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

[Stewarts]

The main advantage of ARM is energy efficiency, which, by definition, x86 processors cannot match because of the architecture. The 80-core Ampere Altra has a TDP of 45 × 210 W, a clock frequency of 3 GHz.

[bug77]

This is not true, SMT or HT does not add execution units. The reason that SMT/HT does not help HPC is because these applications are bottlenecked by the FP execution units or memory bandwidth. Sometimes there is a performance penalty for HPC applications when SMT/HT is enabled, because of cache contention.

Ok, poor choice of words on my part. But as you noted, memory access can and will starve execution. And of course, the cache, I forgot to mention that, thanks for pointing it out.

[andrei_me]

How optimized is those tests for ARM?
Postgres in read tests shows that it scales nicely with threads but looks unoptimized for the ARM arch 🧐

[milkylainen]

That's expected. HT only offers extra execution units, but shares the front end. When there's intensive memory access, the front end becomes a bottleneck and the execution units starve.
When the data fits in the cache, HT can match the performance of physical cores though.

Eh? There are no more execution units available because of HT or some other SMT technique.
Execution units are shared. Front end decode performance is still the same, only that it can still decode one more flow of non stalled/contended context.
But maybe you meant "CPU" with execution unit?

[LoveRPi]

ARMv8 has a long way to go before people start targeting optimizations for it.

[vegabook]

Yet another set of benchmarks puts the lie to ARM being anything other than an architecture for the edge. These guys have had 25 years to show us competitive ARM server side and the fails keep coming. That both Qualcomm and AMD dumped their ARM server ambitions years ago (after working hard at it), and that Intel never even tried, really should have been a clarion call to anyone still dreaming. Amazon, with hundreds of billions of dollars of cash, in its second version, still can't get anywhere close. This is going to cement the general view that ARM just doesn't cut it for serious compute and probably never will.

[milkylainen]

The main advantage of ARM is energy efficiency, which, by definition, x86 processors cannot match because of the architecture. The 80-core Ampere Altra has a TDP of 45 × 210 W, a clock frequency of 3 GHz.

There is very little difference between the architectures regarding ISA efficiency in relation to power consumption.
If anything I'd say that the CISC frontend usually more dense, hence moving less data for the same instructions.
This is true if you're looking at anything above the really, really small CPUs where the actual decode takes a larger part of the implementation.

x86 is not frowned upon because of _ISA_ energy efficiency. It's because the ISA is a horrible legacy mess.
Also. Being CISC the frontend has an unbounded decode space (brr. horrible...).

The x86 power-hungry legacy (compared to ARM) comes from actually building CPUs that can churn a lot of data. It is not harder than that.
Because of the actual complexity today, the ISA frontend is a smaller and smaller part of the CPU.
Most CPUs do translations to internal representations fitting the actual uARCH better, advanced ARMs included.

Assuming CPU-uarch teams are not stupid (they aren't), you can extract basic performance data by looking at the bigger numbers today.
Manufacturing process, clock frequency, number of transistors, caches, type of design (Not comparing GPU to CPU f.ex).
If they are roughly the same, then they're going to have roughly the same power envelope. It's not going to be some magic sauce here.

As ARM systems are catching up to x86 in the data-churning space, so will their power envelope.
Transistors do not toggle easier because electrons overheard that they were in an ARM core...

[CochainComplex]

Don't worry about it, Amazon doesn't build these for consumers. They just need something better tailored for their usage patterns at a possible lower cost.

Also, if you're into CFD simulations, would you rather see a cat litter box commercial instead? :P

Haha litter box for cats

[milkylainen]

Yet another set of benchmarks puts the lie to ARM being anything other than an architecture for the edge. These guys have had 25 years to show us competitive ARM server side and the fails keep coming. That both Qualcomm and AMD dumped their ARM server ambitions years ago (after working hard at it), and that Intel never even tried, really should have been a clarion call to anyone still dreaming. Amazon, with hundreds of billions of dollars of cash, in its second version, still can't get anywhere close. This is going to cement the general view that ARM just doesn't cut it for serious compute and probably never will.

In a sense, yes. But it's not about ARM or x86. It's about all the R&D. It's about the best fabs, etc.
ARM is a IP company.
So they keep making cores and then you have a one hit wonder trying to build a server CPU out of a ARM IP base.
Usually they create a pretty good one-off something-something by piecing together stuff.
But people tend to forget that Intel & AMD has been doing this for what, 40+ years?

Also, people tend to forget that high-end CPUs like our desktop ones are also _high_ volume products.
A "we have built our own high end ARM and put it in a couple of thousand of servers" is not.
They will never compare in price if you're looking to recoup R&D costs over those thousands of units.
I guess they could heavily subsidize the sale of actual servers & CPU's to gain a foothold. But that is another story altogether.

[vegabook]

In a sense, yes. But it's not about ARM or x86. It's about all the R&D. It's about the best fabs, etc.
ARM is a IP company.
So they keep making cores and then you have a one hit wonder trying to build a server CPU out of a ARM IP base.
Usually they create a pretty good one-off something-something by piecing together stuff.
But people tend to forget that Intel & AMD has been doing this for what, 40+ years?

Also, people tend to forget that high-end CPUs like our desktop ones are also _high_ volume products.
A "we have built our own high end ARM and put it in a couple of thousand of servers" is not.
They will never compare in price if you're looking to recoup R&D costs over those thousands of units.
I guess they could heavily subsidize the sale of actual servers & CPU's to gain a foothold. But that is another story altogether.

It would be great if purity of intellect were the determinant of success. You seem to excuse ARM because x86 has a huge head start and enormous advantages of scale. I'll retort that x86 "figured it out" 20 years before ARM did. Intel did that. They designed a long lasting architecture 20 years before ARM did. And to be fair to AMD, they were innovative enough in other spheres of silicon design that they were the "natural" licensee when IBM insisted on a second source. So obliquely, their intellectual efforts then mean they get to own some x86 now. They earned it through intellect. And they've remained competitive when others have not.

ARM came late to a party where the rules in high performance had already been well established. They niched into lower power, good for them, but they also seem to have made decisions, rightly, that hobble them on high-power high-performance. Basically, Intel figured out high performance before ARM did, and therefore rightly, they win even today. The economics/scale are just a result of that and not the cause of it.