ARM performance vs x86 cores - another insight

[watercooled]

When I was reading through Anandtech's article on previewing the Galaxy S20 performance, I noticed it has some x86 comparisons in the SPEC charts. Of course, it's an older synthetic benchmark so it wouldn't be reasonable to use it as a sole data point in comparing two completely different architectures, however it's still very interesting IMO. And especially so given the vastly different power envelopes.

https://www.anandtech.com/show/15609...rmance-preview


ARM cores, even those on mobile devices are IMO very impressive right now. When comparisons like this have been made in the past, some particular people would dismiss any comparisons as useless and anything non-x86 (or non-Intel) as slow - no comparison was good enough, making it effectively (or you could say conveniently) impossible to compare them. And to be clear I'm not claiming the polar opposite; I understand the limitations of replying on a single synthetic benchmark, even if it is a decent set of data points.

Another article worth reading is the one on Amazon's Graviton CPUs: https://www.anandtech.com/show/15578...-intel-and-amd
Sadly there's no Zen2 in there for comparison.

[CAT-THE-FIFTH]

I don't really read too much into those results,ie,they are almost low clocked single threaded best case scenarios,a lack of SIMD testing and Anandtech conveniently ignored Zen2 when reviewing Graviton 2. The problem is Intel due to 10NM problems is stuck with a relatively old core design,ie, Skylake for nearly 5 years. AMD Zen2 is going to be replaced by Zen3 this year. The AT article is talking about Graviton 3 in 2022,but seemingly has forgotten to add,by then Zen4 will be probably out,and Intel will probably be on a non-Skylake design by then. Basically as usual AT seems to base most of what its talking about Intel,and underplays what AMD is doing in all of this,only adding one line about what AMD is doing at the end.

Another problem is a lot of these ARM based devices are always on the best cutting edge nodes,so can throw transistors at the problem,and are still on large monolithic dies. Intel and AMD tend to lag behind on nodes,but are also better prepared if there are problems too.

A lot of the ARM based designs are orientated from the get-go to low power and relative low clocks(and wide designs) which work well for the mobile devices they are orientated towards. They are also not held back by having to incorporate years of legacy extension support,and are not held by legacy OSes and applications.For example in the case of some of the Apple core tests,Apple has the advantage of a closed ecosystem which can extract the maximum performance out of the CPUs,even to the extend I heard they profile certain applications specifically(I think it is the Adobe ones IIRC).

AMD moving to chiplets is really a gamechanger when it comes to yields and binning. There is no free lunch with these standard N1 ARM cores in Graviton 2 as these are not small CPUs,and the problem at least in HPC is there is always a disconnect between seeing wide adoption and launches,unlike even things such as Zen2,ie,they end up competing with later products from Intel and AMD. Also when you come to think of it,the X86 part of an "X86 core" really isn't taking up much die area.There is a lot of experience Intel and AMD have outside base core design aspects,which is probably why they have managed to dominate the HPC market. Remember even IBM and Sun Microsystems,quietly moved away from these areas when they had competitive cores.

Personally,I think ARM based systems such as the Fujitsu A64FX have potential for a greater impact for ARM in HPC areas than Gravition and not because of the actual cores,but because Fujitsu have actual experience in HPC,and is trying to do something revolutionary in its own right. Having a hybrid CPU/GPU as a repeating unit in a homogeneous system,with their own low power connecting fabric,its a simpler system,and probably more easily scalable and efficient system than having 1000s of heterogeneous CPUs/GPUs in a supercomputer.

[DanceswithUnix]

If you haven't had a play with the latest Raspberry Pi yet, I recommend you find whatever lame excuse you need to buy one

Those A72 cores are rather capable, and now really cheap. It will be interesting to see where this goes, I think the ARM eco system is being pushed up in performance by RISC-V starting to eat at the low end of their market.

You still see two main problems with ARM systems though: thermal throttling (hello Pi4) and poor ram performance. That makes the cores look bad, like someone benchmarking a 3900X with a single stick of 2133MHz DDR4.

[watercooled]

I didn't interpret it as Anandtech having any sort of motive, rather making some comparisons based on what they have access to. The chart I posted from the mobile SoC include desktop (highly clocked) variants of both Skylake and Zen2.

As I said in OP, it would be interesting to see more metrics but it's obviously limited what they are able to compare, apples-to-apples, given the different ISA. If nothing else it gives us a sort of performance ballpark - I don't think it's reasonable to continue to dismiss ARM cores as some low performance gadget cores like I see certain people in comment sections refer to them as.

I'm not trying to make a point of whether ARM is/could be a substitute for all uses, rather admiring the performance possible in such a small power envelope. I know that's not all attributable to the ARM ISA either (in fact relatively little of it probably is in the larger cores), but they're some impressive core designs by both ARM and Apple.

I don't think application profiling is anything unique to Apple, and not only do core architects have to take this into consideration, but also what uses may be expected by the time the core is released, perhaps several years into the future in a rapidly-changing environment.

Nor is it really a competitive thing *right now* - the Graviton won't be sold to third parties as I understand it? And maybe apart from Apple, x86 on the desktop has such momentum behind it, I expect it would require more than ballpark performance parity to encourage any sort of shift. Macbooks on the other hand are a very interesting one to watch. I know that's something that's been rumoured and dismissed for a long time, and their cores do seem almost excessively performant for mobile devices.

@DanceswithUnit: I have a couple of Pis now, one recently bought for me as a gift but I've not had much of a chance to use it! I'll have to double-check the version and have a play with it.