Arm Announces ARMv9 Architecture With SVE2

[WolfpackN64]

I can't help but smell money in the new architecture name. Does anyone else think that that perhaps the reason of not just enhancing Armv8 (adding another extension, as they have done previously), and instead designating this as a new architecture, arm (the company) gets new licensing fees? This could increase revenue by billions (and Son wants his money). I wonder how much Apple will need to pay to upgrade their Armv8 architecture license to Armv9.

I think it´s mostly a compliance thing. ARMv9 incorporates most v8 extensions up to ARMv8.6 and the new vector extension made by Fujitsu. If all these extensions become the new baselines, it´ll help the ecosystem and developers since they know what features they can expect from all ARMv9 processors.
Since Apple already incorporates many newer extensions, I don´t think they´ll be paying much extra. Then again, the ARM license costs are probably pocket change for Apple.

[blacknova]

Interesting, according to Ars they do not consider desktops or laptops as a target for the ARMv9. So we'd probably would not see proper ARM motherboards with all props we have for x86. Figures.

[Dawn]

Nope, it does not, or at least the Pi 4 processor.
The Pi 4 processor is only 15% faster than my 10-year-old low-end AMD machine.

Modern ARM cores are many times faster than the Pi4 processor. That is a 6-7 year old uarch running at unusually low clocks. Today's cores are 2-3x faster than the Pi4's Cortex-A72 at the same clock speed, and are commonly shipping up around 3GHz vs the Pi4's stock 1.5GHz. That's licensable cores - Apple's numbers are even higher.

I am who. I care.

Single-threaded performance

Right, but you don't need high clocks to get it, necessarily. Both Apple and ARM favor wider microarchitectures at lower clocks. At same-clock single-thread int, against Zen3, ARM X1 roughly matches - and Apple M1 significantly exceeds - current best x86 core performance (Zen3.) If they can get equivalent or better performance without clocking at 4+ GHz, that is perfectly viable, and Apple has shown that it can be done.

[tildearrow]

Right, but you don't need high clocks to get it, necessarily. Both Apple and ARM favor wider microarchitectures at lower clocks. At same-clock single-thread int, against Zen3, ARM X1 roughly matches - and Apple M1 significantly exceeds - current best x86 core performance (Zen3.) If they can get equivalent or better performance without clocking at 4+ GHz, that is perfectly viable, and Apple has shown that it can be done.

I know, but doesn't ARM always lag behind x86 in IPC?
Apple just did some sort of black magic to achieve x86 IPC on ARM.

[Jumbotron]

Modern ARM cores are many times faster than the Pi4 processor. That is a 6-7 year old uarch running at unusually low clocks. Today's cores are 2-3x faster than the Pi4's Cortex-A72 at the same clock speed, and are commonly shipping up around 3GHz vs the Pi4's stock 1.5GHz. That's licensable cores - Apple's numbers are even higher.



Right, but you don't need high clocks to get it, necessarily. Both Apple and ARM favor wider microarchitectures at lower clocks. At same-clock single-thread int, against Zen3, ARM X1 roughly matches - and Apple M1 significantly exceeds - current best x86 core performance (Zen3.) If they can get equivalent or better performance without clocking at 4+ GHz, that is perfectly viable, and Apple has shown that it can be done.

CORRECT !! I got scorched for declaring the "Age of ARM" has arrived and x86 should now be considered a "legacy" arch and ISA when Apple debuted the ARM based M1. Now in 2021 the number of ARM based Chromebooks is going up 6 fold or more. Apple will be x86 free with M1 and higher derivatives across every single platform including desktop this year. Microsoft is loading up on Qualcomm SoCs for their WinARM laptops and eventually desktops. And speaking of Qualcomm, they will debut Nuvia based Qualcomm SoCs by end of year 2021 ramping into 2022 that will go toe to toe with Apple's M1. Nvidia should have completed it's purchase of ARM by end of year and will bring Geforce Ampere GPUs to their ARM SoCs and the EU will be building out the European Continental Supercomputer with Next Gen ARM N1 SoCs called Poseidon which will have CCIX 3.0 and CXL 3.0 interconnects.

Apple will be at 3nm in 2022.

The rest of ARM and Qualcomm will all be at 3nm by 2023.

AMD will be struggling at 5nm but better off for production as everyone else (ARM based) will be at 3nm.

And poor Intel will be struggling at 7nm in 2023 with the rest of their line stuck on 10++++++++++++++++

[kvuj]

I think it´s mostly a compliance thing. ARMv9 incorporates most v8 extensions up to ARMv8.6 and the new vector extension made by Fujitsu. If all these extensions become the new baselines, it´ll help the ecosystem and developers since they know what features they can expect from all ARMv9 processors.
Since Apple already incorporates many newer extensions, I don´t think they´ll be paying much extra. Then again, the ARM license costs are probably pocket change for Apple.

I really hope so. It's about damn time we start using these awesome instructions in our CPUs. Maybe armv9 Linux distros / Windows build, which would be insanely cool (kinda like glibc hwcaps).

I think Arch already does that for board specific kernels but I'm not sure they recompile every package for every ARM level.

[rmfx]

Full compatibility with Armv8, but since Armv8 have full compatibility with v7, does that mean v9 has to support v7 bloat and more or they go toward a cleaner ISA with v9/v8 only??

[wizard69]

Interesting, according to Ars they do not consider desktops or laptops as a target for the ARMv9. So we'd probably would not see proper ARM motherboards with all props we have for x86. Figures.

What does Ars know. That web sites credibility is getting thinner every year.

The problem these days with ARM isn't performance but rather getting a manufacture to adopt and make hardware that isn't crap. I'd love to see a laptop or compact desktop based on a chip built from these designs. Build it from the ground up to run Linux and leave Microsoft to pound salt.

I literally ran out an got an M1 Air when Apple delivered and it lives up to every expectation I've had for an ARM based laptop. I have not had an Apple Mac in years but, for all of their wrongs I'm not going to object when they get something right. It is just incredibly nice to get great performance out of a passive cooled laptop. Now imagine a manufacture taking one of these new ARM chips and designing an open laptop

[Dawn]

I know, but doesn't ARM always lag behind x86 in IPC?
Apple just did some sort of black magic to achieve x86 IPC on ARM.

No. That's kind of what I'm saying. ARM X1 is basically Zen3-class integer ST "IPC" (I don't like that term - "iso clock performance" is better), and Apple M1 is significantly higher "IPC" than any currently available x86 microarchitecture. "ARM cores are tiny slow designs" hasn't been true for a while, arguably since before the Cortex-A76 came out; both ARM and its major architectural licensee (Apple) prefer wide, relatively low-clocked, very-high-IPC designs rather than chasing high clocks with narrower cores.

I've done a fair amount of testing myself of a 2.5GHz Neoverse N1 (grav2) vs 3.3GHz Zen2. The Zen2 usually wins at single-thread, but not by much. I often see higher "IPC" on the N1, which is an old core, than on Zen2. Newer Neoverse cores are expected to be another 40-50% higher.

Basically, old stereotypes are no longer as true as they once were. The ARM Austin design group has been executing consistently well. That could change - and at some point it probably will - but right now, ARM's high-end cores have absolutely been delivering. The improvements in ARM cores since the A72 and A73 are really impressive.

[ezst036]

I'm really excited. They need ATX motherboards and socketed cpus pronto!