Linux 6.1 Drops BF16 Support For Cortex-A510 Due To Hardware Bug

[phoronix]

Phoronix: Linux 6.1 Drops BF16 Support For Cortex-A510 Due To Hardware Bug

The 64-bit Arm (ARM64 / AArch64) architecture changes were merged last week for the ongoing Linux 6.1 merge window...

https://www.phoronix.com/news/Linux-6.1-ARM64-Updates

[coder]

Wow, that's pretty bad. I'm guessing some SoC-specific patches will be incoming, to selectively re-enable it on chips that don't have the issue. So far, ARM SoCs are ISA-symmetric (except for some ARMv9 cores not supporting AArch32), meaning most software will break if BF16 isn't disabled on all cores.

Heh, who could've possibly imagined cores sharing a FPU would've invited any bugs?

[patrick1946]

Wow, that's pretty bad. I'm guessing some SoC-specific patches will be incoming, to selectively re-enable it on chips that don't have the issue. So far, ARM SoCs are ISA-symmetric (except for some ARMv9 cores not supporting AArch32), meaning most software will break if BF16 isn't disabled on all cores.

Afaik Snapdragons are shared and Mediateks are not. Looks not good for my phone. 😏

[coder]

Afaik Snapdragons are shared and Mediateks are not. Looks not good for my phone. 😏

The practical implications are somewhat dubious, since BF16 is so low-precision that its applications are really limited to deep learning. Given that basically every SoC with A510 cores also has a NPU and GPU that will be both higher-performance and much more efficient, I doubt the CPU cores would've been used much for it, anyhow.

I think you probably couldn't have chosen a better feature in which for them to have a critical defect. Let's hope this well doesn't go any deeper, though. If SVE2 turns out to have some fundamental defect, that would be a far bigger liability.

[patrick1946]

The practical implications are somewhat dubious, since BF16 is so low-precision that its applications are really limited to deep learning. Given that basically every SoC with A510 cores also has a NPU and GPU that will be both higher-performance and much more efficient, I doubt the CPU cores would've been used much for it, anyhow.

I think you probably couldn't have chosen a better feature in which for them to have a critical defect. Let's hope this well doesn't go any deeper, though. If SVE2 turns out to have some fundamental defect, that would be a far bigger liability.

Yes, I share your context. I even think SVE will be not widely used in the near future. Googles new Tensor G2 has no SVE support. Clang afaik does not generate automatically SVE code. So I think my phone will be quite old before we get widespread SVE usage.

[hajj_3]

I wonder whether arm will quickly release a cortex-a511 to fix this bug?

[dragorth]

So, is the bug intermittent? This article makes it sound as if they have a reproducible test to show this is an issue. Since that is the case, can't they just create a test specifically for this CPU family that will check once to see if it needs to be turned off? Obviously this doesn't belong in the kernel, but a simple program that sets a variable to be passed to teh kernel at boot, run once at install or first boot would probably solve this, no?

[coder]

I even think SVE will be not widely used in the near future. Googles new Tensor G2 has no SVE support.

SVE2 is a mandatory part of the ARMv9-A spec. Anything using those cores will have it. The A510 is ARMv9-A.

[patrick1946]

SVE2 is a mandatory part of the ARMv9-A spec. Anything using those cores will have it. The A510 is ARMv9-A.

Yes but the Tensor G2 ist nit using the A510 but the A55. 😉