = 2287 (100.00%) ✅
NBC | Cortex-X3 @ 3.360 Hz = 2107 (92.13%)

It's the X4. Thus, Bar 1 | "2023 Premium Android" baseline at 1x is the Cortex-X4.

//

Bar 2 estimate via Step 3 = 2890.91 (100%)
NBC | Apple A17 Pro @ 3.780 GHz = 2930 (101.35%) ✅
NBC | Apple A16 @ 3.460 GHz = 2644 (91.46%)

By frequency alone, it's the A17 Pro. But the scores also nearly precisely match. Thus, Bar 2 | "Best-in-Class 3.8 GHz" is the A17 Pro.

So I believe @SarahKerrigan had it right from the get-go and the math lines up:

2023 Premium Android → TCS23 | Cortex-X4; Cortex-X4 phones launched in late 2023.

2023 Best in Class → A17 Pro launched in late 2023. Apple is the well-known "Best in Class" Arm uArch.

Conclusion

With slightly more accurate math, Bar 3 & 4 put X925 perf in GB6.2 perf quite high:

2,939 pts @ 3.8 GHz = 773 Pts / GHz for the Cortex-X925
3,110 pts @ 3.8 GHz = 818 Pts / GHz for the Cortex-X925 tweaked

From my previous chart,
M3 IPC = 774 Pts / GHz
M4 IPC = 844 Pts / GHz

Give this man an award🏆

 

[ikjadoon]

Thanks a lot!

Though this kind of computations is prone to errors, this seems to show Arm has caught up in "IPC" vs Apple. Now the question is: how high will Arm customers go with frequency?

Well, they have. X925 would have the same IPC as A17 Pro. And a stone's throw away from M4 (if you exclude SME).

Give this man an award🏆

Cheers, happy to share. 👍

The perf numbers are dependent on which X4 variant is Arm's baseline:

• 3.40 GHz X4 variant (Dimensity 9300+)
• 3.39 GHz X4 variant (Snapdragon 8 Gen 3 for Galaxy)
• 3.30 GHz X4 variant (Snapdragon 8 Gen 3)
• 3.25 GHz X4 variant (Dimensity 9300)
• 3.20 GHz X4 variant (Exynos 2400)

Doesn't say anything about power either.

Yes, that was implied by my frequency question. I think freq will be limited by power in devices such as phones.

Yes. To clarify, the analysis uses 3.8 GHz because it's Arm's only reference clock in its charts. Not that 3.8 GHz is the "right" frequency or the shipping frequency.

//

Unfortunately, we won't see power data for at least 6 months and, if prior gen. mid-cycle refreshes are any indication, peak smartphone frequency won't be known until H2 2025—a year to go.

With the X925 likely featured in upcoming Windows on Arm SoCs, I'd not be surprised for X925 designs to set the high bar on stock Arm core frequencies.

And, OEMs love bumping clocks well past the efficient part of the curve to beat XYZ competitor. Arm's confirmed Samsung / TSMC 3nm-class nodes should add safe headroom.

(Curiously, Arm gave no mention of any X925 cores on Intel 18A. Coming later? Delayed? Never?)

Even as Arm promotes 3.8 GHz, I'd be happy with sub-3.6 GHz implementations: esp. on 3nm-class nodes, surely we can have some fanless WoA designs finally...

// the little we have

Arm has some more unlabelled data—gotta love it. Peak to peak, X925 eats +22.6% power to gain +30.0% perf vs X4, but these curves may well be cropped.



Hooray for not labelling charts. While I'm curious, I'd also have liked to see energy data and any absolute data on their L3 additional power-gating.

 

[ikjadoon]

Another interview with ARM CEO Rene Haas;

Client Challenge

This man loves getting interviewed.

This interview has one interesting disclosure: Arm has finally straightforwardly & officially confirmed a "second generation" (aka non-Qualcomm) Windows on Arm SoCs.

We know there’s going to be some other [Windows AI PC] systems coming out in the upcoming years. So, while the first-generation systems are going to be interesting, the second generation systems are going to be even more [so]. And folks who bought the first ones are probably going to be a little bit green with envy when they see what the second ones look like.

1. Multiple (?) "second generation" Windows on Arm SoCs in development.
2. Launches are in "upcoming years".
3. Claims the "second generation" systems are more interesting.

#3 seems like a dig at Qualcomm's Oryon, without acknowledging Oryon will have been shipping for at least 6+ months to years earlier. Maybe Arm has some frustration at MS/QC's alleged exclusivity agreement, but whatever Arm ships will go head-to-head with with laptops equipped with M4 (4C+6c), Oryon (12C), Lunar Lake (4C+4c), and Kraken Point (Zen5 4C+4C APU).

Surely, Cristiano Amon, SoCs releasing in 2025 should be more interesting than SoCs released in 2024.

 

[soresu]

This interview has one interesting disclosure: Arm has finally straightforwardly & officially confirmed a "second generation" (aka non-Qualcomm) Windows on Arm SoCs.

1. Multiple (?) "second generation" Windows on Arm SoCs in development.
2. Launches are in "upcoming years".
3. Claims the "second generation" systems are more interesting.

#3 seems like a dig at Qualcomm's Oryon, without acknowledging Oryon will have been shipping for at least 6+ months to years earlier. Maybe Arm has some frustration at MS/QC's alleged exclusivity agreement, but whatever Arm ships will go head-to-head with with laptops equipped with M4 (4C+6c), Oryon (12C), Lunar Lake (4C+4c), and Kraken Point (Zen5 4C+4C APU).

Surely, Cristiano Amon, SoCs releasing in 2025 should be more interesting than SoCs released in 2024.

The only thing that comes to mind is probably talking about DLSS from nVidia offerings allowing for better gaming performance with a well tuned superscaling/AA solution.

Also nVidia SoC's with their GPUs would probably play better with Windows games that are basically predicated on nVidia and AMD hardware.

I wouldn't be surprised if QC have an uphill battle to get a decent legacy catalogue of x86 games to run even with Prism, not to mention the optimisations are all for NV and AMD GPUs so even if they technically run, they won't run well and anything older than a few years won't get patches at this point with perhaps the exception of stuff like GTA5 which won't stop being farmed by the dev until GTA6 has a decent player base, and maybe a DLC or 2.

 

[MarkizSchnitzel]

Couldn't 2nd gen simply refer to.. AI?

Surely in a few years there will actually be AI software useful for John Smiths, and NPUs will be more powerful still.
E.g. I expect to have a useful locally run assistant in a few years?

 

[FlameTail]

ARM is now getting a major new customer for their CPUs and GPUs.

GOOGLE PIXEL

The Tensor G5 will be fully designed in house by Google, and manufactured on TSMC 3nm.

Until now (Tensor G1-G4) were co-designed with Samsung LSI, and manufactured at Samsung Foundry.

 

[DrMrLordX]

ARM is now getting a major new customer for their CPUs and GPUs.

GOOGLE PIXEL

The Tensor G5 will be fully designed in house by Google, and manufactured on TSMC 3nm.

Until now (Tensor G1-G4) were co-designed with Samsung LSI, and manufactured at Samsung Foundry.

Uh, don't you mean TSMC is getting a major new customer?

 

[ikjadoon]

(Curiously, Arm gave no mention of any X925 cores on Intel 18A. Coming later? Delayed? Never?)

Correcting myself: Intel did give a promotional statement to Arm on the X925 / X725 launch, just like Samsung Foundry & TSMC:

Intel - “Intel Foundry is deeply engaged with Arm, and this announcement demonstrates progress on the Arm Client. We are collaborating closely on leading-edge technology nodes, including Intel 18A, to provide best-in-class power, performance, and area metrics and enable next-generation mobile SoC products based on Arm’s new Compute Subsystems (CSS) for Client.” – Suk Lee, Vice President and General Manager of the Ecosystem Technology Office, Intel Foundry Services

Samsung Foundry - “GenAI capabilities embedded in mobile devices are in major demand amongst customers, requiring top-tier silicon technologies integrated with leading computing solutions. Combining Arm’s Cortex-X925 CPU solutions with Samsung Foundry’s latest 3nm GAA process node delivers on this demand. Our long-standing partnership with Arm allowed for early and tight collaboration in the areas of DTCO and PPA maximization for an on-time silicon delivery that hit performance and efficiency demands.” – Jongwook Kye, Executive Vice President and Head of Foundry Design Platform Development, Samsung Electronics

TSMC - “The AI-optimized Arm CSS is a prime example of the collaboration between Arm and TSMC in helping designers achieve unprecedented levels of performance and power efficiency, pushing the boundaries of semiconductor innovation for AI. Together with Arm and our Open Innovation Platform® (OIP) ecosystem partners, we empower our customers to accelerate their AI innovation using the most advanced process technologies and design solutions.” – Dan Kochpatcharin, Head of the Ecosystem and Alliance Management Division, TSMC

//

Looking back at Neoverse V3 line: Cortex-X4 based or Cortex-X925 based? The timeline, to me, suggests X4 (e.g., Neoverse reveals have typically been after the consumer uArch reveal). Unfortunately, the numbers won't help much here: it's a mess as Arm vaguely claimed "+12% Integer performance" for V3 over V2 without explaining which test (SPECT rate or SPEC; Arm sometimes gives both #s) nor whether that is even iso-frequnecy.

Neoverse V1 (X1 based):
Neoverse V2 (X3 based): +13% SPECint2017
Neoverse V3 (?? based): +12% "Integer performance" (is this SPEC rate or SPEC? Surely not 2006? Shouldn't be GB6. Nonsense to avoid naming it, when Arm knows it)

But, there was a curious commit to the Linux Kernel a few weeks ago by Arm engineers, as noted by Phoronix

A number of Arm Ltd CPUs suffer from errata whereby an MSR to the SSBS special-purpose register does not affect subsequent speculative instructions, permitting speculative store bypassing for a window of time. We worked around this for Cortex-X4 and Neoverse-V3, in commit:

Now, later the SSBS workaround was added for the X2, X3, X925, A710, A720, and V2, heh. But that they initially found the vuln in both X4 & V3 makes it seem like their lineages are connected.

 

[soresu]

Neoverse V3 (?? based): +12% "Integer performance" (is this SPEC rate or SPEC? Surely not 2006? Shouldn't be GB6. Nonsense to avoid naming it, when Arm knows it)

Pretty sure V3/Poseidon is X4 based.

 

[FlameTail]

When will ARM cores move to wider vector pipelines?

Except for Neoverse V1, every other ARM core has had 128b vectors in recent history.

 

[adroc_thurston]

When will ARM cores move to wider vector pipelines?

Not for a long while.

Except for Neoverse V1, every other ARM core has had 128b vectors in recent history.

A64fx has a chungus 2*512b FMA setup.

 

[LightningZ71]

A64fx is about as niche as you can get these days...

 

[soresu]

A64fx is about as niche as you can get these days...

It is, but apparently Fujitsu's followup chip Monaka will be targeting a wider use case than just HPC/supercomputer stuff.

Possibly they want to leverage it into a datacenter chip.

It's not coming till 2027 though, so we will probably have another 2x Cortex X/Neoverse V generations of improvement before that, though I doubt we will see anything mind blowing on the SIMD front given they just added another 50% after 4 generations.

 

[soresu]

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

AnandTech Forums: Technology, Hardware, Software, and Deals

Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

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.