Qualcomm revealed X Elite's benchmark scores

[leman]

I don't consider any PC to be "AI-enabled" unless it can run a model as good as GPT-4 locally at a similar token/s to ChatGPT. It's my own personal requirement.

So probably 2-3 more years.

Unlikely. RAM bandwidth just won’t be there. Maybe a heavily quantized model.

 

[senttoschool]

Unlikely. RAM bandwidth just won’t be there. Maybe a heavily quantized model.

In 2-3 years, I'm guessing that a computer using a chip in the class of a M6 Pro should be able to run it. The software side is moving really fast and small (7B) models are approaching GPT3.5 in output quality.

2-3 more years of software and hardware optimization should get us to GPT4 level locally for a M Pro class of devices. I'm betting that all chip makers are completely focused on improving the NPU/GPU for the next 2-3 generations of chips.

 

[Homy]

QC now have to start worrying about M4.

Apple Debuts Next-Generation M4 Chip

Apple today unveiled the M4 chip, its newest system on a chip for the 2024 iPad Pro lineup. Apple said that the M4 chip features a new...

www.macrumors.com

 

[MayaUser]

QC now have to start worrying about M4.

Apple Debuts Next-Generation M4 Chip

Apple today unveiled the M4 chip, its newest system on a chip for the 2024 iPad Pro lineup. Apple said that the M4 chip features a new...

www.macrumors.com

dont worry, we will get a new chart from them comparing with the M4, maybe they will invent some numbers for the upcoming M5

 

[rampancy]

I'm sure we'll expect all of the usual folks (especially the ones on YouTube) saying that Tim Cook is entirely responsible for this, that Apple is doomed, the X Elite is going to better than the M4, M5 and M6 combined, and that all of our Apple Silicon machines are now instantly junk, so we all have to toss them in the bin right now and get Surface Pro X's instead.

 

[donth8]

First M4 Geekbench results!!! Single core did improve as well compared to M3.

Geekbench Search - Geekbench

 

[MayaUser]

If thats true....Qualcomm has left the building and we will not get from them a chart Elite vs M4

 

[leman]

So much to "all CPU design talent has left Apple for Qualcomm"

 

[jido]

I think that M4 score is skewed because Geekbench is using the wrong metric for “object detection”… score is just too high.

 

[leman]

I think that M4 score is skewed because Geekbench is using the wrong metric for “object detection”… score is just too high.

It's not a wrong metric. Apple has finally opened up their matrix coprocessor to user software and GB6 can take advantage of that. It does push the score higher, but even if you remove that subtest you still have 10-20% better scores on average than M3 in the majority of subtests.

 

[Xiao_Xi]

Apple has finally opened up their matrix coprocessor to user software and GB6 can take advantage of that.

How can user software access the matrix coprocessor without any modification?

 

[leman]

How can user software access the matrix coprocessor without any modification?

The software obviously needs to support the extensions. Presence of specific functionality (like SME) can be detected at runtime, at which point the software can invoke an optimized routine.

Note: I am unsure how this is actually done in practice. On Apple platforms, one can use sysctl family of functions. I imagine that there is also a CPU control register that allows one to check for a feature, but I was unable to figure this out with a cursory search.

 

[Xiao_Xi]

On Apple platforms, one can use sysctl family of functions. I imagine that there is also a CPU control register that allows one to check for a feature, but I was unable to figure this out with a cursory search.

Maybe FEAT_SME?

Previous SoCs don't have it.

arm-cpusysregs/docs/apple-m1-features.md at main · lelegard/arm-cpusysregs

Access Arm64 CPU system registers. Contribute to lelegard/arm-cpusysregs development by creating an account on GitHub.

github.com

The software obviously needs to support the extensions. Presence of specific functionality (like SME) can be detected at runtime, at which point the software can invoke an optimized routine.

It is quite timely that the latest version of Geekbench has introduced support for Arm Scalable Matrix Extensions (SME) instructions.

Geekbench

www.geekbench.com

 

[leman]

Maybe FEAT_SME?

That's the one.

It is quite timely that the latest version of Geekbench has introduced support for Arm Scalable Matrix Extensions (SME) instructions.

Right? It is quite interesting, especially given the fact that no other ARM architecture (including ARM's upcoming Neoverse server cores) support SME.

 

[Xiao_Xi]

For those interested in an ARM-based PC laptops, Windows Report has posted photos of new Dell laptops.

Exclusive Images: This is Dell's upcoming lineup with Elite X & Intel Ultra CPUs

Dell is officially testing the Qualcomm waters by equipping its upcoming devices with Elite X processors, and we have exclusive images of them.

windowsreport.com

 

[Xiao_Xi]

It looks like a bunch of interesting internal Dell documents have been leaked.

The company claims that Snapdragon-based laptops will deliver 16.8 to 29.9 hours, depending on the screen type. To be precise, the FHD+ option at 29.6 hrs, QHD at 21.9 hrs, and OLED at 16.8 hrs.

Based on the 2023 estimates, Dell XPS with 10-core Nuvia (Snapdragon X Plus) would cost 145 USD, while the predecessor based on Core i7-1360P was valued at 293 USD.

Unfortunately, there is no information on the expected performance of Qualcomm's SoC.

https://videocardz.com/newz/snapdragon-x-series-chips-cost-only-half-as-much-as-intel-raptor-lakes-with-battery-life-up-to-98-higher

 

[crazy dave]

It looks like a bunch of interesting internal Dell documents have been leaked.


Unfortunately, there is no information on the expected performance of Qualcomm's SoC.

https://videocardz.com/newz/snapdragon-x-series-chips-cost-only-half-as-much-as-intel-raptor-lakes-with-battery-life-up-to-98-higher

This slide is all kinds of hilarious - the target customer is such corporate speak. I'm sure Apple has similar internal target customer descriptions that would be equal parts hilarious and horrifying to read. It's just funny to see it in black and white - what they think of us, you know beyond dollar signs.

Also, target market position as compared to the ... checks notes ... M2 Air and M1 Pro. Oh dear ... a little delayed getting out to market were we?

 

[Icelus]

Qualcomm Talks Up Their Linux Support For The Snapdragon X Elite - Phoronix

www.phoronix.com

Upstreaming Linux kernel support for the Snapdragon X Elite

Latest SoC is getting traction on Windows-based laptops, and we’re also consistently upstreaming to mainline Linux. See areas of focus and try our Debian installer.

www.qualcomm.com

 

[Xiao_Xi]

Chips and Cheess has attempted to describe the Oryon core through LLVM patches.

https://chipsandcheese.com/2024/05/15/qualcomms-oryon-llvm-patches/

list<Predicate> UnsupportedFeatures = !listconcat(SVEUnsupported.F,
                                                    SMEUnsupported.F,
                                                    MTEUnsupported.F,
                                                    PAUnsupported.F,
                                                    [HasPAuth, HasCSSC]);

llvm-project/llvm/lib/Target/AArch64/AArch64SchedOryon.td at 8aebe46d7fdd15f02a9716718f53b03056ef0d19 · llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - llvm/llvm-project

github.com

ARMCPUTestParams<AArch64::ExtensionBitset>(
            "oryon-1", "armv8.6-a", "crypto-neon-fp-armv8",
            (AArch64::ExtensionBitset(
                {AArch64::AEK_CRC,     AArch64::AEK_FP,      AArch64::AEK_PAUTH,
                 AArch64::AEK_FCMA,    AArch64::AEK_JSCVT,   AArch64::AEK_SIMD,
                 AArch64::AEK_RAS,     AArch64::AEK_LSE,     AArch64::AEK_RDM,
                 AArch64::AEK_RCPC,    AArch64::AEK_DOTPROD, AArch64::AEK_SM4,
                 AArch64::AEK_SHA3,    AArch64::AEK_BF16,    AArch64::AEK_SHA2,
                 AArch64::AEK_AES,     AArch64::AEK_I8MM,    AArch64::AEK_RAND,
                 AArch64::AEK_PROFILE, AArch64::AEK_CRYPTO})),
            "8.6-A")),

llvm-project/llvm/unittests/TargetParser/TargetParserTest.cpp at 8aebe46d7fdd15f02a9716718f53b03056ef0d19 · llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - llvm/llvm-project

github.com

 

[deconstruct60]

Right? It is quite interesting, especially given the fact that no other ARM architecture (including ARM's upcoming Neoverse server cores) support SME.

Merged SME support back in 2022

ARMv9 Scalable Matrix Extension Support Lands In Linux 5.19 - Phoronix

www.phoronix.com

and 2023

Linux 6.3 Preparing Support For Arm's Scalable Matrix Extension 2 (SME2) - Phoronix

www.phoronix.com

relatively limited footprint benchmarks like GB , SPEC , etc get run is processor design simulations.

Arm is not openly talking about Neoverse v4 but that doesn't mean it isn't being worked on. Neoverse v3 has SVE2. Kind of makes sense to roll that out first since it is a precursor SME.

 

[deconstruct60]

Chips and Cheess has attempted to describe the Oryon core through LLVM patches.

https://chipsandcheese.com/2024/05/15/qualcomms-oryon-llvm-patches/

list<Predicate> UnsupportedFeatures = !listconcat(SVEUnsupported.F,
                                                    SMEUnsupported.F,
                                                    MTEUnsupported.F,
                                                    PAUnsupported.F,
                                                    [HasPAuth, HasCSSC]);

llvm-project/llvm/lib/Target/AArch64/AArch64SchedOryon.td at 8aebe46d7fdd15f02a9716718f53b03056ef0d19 · llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - llvm/llvm-project

github.com

ARMCPUTestParams<AArch64::ExtensionBitset>(
            "oryon-1", "armv8.6-a", "crypto-neon-fp-armv8",
            (AArch64::ExtensionBitset(
                {AArch64::AEK_CRC,     AArch64::AEK_FP,      AArch64::AEK_PAUTH,
                 AArch64::AEK_FCMA,    AArch64::AEK_JSCVT,   AArch64::AEK_SIMD,
                 AArch64::AEK_RAS,     AArch64::AEK_LSE,     AArch64::AEK_RDM,
                 AArch64::AEK_RCPC,    AArch64::AEK_DOTPROD, AArch64::AEK_SM4,
                 AArch64::AEK_SHA3,    AArch64::AEK_BF16,    AArch64::AEK_SHA2,
                 AArch64::AEK_AES,     AArch64::AEK_I8MM,    AArch64::AEK_RAND,
                 AArch64::AEK_PROFILE, AArch64::AEK_CRYPTO})),
            "8.6-A")),

llvm-project/llvm/unittests/TargetParser/TargetParserTest.cpp at 8aebe46d7fdd15f02a9716718f53b03056ef0d19 · llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - llvm/llvm-project

github.com

SME is Arm v9 ... Given Qualcomm/Nuvia and Arm have been suing each other over that last couple of years how likely are Qualcomm going to be up to date on licensing?
That they are v8 is a surprise?

Nuvia was somewhat aiming at Neoverse v1/n1 maybe v2. BF16 would be a bigger deal back in the v1/n1 era.