Benchmarking Amazon's ARM Graviton CPU With EC2's A1 Instances

[FireBurn]

The frames per dollar result doesn't look right to me

[Michael]

The frames per dollar result doesn't look right to me

Keep in mind that for these on-demand cloud tests where the cost is hourly (rather than a fixed cost like graphics card / CPU), it's also factoring in the amount of time spent running each particular benchmark for the particular cost.

[vegabook]

Surprised and disappointed that EPYC is still getting comprehensively spanked by Intel, even after factoring in cost. Even with the optimistic 15-20% uplifts from EPYC Rome the battle looks tough for AMD. As for ARM, well, is anyone surprised that this server-fish-out-of-water gets taken behind the woodshed and beaten senseless, yet again? I have no clue why Amazon is pitching ARM in the data centre when we've seen for years now that it's absolutely hopeless in that environment, to the extent that Qualcomm gave up and AMD tinned its entire ARM division before even showing product, when it saw how crap it was.

[edwaleni]

Performance isn't the issue for me, its the pricing. Amazon needs to put a little more back into the pricing of these hosts, meaning they need to lower it.

This may be a toe in the water exercise for now to see how people will use the offering.

[Zola]

Considering how A76 indeed can deliver 2x performance of an A72 while being only around 30% larger in it's silicone footprint & how an A76 on 2.6 GHz based on the 7 nm TSMC first gen Uses only 2.5W of power we are already in the position to say how ARM architecture is largely more advanced than anything CISC can offer. With introduction of larger NEON SIMD area extensions next year ARM will take a leed in every way. If someone starts making the ARM server chips with a dirty cheap cost approach(in both designing & manufacturing) on let's say 22nm FD-SOI it will be a doomsday for Intel (finally).

[FireBurn]

Keep in mind that for these on-demand cloud tests where the cost is hourly (rather than a fixed cost like graphics card / CPU), it's also factoring in the amount of time spent running each particular benchmark for the particular cost.

Like I said the numbers still look way to high, frames per second per dollar per hour? Or is it frames per dollar?

[Weasel]

Considering how A76 indeed can deliver 2x performance of an A72 while being only around 30% larger in it's silicone footprint & how an A76 on 2.6 GHz based on the 7 nm TSMC first gen Uses only 2.5W of power we are already in the position to say how ARM architecture is largely more advanced than anything CISC can offer.

People like you put out this kind of bullshit every time and always get destroyed in benchmarks. x86 is superior, partly because it's CISC.

Not sure how hard for you guys to understand this is, but with how many companies all stick to this crappy RISC mantra, and they are all behind x86, should tell you something. There's 3 cases here:

• All companies investing in crappy RISC CPUs are beyond incompetent (unlikely).
• Intel is just a god when it comes to proper CPU design (unlikely).
• RISC is just shit compared to x86/CISC. (likely)

Pick 1.

(and please don't start with supercomputers, you guys clearly don't understand what they are)

[Zola]

People like you put out this kind of bullshit every time and always get destroyed in benchmarks. x86 is superior, partly because it's CISC.

Not sure how hard for you guys to understand this is, but with how many companies all stick to this crappy RISC mantra, and they are all behind x86, should tell you something. There's 3 cases here:

• All companies investing in crappy RISC CPUs are beyond incompetent (unlikely).
• Intel is just a god when it comes to proper CPU design (unlikely).
• RISC is just shit compared to x86/CISC. (likely)

Pick 1.

(and please don't start with supercomputers, you guys clearly don't understand what they are)

It's like that for people like you that have a shits instead of brain.
Hire you go shitbrain:

https://www.anandtech.com/show/13614/arm-delivers-on-cortex-a76-promises

By the way Apple Vortex is already faster than anything Intel have & it's RISC ARM core.
A76 is licensable IP core available as POP IP for TSMC 7 nm FinFET & so it's a DinamiQ which can be connected to a IO hub the same way Intel & AMD are doing so you don't need a big investment to make 64 core's ARM server CPU, even if it's only 75% of the performance of an Intel's core's (& matches AMD's) it archives this using only 25% power.
From a next year the ARM holding server core design will be licensable & with 512 bit NEON SIMD extension. It will be more then competitive.
CISC whose a shit 20 years ago & we got stuck with it thanks to Intel's uncompetitive practices but it will all fall down soon enough.

[Weasel]

It's like that for people like you that have a shits instead of brain.
Hire you go shitbrain:
https://www.anandtech.com/show/13614...x-a76-promises

I'm not interested in benchmarks between midgets. Being the tallest midget isn't exactly a feat, you know.

By the way Apple Vortex is already faster than anything Intel have & it's RISC ARM core.

Prove it.

A76 is licensable IP core available as POP IP for TSMC 7 nm FinFET & so it's a DinamiQ which can be connected to a IO hub the same way Intel & AMD are doing so you don't need a big investment to make 64 core's ARM server CPU, even if it's only 75% of the performance of an Intel's core's (& matches AMD's) it archives this using only 25% power.

Cool story, heard it every month for the past 5 years, and benchmarks always disappoint. Keep living in that delusional bubble.

[Zola]

Prove it.

OK then. Let's compare Vortex to Pentium G4400.
You might say Pentium is cut off & how it's not a fair comparison but in reality so is Vertex that doesn't utilise latest ARM architecture. Anyway it's two OoO six instructions wide RISC ARM vs OoO six instructions wide CISC from Intel with hyper trading. Both equipped with 256 bit SIMD's.

https://www.spec.org/cpu2006/results...111-38699.html

https://www.spec.org/cpu2006/results...214-38404.html

Now when you normalise scores to the frequency as Vertex runs on 2.5 GHz vs 3.3 GHz of Pentium you get a 31% difference meaning how Vertex is actually faster. Neither Apple nor Intel improved much in the last year period of time regarding it's IPC. Still Intel's chip has hyper thread advantage but loses in small margins.
Now let's talk about efficiency. The Vertex is based on the TSMC 7 nm first gen process which is rushed up so it has only about 20% higher density than TSMC's second gen 10 nm one. The Pentium on the other hand uses Intels 14 nm+ proces, in best case TSMC 7nm has 25% density advantage. Let's also keep in mind how CPU running on 3.3 GHz will duble it's power consumption to the one running on 2.5 GHz. The golden value for FinFET structures still is 2GHz. But Vertex SoC uses less than 5W of power while Pentium sucks around 50W, that's the order of magnitude advantage. Now let's cut the difference in those factors. 2x5(frequency difference) +2.5(manufacturing process difference)=12.5W vs 50W that's 25% power consumption for a litle bit better performance. This actually means how Vertex core is only ¼ the size of Intels one as their is no magic in match. That also means it's four time's cheaper to manufacture on the large scale without profit fees.

Now keep in mind that the A76 is actually even better performing design and more power efficient for it's size & the fact it's a four instruction wide OoO design.
The Kirin 980 results aren't really representive for A76 as Huawei/HiSilicon used EAV+ scheduler which diminish the performance a lot.

So in the end when you put comparable similar designs against each other ARM wins in a huge margins (4x efficiency with a slight performance advantage).