Microsoft arms Azure VMs with Ampere Altra chips

Microsoft claims its latest Arm-powered Azure virtual machines can provide up to 50 percent better price-performance than similar instances using x86 processors. The computing giant announced Monday it is now previewing new D- and E-series VMs that use the Arm-compatible Altra server processors from chip startup Ampere …

[Nate Amsden]

what

Who would ever realistically compare a 8 core CPU with hyperthreading against a 16 core CPU (hyper threading or not)? Also the article makes it sound like the cost of a 8x x86 CPU(with HT enabled) is same/similar as a 16 x x86 CPU VM. I assume this is not the case(I have never used Azure, the fixed allocation models of all of the big public clouds have been a big turn off for me starting ~12 years ago so I really haven't paid much attention to them over the years).

Things would be a lot simpler if they just spit out some numbers from some benchmarks to compare the systems. Benchmarks are of course questionable by themselves but the performance claims being made here seem even more vague than benchmark numbers.

However if a single modern ARM CPU core can compete with a single modern X86 CPU core in server workloads that would be interesting, historically anyways it seemed ARM's designs were for just tons of cores on the chip(more than the standard x86 anyway), as an aggregate they may very well be on par with x86 (historically again they have had similar power usage from what I've read, that being 150W+/socket), but you're not comparing core-to-core performance( because the chips don't have the same number of cores - which in many cases doesn't matter I just mention that because the article seems to focus in on core-to-core performance).

Never personally been a fan of hyper threading myself mainly because it's not easy to assume how much extra capacity those threads give(but I haven't disabled it on any of my systems I just measure capacity based on actual cores rather than some funny math to adjust for extra threads).

[Charlie Clark]

Re: what

I think high-end x86 still outperform ARM in some single-threaded disciplines, but ARM has the edge in many other areas due to the degree of hardware acceleration, and, with Apple at least, faster memory.

And I agree with you, if anyone is to criticise the marketing, then hyperthreading is the place to start: any advantages disappear immediately for CPU intensive work.

[matjaggard]

Re: what

According to the CPU charts at cpubenchmark about the 4th fastest (depending on whether you count K models separately) single core available is an ARM.

[diodesign]

VM costs

"the article makes it sound like the cost of a 8x x86 CPU (with HT enabled) is same/similar as a 16 x x86 CPU VM. I assume this is not the case"

It likely costs more. We've updated the article with Ampere's POV on this.

C.

[mark l 2]

I suspect that all these cloud providers are going to try and push more ARM VMs going forward because the cost per Watt is lower than x86 and with energy prices sky rocketing that is starting to make a big difference for them.

Oracle Cloud are letting you play around with a VM with 4 Arm Ampere A1 cores and 24 GB of memory on their always free tier. Which is a really decent spec for a free VM if you can put up with it being offered by Oracle.

[Ken G]

Which is a really decent spec for a free VM

if you can put up with it being offered by Oracle.

if you have an exit strategy to another host.

[Charlie Clark]

Many data centres now have their own energy generation units (usually renewables because these are cheap and tax efficient) on site. But the simple economics of ARM systems with better per core performance in cheaper and smaller packages match customer expectations better. Virtualisation, essentially timesharing, makes sense on individual machines, but when you're billing for use, dedicated resources make more sense.

[matjaggard]

Energy use is still extremely important to them, if only because of the cooling required for each watt of power the CPUs are turning into heat.

[orly_andico]

AWS Graviton2 is similar

AWS says up to 40% better price-performance for Graviton2 versus Intel.

They are comparing vCPU to vCPU, so indeed comparing say an 8 core Intel, with a 16 core ARM64.

BUT: that 16 core ARM64 is still **CHEAPER** than the 8 core Intel. So it is a valid comparison.

Power draw is also much less (even with twice the cores) so sustainability wise ARM is a win.

It's basically the same story with Apple M1. Core to core the M1 is actually slower than Intel but at half the power draw. If you run coremark on the M1, the raw CPU performance is significantly less than Intel: the performance parity is from the on-module RAM.

[Nate Amsden]

Re: AWS Graviton2 is similar

Graviton2 is a different situation I think. I believe that chip is designed by Amazon, means they reap the benefits more of vertical integration(mainly cost savings, not having to pay higher margin costs to another supplier).

[Anonymous Coward]

Meaningless?

The CPU wars are filled with clickbait headlines and vague claims based on the manufacturers' tests. I just read another article (not on ElReg, mea culpa) about Intel having the fastest gaming chip.

Ultimately, it always boils down to your individual use case and your computing infrastructure (cloud and on prem).

Still, these articles make for good popcorn reading.