Arm Announces ARMv9 Architecture With SVE2

[hotaru]

Yeah ARM won’t be interesting to me until there is a standardized boot process similar to x86.

there already is one, since 2009 for 32-bit and since 2013 for 64-bit.

[sykobee]

It's a good idea to set a new baseline of ISA extensions as mandatory with a new v9 ISA. This doesn't necessarily replace v8 as an option in the market, as that will be more than enough for many SKUs going forward.

As for the Pi, it's cheap, so it uses the cheap-to-license older ARM cores, A72 in this case. It fabs them on a cheaper older process, 28nm in this case. It absolutely cannot be used as a comparison to x86 IPC / iso-performance.

Maybe the next Pi will be a 12nm SoC using A76s, but it depends on the licensing fee. It might go to a 2+4 A76+A55 for example. Faster A72s on 12nm would not be a disaster though. What it does need is a better GPU configuration, which 12nm would allow - hopefully at least doubling the compute and increasing the speed on top of that.

I wonder when ARM will have ARMv9 cores ready to licence? I expect they will be in the A8x range, e.g., A82. But the small core would also need to be done (narrow SVE2 implementation) but ARM already clutters the A6x namespace (A65), although I guess they could start with the A62. Maybe they will start at a round number this time, A80/A60. Edit: A9x would make more sense! e.g., A90+A60.

[PerformanceExpert]

Performance, already does. X1 beats current x86 cores clock for clock and competes decently core for core. Apple's numbers are higher. I've spent considerable time on Neoverse N1's; generally N1 compares favorably, core for core, with Zen2, though falls behind at higher core counts because the N1 vendors opted for undersized cache configs.

And that's what they're doing with a core that's effectively two - almost three, with Matterhorn coming out - generations old.

Clock speeds, who cares? Power6 was hangin' out at 5GHz back in 2008 and it still lost to Nehalems at 3.5.

Indeed, it's incredible how quickly Arm has taken the server performance lead. Ampere Altra beats AMD Milan by 1% on 1S SPECINT_rate, and that's using the relatively old Neoverse N1 core and despite having only 1/8th of the L3 cache!

And things are not looking good for x86 if we consider the much faster 128-core Altra Max and next generation Neoverse designs that come out this year... I'd expect Arm to take over much of the hyperscalers and a significant chunk of the overall server market. Graviton is already 15% of AWS and still growing.

[Dawn]

I wonder when ARM will have ARMv9 cores ready to licence? I expect they will be in the A8x range, e.g., A82. But the small core would also need to be done (narrow SVE2 implementation) but ARM already clutters the A6x namespace (A65), although I guess they could start with the A62. Maybe they will start at a round number this time, A80/A60. Edit: A9x would make more sense! e.g., A90+A60.

Based on history, May of this year is very likely for Matterhorn and its small-core friend - and shipping products around the new year.

[PerformanceExpert]

ARMv8-A doesn't mandate ARMv7-A compatibility. Fujitsu left it out, on their HPC CPU, though I don't know of any other such examples. I'll bet Apple could've done (and might do), since they have plenty of headroom to emulate v7 in software.

There are many ARMv8 cores that don't implement ARMv7 - iOS dropped 32-bit support in 2017 and the latest Ax cores no longer support it (there is no emulation - developers needed to recompile their applications for AArch64). Android will also become 64-bit only in the next few years. On the server side Centriq, Thunder X1/2/3, A64FX and Huawei's custom Arm cores don't support ARMv7.

[phuclv]

there already is one, since 2009 for 32-bit and since 2013 for 64-bit.

UEFI isn't quite commonly used in ARM. People usually use U-boot or similar things

But what ARM really needs is a standard BSP, so that each new board model doesn't require explicit support in the kernel. Imagine if every PC motherboard required its own kernel patches! That's where the ARM ecosystem is currently at!

yes the lack of a standard board model is what irritates me. That's also the reason why Android's long term support was so bad because Linux insisted drivers to be in-tree but no ARM vendors do that and use proprietary drivers instead so after some time the device becomes a brick without proper security updates

[WorBlux]

there already is one, since 2009 for 32-bit and since 2013 for 64-bit.

And used in how many consumer devices? A couple hobby SBC's and Microsoft's ARM surface (which was bootloader locked anyways).

The problem isn't that there aren't standards to pick from, it's that chip designers have no interest in long term support (seemingly aside from NXP, and the hobby community around certain Broadcom modules)

Though UEFI is pretty heavy for a lot of mobile and embedded applications as well. Though something with a few more features than busybox+device tree would be nice.

[carewolf]

There are many ARMv8 cores that don't implement ARMv7 - iOS dropped 32-bit support in 2017 and the latest Ax cores no longer support it (there is no emulation - developers needed to recompile their applications for AArch64). Android will also become 64-bit only in the next few years. On the server side Centriq, Thunder X1/2/3, A64FX and Huawei's custom Arm cores don't support ARMv7.

Please dont confuse ARMv8 and AArch64 aka ARM64. There are also ARMv8 32bit extensions, but they are useless without ARMv7 and earlier stuff.

[carewolf]

And for love of everything stop referring to AArch64 as ARMv8. It is not only wrong, it is contagiously wrong, causing confusion and nonsense.

[Vistaus]

I would be surprised if v9 achieves x86 levels of performance and clock speeds...

Doesn't v8 already do that with the Apple M1?