Google Axion CPU With GCE C4A vs. AWS Graviton4 Performance

[phoronix]

Phoronix: Google Axion CPU With GCE C4A vs. AWS Graviton4 Performance

Last week Google announced the general availability of their C4A instances powered by their in-house Axion processors. I delivered launch-day benchmarks looking at the Google Axion CPU performance with the C4A instances compared to other Google Cloud instance types powered by Ampere Altra and Intel Xeon. In this article is a look at how the Google Axion processor performance compares to the competing Amazon/AWS Graviton4 processor.

https://www.phoronix.com/review/google-axion-graviton4

[coder]

No price-normalized results?

Current on-demand pricing [1] for r8g.12xlarge is $2.828/hr, but the spot price is just $0.6022. The difference is that a spot instance can get shutdown out from under you, with little advanced notice (2 minutes).

Current on-demand pricing for c4a-standard-48 is $2.155/hr (spot price: $0.5923 + $0.2688 for 192 GB = $0.8611, IIUC).
Current on-demand pricing for c4a-highmem-48 is $3.489/hr (spot price: $0.5923 + $0.2688 for 192 GB = $1.1299).

So, according to that, standard Axion scores 15384 geomean/$, highmem scores 9494 geomean/$, and the Graviton 4 scores 10713 geomean/$.
For the spot prices, standard Axion scores 38501 geomean/$, highmem scores 29316 geomean/$, and the Graviton 4 scores 50311 geomean/$.

For most purposes, I'd guess perf/$ matters more than outright performance (within reason), although if you don't need 384 GB and can deal with the spot instance issues, then m8g.12xlarge [3] is really a killer option at 136719 geomeans/$!

Notes:

1. I used the Northern Virginia datacenter prices for both service providers.
2. Since AWS EC2 r8g12xlarge has 384 GB of memory, it compares most directly to the the highmem Axion instance.
3. If you don't need that much RAM, the m8g.12xlarge is the same, but with only 192 GB of memory and costs $2.154/hr (spot: $0.2216/hr), which is suspiciously close to the c4a-standard-48 on-demand pricing (but way lower spot price).
4. For comparison, the m7i.12xlarge is the same as that, but with a 3.2 GHz Intel Sapphire Rapids and costs $2.419/hr (spot: $0.7448).

[Michael]

No price-normalized results?

Current on-demand pricing for r8g.12xlarge is $2.828/hr, but the spot price is just $0.6022. The difference is that a spot instance can get shutdown out from under you, with little advanced notice (2 minutes).

Current on-demand pricing for c4a-standard-48 is $2.155/hr (spot price: $0.5923 + $0.2688 for 192 GB = $0.8611, IIUC).
Current on-demand pricing for c4a-highmem-48 is $3.489/hr (spot price: $0.5923 + $0.2688 for 192 GB = $1.1299).

So, according to that, standard Axion scores 15384 geomean/$, highmem scores 9494 geomean/$, and the Graviton 4 scores 10713 geomean/$.
For the spot prices, standard Axion scores 38501 geomean/$, highmem scores 29316 geomean/$, and the Graviton 4 scores 50311 geomean/$.
For most purposes, I'd guess perf/$ matters more than outright performance (within reason).

Notes:

• I used the Northern Virginia datacenter prices for both service providers.
• Since AWS EC2 r8g12xlarge has 384 GB of memory, it compares most directly to the the highmem Axion instance.

Right no normalized results since particularly cross-cloud pricing comparisons can get rather cloudy depending upon all the variables and not all of the options being 1:1 across vendors... So in this case just focus on the performance results and those with their desired pricing metrics can in turn use the OB link or manually input their perf-per-dollar-per-hour and get their desired output.

[pWe00Iri3e7Z9lHOX2Qx]

Nice job Google. Now go make your / Samsung's consumer Tensor cores competitive. You were already meaningfully behind there, and the NUVIA based Oryon cores debuting in phones soon are really going to put a hurt on Tensor.

[coder]

Nice job Google. Now go make your / Samsung's consumer Tensor cores competitive. You were already meaningfully behind there, and the NUVIA based Oryon cores debuting in phones soon are really going to put a hurt on Tensor.

FWIW, it surprised me that the newest Pixel phones that just launched have a SoC made on a TSMC N4 node, while the Snapdragon Elite is already made on a second generation N3 node (N3E).

That said, caveat emptor:

https://www.androidauthority.com/real-world-snapdragon-8-elite-benchmarks-3494890/