Discussion ARM unveils Cortex A78/X1

[Tabalan]

Hi,
this topic is about yesterdays ARM big announcement - Cortex A78 (which was expected by everyone) and Cortex X1 (which was surprise, I couldn't find any leaks or gossips about new, big core from ARM, closest thing I could find are some extremply sparse info about next year Mattlehorn core). Link to AnandTech article:

Arm's New Cortex-A78 and Cortex-X1 Microarchitectures: An Efficiency and Performance Divergence

www.anandtech.com

What surprised me is the fact that ARM decided to unveil Cortex X1 now. Currently, in mobile you have low performance Cortex A55, high performance A78 (A77 is described as "Leadership performance and efficiency for 5G mobile solutions ") and even higher performance X1. From my point of view, we lack medium core, which cxould be used in mid range SoC (Snapdragon 600, Kirin 600 and so on). I'm aware that Cortex A78 couldn't have same performance as A77 (but lower area and power consumption, shifting into middle core), because it would force ARM clients into noticeably bigger and more power hungry X1. This (pretty much forcing into X1) could negatively reflect on ARM-client relationship in future.
Next year ARMv9.0 should be announced, so, honestly, I expected new cores (huge, medium, small) to be released then. I think new small core will be mandatory, because there might be some problems with big.LITTLE combining ARMv8.2 and v9.0 cores (at least I assume so).

Another things, it's interesting to whether manufacturers will decide to go for X1 or stick with A78. I think all of them might jump on X1 train to not be left behind in benchmarks ("...but new Exynos is so slow, new Kirin/Snapdragon is 20% faster") with tri-cluster solutions (A78 would act as medium core, X1 would replace current A77 with higher clock speed), while mid-high (Kirin 800, SD700) series will adopt A78 + A55 combo.

What are your thoughts on this?
<<I'm just tech enthusiast, so pretty sure I'm wrong somewhere. Point it out or disregard it >>

Edit:
"What’s really shocking here is how close Arm would be getting to Intel and AMD’s current best desktop systems in terms of performance. If both incumbent x86 vendors weren’t already worried about Arm’s yearly rate of improvement over the last few generations, they should outright panic at these figures if they actually materialize – and I do expect them to materialize. "
Love this part. Considering vast majority of people use browser, Office and some minor programmes, is it pretty much safe to say we reached the point where we can shift from x86 to ARM? Yes, pro software and games would stay on x86 for now, but everyone else could migrate to ARM without any problem.

 

[Tabalan]

I already mentioned that 50-60% A76 to X1 boost when I said it was similar to the Cosmos -> N1 IPC uplift.

Under some SIMD heavy loads such as encoding video the X1 core should near double the performance of A77, let alone A76 when you account for the doubling of execution units from 2 to 4.

Considering the sheer amount of video processing that is necessary today in servers that isn't a small thing at all - I'm sure there are various other uses for heavy SIMD code in servers too.

Yes, but according to your slide N1 to N2 should bring 30% perf increase. That number aligns with Cortex A76->A78 performance uplift. Ofc, ARM can change their mind and give us N2 based on X1, but that should double performance uplift from 30% to 50-60%.

Also, one more thing. December last year Nvidia announced Tegra Orin (not released yet), which was supposed to be using ARM Hercules cores (A78) and next gen GPU (Ampere). I wonder if they stick to A78 or switch to X1.

 

[soresu]

Yes, but according to your slide N1 to N2 should bring 30% perf increase.

I took the Neoverse promise to be the same as ARM's promise to provide 15% generational IPC increase for the big cores after A76 - ie minimum, not maximum.

That's just good PR - under promise, over deliver.

Also, one more thing. December last year Nvidia announced Tegra Orin (not released yet), which was supposed to be using ARM Hercules cores (A78) and next gen GPU (Ampere). I wonder if they stick to A78 or switch to X1.

They mentioned Orin again very recently during their Ampere announcements, nothing about a different core was mentioned - if it was the case it would have been announced in lockstep with the X1 announcement most likely.

It's possible that even the A78 cores are surplus to purpose beyond a strong backbone for Orin - the high number of cores (12) suggests a desired thread count, rather than sheer compute grunt for all those sensor processes that need routing to the Ampere GPU in an autonomous driving scenario.

I expect future ARM based NAS systems will continue to use Axx big cores going forward rather than Xn cores - at least for the smaller systems that SMB's and consumers use rather than the huge server racks for enterprise and datacenters.

 

[moinmoin]

Why not, if the chip is a normal AoW ig.LITTLE design that only fires x86 cores up exclusively for tasks normally emulated on Arm. These tasks would remain on zen cores and have their own memoryspace.

I don't think this is a very good idea, I just don't get technical reasons why it couldn't really be done or why would it be so taxing battery life as some claim, if those processes never migrate between different core-architectures.

EDIT:
To further clarify what I meant (as people constantly seem to ignore it)

Don't think of it as a hybrid chip. Think of it more as a totally normal run-of-the-mill Arm big.LITTLE design (A55 + A78) with a dedicated x86 accelerator for emulation.

It's far from technically accurate but perhaps gets the point across.

My keyword was "seamless" which is what I wouldn't call what you are talking about.

Your line of thought is more similar to the concept of co-processors in the way FPU started out as part of the x86 platform.

 

[Tabalan]

I took the Neoverse promise to be the same as ARM's promise to provide 15% generational IPC increase for the big cores after A76 - ie minimum, not maximum.

That's just good PR - under promise, over deliver.

They mentioned Orin again very recently during their Ampere announcements, nothing about a different core was mentioned - if it was the case it would have been announced in lockstep with the X1 announcement most likely.

It's possible that even the A78 cores are surplus to purpose beyond a strong backbone for Orin - the high number of cores (12) suggests a desired thread count, rather than sheer compute grunt for all those sensor processes that need routing to the Ampere GPU in an autonomous driving scenario.

I expect future ARM based NAS systems will continue to use Axx big cores going forward rather than Xn cores - at least for the smaller systems that SMB's and consumers use rather than the huge server racks for enterprise and datacenters.

Ohhh, I'd love for ARM to over deliver, trust me. +60% perf 2nd time in a row? Yes, please. I'm super curious how such N2 would fare vs Icelake-SP Xeon and Zen2/3 Epyc. From article on Phoronix I assume N2 would be winning most of the time.

100+ Benchmarks Of Amazon's Graviton2 64-Core CPU Against AMD's EPYC 7742 - Phoronix

www.phoronix.com

About Orin - yep, even sticking to 12x A78 would bring much higher performance. Carmel was equivalent to Cortex A75 and Tegra had maximum of 8 such cores (there were versions with fewer).

 

[Tabalan]

By the way, there is no roadmap for Cortex A series past 2020. Couldn't find even single one. This makes This makes me wonder is ARM will rename/reposition (performance tier wise) Cortex A successors.
For servers we have roadmap till 2021 (Zeus in 2020, Poseidon in 2021).

 

[soresu]

By the way, there is no roadmap for Cortex A series past 2020. Couldn't find even single one. This makes This makes me wonder is ARM will rename/reposition (performance tier wise) Cortex A successors.
For servers we have roadmap till 2021 (Zeus in 2020, Poseidon in 2021).

Not surprising given that Enyo/A76 was completely out of nowhere, unless Ares was its original name prior to Neoverse happening that is.

It's not odd at all to keep the Axx branding as the X does sound separate and meaningful.

However - I do think that they need to correct the mistake of putting little cores in the same A5x as A57.

Given the rumoured change to v9-A is coming, and likely for both big and little - it would be wise to ditch any association with present v8-A cores completely.

Perhaps something like A17x for big cores and A 15x for little cores?

OTOH they could create a new name branding for big and little to separate them from Cortex X, similar to Sempron and now Athlon for AMD, or Celeron/Pentium for Intel.