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

[soresu]

Arm would have a great core in the Cortex-X5 and +15% is a big enough to hype about, even if it is six months earlier than usual

The research article quoting Blackhawk's "5 year high IPC increase" was just days ago.

ARM 2023 IP announcements were end of May just as they were from 2016 -> 2021, rather than the anomalous June that the 2022 IP PR was.

So very probably just 5 months away from the paper launch.

Arm is pulling some theatrics in the post-IPO world (launch IPC may underwhelm in 6-12 months, Arm backtracks claiming its "early estimate didn't pan out"

This close to IP release it seems unlikely that their numbers would be significantly off.

They wouldn't be the first to try such a thing, but they should know from past examples #cough# Bulldozer #cough# that it doesn't play well.

 

[ikjadoon]

This is great. It's amazing you can do all those calculations, and then consistently put it forth in well formatted comments. I certainly cannot muster the energy or time to do all of that.

Ah, cheers. It's not ideal with all the thermal inconsistencies (and it makes sense why Andrei, while reviewing for AnandTech, used active cooling for his reviews of mobile SoCs).

But, really, it's Notebookcheck's work. Initially, I thought they truly dredged up ancient phones stored in a drawer, but one of their Cortex-A75 devices was reviewed 9 days ago.

Luckily, I worked on bit by bit between work, haha.

How much that accounts for is unknown (but in the low single digits unless you choose specific tests) but the more your benchmark components are sensitive to memory bandwidth (memory latency is unlikely to matter much in the smallish subtests GB6 is doing, especially at today's L3/SLC sizes) the more of an effect it will have.

On that note, I did find AnandTech's tests of Geekbench 5 (none on Geekbench 6) with Alder Lake DDR4 vs DDR5, which on 1T shows a very similar (<2%) result, even with +49% copy bandwidth on the DDR5 testbed (individual read/write +69%) vs DDR4.

DDR4-3200 CL22
DDR5-4800B CL40

I believe Geekbench 1T is not very sensitive to bandwidth.

The same should apply to Geekbench 6 (i.e., what Arm's Blackhawk perf claim is based on): identical 5950X system, just 1x8GB (single channel) vs 2x8GB (dual channel): 1T hardly budges, just 2.3% in favor of dual-channel.

//

Contrast that with web browsing, e.g., Speedometer 2.0 (+18.0%) or GB5 nT (+25.3%):

I was honestly surprised JavaScript tests were sensitive to memory bandwidth with such a notable win for the DDR5 testbed. I don't see that reproduced in other web benchmarks, either: DDR5 scaling or DDR4 vs DDR5.

The research article quoting Blackhawk's "5 year high IPC increase" was just days ago.

ARM 2023 IP announcements were end of May just as they were from 2016 -> 2021, rather than the anomalous June that the 2022 IP PR was.

So very probably just 5 months away from the paper launch.

This close to IP release it seems unlikely that their numbers would be significantly off.

They wouldn't be the first to try such a thing, but they should know from past examples #cough# Bulldozer #cough# that it doesn't play well.

Ah, great point. I remembered reading the news articles in "kind of June-ish" and thought, "OK, June is the 6th month, so it's 6 months!" but you're right that that's not accurate: we're actually just five months away.

Agreed; it'd be a sad turn of events for Arm to fall into that trap. I mean, I don't know how many people read the note, but here at AnandTech Forums, we'll definitely lambast them. That's too big a claim to try to walk back.

I'm also curious why only Moor Insight & Strategy reported it: no other analysts were invited to the meeting? Or they were, but didn't think this was interesting?

As @FlameTail mentioned, looks like Arm is at CES, but I didn't see any events on the public CES calendar, either...maybe an invite-only event?

 

[ikjadoon]

Arm confirmed the Moor Insight report today to The Reg,

A spokesperson at Arm told The Register: "I can confirm the information in the post from Patrick Moorhead is accurate. We will provide more details on Blackhawk when it launches later this year. No further comment at this time."

Unrelatedly, it mentions Samsung "moving" to Arm's stock cores, but Samsung's custom uArch program was disbanded since the final M5 release in 2020 (Exynos 990) and the M6 cancelled. Or is the M7 rumored? Samsung apparently denied it in March 2023. From the original research note, it's a little too optimistic, I think.

Samsung has had performance challenges with its custom CPU cores as well, so it’ll have some very big decisions to make on its next designs.

//

Don't want to get my hopes up too much, especially as we don't know power or who'll be shipping SoCs with X5 in 2025.

At least five months isn't too long for at least official information.

 

[ikjadoon]

Long interview with Rene Haas, Arm CEO, came out today by Stratechery. Just skimmed it, but a few things

• Confirms Qualcomm has exclusivity on Windows on Arm and it will expire at the end of 2024. CTRL+F "exclusivity that Qualcomm has with Windows times out"
• Transparent admission that in Arm licensing "the price doesn't matter" b/c ecosystem barriers prevent easy ISA switches. Some validity, sure, but why poke that dragon? It'll bite back at some point.
• Claims we'll see more product announcements of companies moving to Arm "over the next 12 to 18 months" and later "the growth is not just with Microsoft and AWS".

Some vague comments on the Qualcomm litigation, but can't parse anything useful.

 

[FlameTail]

Unrelatedly, it mentions Samsung "moving" to Arm's stock cores, but Samsung's custom uArch program was disbanded since the final M5 release in 2020 (Exynos 990) and the M6 cancelled. Or is the M7 rumored? Samsung apparently denied it in March 2023. From the original research note, it's a little too optimistic, I think.

Interesting, very interesting.

There was this leaker on X- he wasn't the most reliable leaker, and often said some crazy stuff. He had a leak about Samsung's 2025 Exynos chip (Exynos 2500?).

He said it uses semi custom ARM cores, something allegedly called Cortex X5c and some stuff. It seems this is something like an Built-On-Cortex program on steroids. You know the BoC program? It allows ARM Cortex core licensees to make minor tweaks to ARM's Cortex cores in their implementation.

Let me see if I can find his tweets.

 

[FlameTail]

Long interview with Rene Haas, Arm CEO, came out today by Stratechery. Just skimmed it, but a few things

• Confirms Qualcomm has exclusivity on Windows on Arm and it will expire at the end of 2024. CTRL+F "exclusivity that Qualcomm has with Windows times out"
• Transparent admission that in Arm licensing "the price doesn't matter" b/c ecosystem barriers prevent easy ISA switches. Some validity, sure, but why poke that dragon? It'll bite back at some point.
• Claims we'll see more product announcements of companies moving to Arm "over the next 12 to 18 months" and later "the growth is not just with Microsoft and AWS".

Some vague comments on the Qualcomm litigation, but can't parse anything useful.

That's one solid chunk of a juicy interview.

 

[SpudLobby]

https://x.com/hjc4869/status/1706155380924142004?s=46

Gonna go ahead and bet that when he says two A17 Pro p cores = one X Core (prob X4) on area, he is including the L2 with the X core but not the L2 on the A17 Pro P cores.

Like, the Dimensity 9000 X2 was ~ 2.10 mm^2 with 1MB of L2 and Avalanche with just L1 was 2.55mm^2.

Say the scaling from Avalanche or Everest in A16 (similar sizes ish) is ~ .65 their current size with N3B, you’d get about 1.65mm^2 (with no L2) and then * two, ~ 3.315mm^2.

And the X4 has two generations of change from the X2 and is wider with 2MB of L2 from Qualcomm at least — and on N4P not N3 so I could see it being like 3-3.5mm^2 including that.

I do not really see how it could be the case that two A17 P cores on N3B are the same size as one Arm X core on N4P if neither have L2 included. The Apple cores are straight up just as big on width and have much larger L1 structures by a factor of almost 3, larger ROBs etc.

Even if so it’s because of N3B.

Wild to think about Apple p core sans L2 being in the 1.5mm^2 +- some range though. We’ve come a long way.

 

[FlameTail]

A17 P-core is like only 2.2 mm².

Also IIRC, P-cores in M chips are slightly larger than the ones in A chips, to enable the higher clock speeds. Can't confirm if this is true. Can someone take a measuring stick to this image?

M3 is 146 mm².

 

[soresu]

I'm also curious why only Moor Insight & Strategy reported it: no other analysts were invited to the meeting? Or they were, but didn't think this was interesting?

That article isn't the first time I've seen something about Blackhawk being a big step forward - the Korean URL link which originally called it Logan mentioned something similar like "biggest IPC jump in X series"

I guess we'll see soon enough.

I just hope that we get an SBC style implementation soon whatever it is, why are we still stuck on 5-6 year old A76 based SBCs in 2024 😭

 

[FlameTail]

What is SBC

 

[soresu]

Found this on Youtube....

 

[soresu]

What is SBC

Single board computer.

Basically Raspberry Pi or Odroid like super smol computers.

 

[poke01]

That article isn't the first time I've seen something about Blackhawk being a big step forward - the Korean URL link which originally called it Logan mentioned something similar like "biggest IPC jump in X series"

I guess we'll see soon enough.

I just hope that we get an SBC style implementation soon whatever it is, why are we still stuck on 5-6 year old A76 based SBCs in 2024 😭

If this X5 turns out to be powerful than Qualcomm wasted billions by buying Nuvia

 

[poke01]

With licensing ending in 2024 and if Nvidia and AMD do get on board with ARM SoC using X5 in 2025 then it will very interesting.

 

[soresu]

If this X5 turns out to be powerful than Qualcomm wasted billions by buying Nuvia

Between OG Kryo, Mongoose 1-5 and the several ARM server cores that fell flat on their face it wouldn't be the first time that a lot of money was wasted.

 

[Doug S]

Also IIRC, P-cores in M chips are slightly larger than the ones in A chips, to enable the higher clock speeds.

Doubt it. The higher clock rate is easily explained by having a higher power budget and being a little further up the performance curve.

They may be a little larger because they are in a bigger cluster = more ports in the caches for snooping etc.

 

[FlameTail]

Between OG Kryo, Mongoose 1-5 and the several ARM server cores that fell flat on their face it wouldn't be the first time that a lot of money was wasted.

NOOOOOOOOOOOOO!

If this X5 turns out to be powerful than Qualcomm wasted billions by buying Nuvia

Current Cortex designs lag behind Apple and Oryon not only in performance, but also efficiency. They widened the core in X4 and the power consumption ballooned. X5 needs to bring not only a huge performance improvement, but also a huge performance-per-watt improvement.

I have no doubt ARM can deliver the former but I am skeptical about the latter. I think Oryon's saving grace will be it's superior efficiency.

The tragedy of Oryon is how significantly delayed it has taken to arrive to market.

 

[DrMrLordX]

I just hope that we get an SBC style implementation soon whatever it is, why are we still stuck on 5-6 year old A76 based SBCs in 2024 😭

SBCs are built to compete in a space occupied by Arduinos and Raspberry Pis. A new mobile-class SoC featuring at least one X5 core on an SBC would cost quite a bit more than an Orange Pi (or similar) and bring more compute power to the table than people "need" for their small embedded projects.

 

[soresu]

SBCs are built to compete in a space occupied by Arduinos and Raspberry Pis. A new mobile-class SoC featuring at least one X5 core on an SBC would cost quite a bit more than an Orange Pi (or similar) and bring more compute power to the table than people "need" for their small embedded projects.

Yeah I know, but back before Huawei and HiSilicon's future got kiboshed they were producing HiKey SBC kits using their Kirin SoCs that were far ahead of the competition at the time.

Hopefully when Mac mini like boxes start getting churned out by OEMs using SDXE or Mediatek/nVidia/Samsung equivalents they will be open and not completely locked to WoA.

 

[soresu]

Current Cortex designs lag behind Apple and Oryon not only in performance, but also efficiency. They widened the core in X4 and the power consumption ballooned

IMHO a significant part of this is that the ground up new core got delayed so they ended up extending the µArch that goes all the way back to A76 roots at their Austin design team more than once.

It wouldn't surprise me to find that X5 backtracks on µArch width a bit.

It wouldn't be the first time as they did with it A72 -> A73, and A15 -> A17 before that.

 

[ikjadoon]

Found this on Youtube....

He shares he had a meeting with Arm with some Q&A time. Interestingly, this quote at 2:50 goes a bit further than closing the gap:

"The goal here is literally that this Blackhawk CPU will outperform the leading edge from Qualcomm and Apple."

A bit clearer why he said "best smartphone CPU" in the research note, but I assumed he got walked it back with "closing the gap". Maybe just 49 (= closing the gap) vs 51 (= outperform).

Good to know he clarifies that Arm didn't share any power or cost figures, but he's assuming they will be reasonable.

//

There was this leaker on X- he wasn't the most reliable leaker, and often said some crazy stuff. He had a leak about Samsung's 2025 Exynos chip (Exynos 2500?).

He said it uses semi custom ARM cores, something allegedly called Cortex X5c and some stuff. It seems this is something like an Built-On-Cortex program on steroids. You know the BoC program? It allows ARM Cortex core licensees to make minor tweaks to ARM's Cortex cores in their implementation.

Let me see if I can find his tweets.

That is interesting. Might there be a few named variants of the X5? We have plenty of core variants due to Arm's customization options, but they seem averse to branding any of these particular configurations.

@soresu's OP post in this topic does mention an upcoming "Poseidon VN" to be situated between the V series and N series. Poseidon also seems to be the same Cortex-X5 generation (as it'll power the next "V3" core and should come with PCIe Gen6, so it lines up with 2025+).

So Arm might do that in consumer, too? Like a "Cortex XA-5" core that's cheaper to license or smaller area than the Big Kahuna X5?

I still feel like we're waiting for the other shoe to fall on Blackhawk: what's the gotcha here? Power? Area? Licensing? SoC bottlenecks?

//

I'd love to see some leftover smartphone mainboards repurposed into SBCs or even mini PCs. Somehow, these $550 phones end up with a Cortex-X3 core, even after a BOM includes a fancy screen, cameras, modems, etc.

 

[SarahKerrigan]

He shares he had a meeting with Arm with some Q&A time. Interestingly, this quote at 2:50 goes a bit further than closing the gap:



A bit clearer why he said "best smartphone CPU" in the research note, but I assumed he got walked it back with "closing the gap". Maybe just 49 (= closing the gap) vs 51 (= outperform).

Good to know he clarifies that Arm didn't share any power or cost figures, but he's assuming they will be reasonable.

//



That is interesting. Might there be a few named variants of the X5? We have plenty of core variants due to Arm's customization options, but they seem averse to branding any of these particular configurations.

@soresu's OP post in this topic does mention an upcoming "Poseidon VN" to be situated between the V series and N series. Poseidon also seems to be the same Cortex-X5 generation (as it'll power the next "V3" core and should come with PCIe Gen6, so it lines up with 2025+).

So Arm might do that in consumer, too? Like a "Cortex XA-5" core that's cheaper to license or smaller area than the Big Kahuna X5?

I still feel like we're waiting for the other shoe to fall on Blackhawk: what's the gotcha here? Power? Area? Licensing? SoC bottlenecks?

//

I'd love to see some leftover smartphone mainboards repurposed into SBCs or even mini PCs. Somehow, these $550 phones end up with a Cortex-X3 core, even after a BOM includes a fancy screen, cameras, modems, etc.

Named variants have happened before, namely A78C (which I don't recall ever really getting any design wins.)

 

[FlameTail]

"The goal here is literally that this Blackhawk CPU will outperform the leading edge from Qualcomm and Apple."

Wow. That is both exciting and worrying to me. Worrying because, as you all know- I am rooting for Oryon.

 

[ikjadoon]

Named variants have happened before, namely A78C (which I don't recall ever really getting any design wins.)

That's a good point. I'd never heard about it again, sans a recent rumor the A78C may power the Switch 2. I guess it'd make a bit more sense as NVIDIA seems like the A78 custom variants.

Initially thought of the die-size-optimized A720 that is also just called A720 without any moniker, even with ~10% lower perf than the standard A720.

And another variant, you're right: Arm also had a A78AE; it ended up in NVIDIA's SBCs.

//

Wow. That is both exciting and worrying to me. Worrying because, as you all know- I am rooting for Oryon.

Aha, no worries; Oryon will be first to market and Oryon appears as a very premium, very high-perf, low-power SoC: 12 P-cores with strong IP (GPU, NPU, mem. controllers, ISP, etc.). As a consumer CPU in a laptop, it'll be fantastic, I'm sure.

And, who knows if any Arm vendor will stuff an SoC with 12x Cortex-X5 cores or if any X5-packing SoC gets a design win on a Windows PC (the NVIDIA & AMD rumor notwithstanding).

I mean, if Blackhawk delivers, it should be good, e.g., Oryon's pricing may improve if Cortex-X5 succeeds, and another serious Arm CPU for Windows should encourage more WoA development.

 

[Doug S]

Wow. That is both exciting and worrying to me. Worrying because, as you all know- I am rooting for Oryon.

Why would you root for Oryon but against ARM? If you want to see high performing non Apple CPUs having ARM cores winning would be better than having Qualcomm cores winning, because a lot of OEMs will have access to them.