By: Wilco (wilco.dijkstra.delete@this.ntlworld.com),
Room: Moderated Discussions
Travis Downs (travis.downs.delete@this.gmail.com) on December 3, 2019 1:54 pm wrote:
> There are some interesting performance claims in the AWS announcement:
>
>
>
> Note that these are "per vCPU", which means these are core to hyperthread comparisons,
> since one vCPU is one core on Gravitron, while one vCPU is one thread on x86
> (which is why x86 instances sizes are always a multiple of 2).
>
> So 40% better means m5 is 43% better (1/(1.4/2)) "per core" for m5 (Skylake-SP)s. I'm not saying the
> way they've positioned it is wrong, because you pay per vCPU, so it makes sense to compare it that way
> - but I point this out here so no one thinks Neoverse uarch somehow passed SKX in performance.
Your math is off, you have to factor in the speedup due to hyperthreading. Assuming it gives a 30% boost on average, it's 1.3/2 * 1.44 = 94%. So per-core performance is on par with a 3.1GHz Skylake-SP (AWS m5 instances use Platinum 8175 which has 3.1GHz all-core turbo). Enabling hyperthreading lowers per-thread performance, hence the significant vCPU difference.
If you're interested in throughput then the Graviton2 provides roughly double the throughput of the 24-core Skylake-SP with hyperthreading enabled. Or 2.5 times with hyperthreading disabled.
Take your pick, but either way, this counts as beating Skylake by a huge margin. Then there is power and cost (24-core Xeon Platinum is ~$5000...).
Wilco
> There are some interesting performance claims in the AWS announcement:
>
>
>
> Note that these are "per vCPU", which means these are core to hyperthread comparisons,
> since one vCPU is one core on Gravitron, while one vCPU is one thread on x86
> (which is why x86 instances sizes are always a multiple of 2).
>
> So 40% better means m5 is 43% better (1/(1.4/2)) "per core" for m5 (Skylake-SP)s. I'm not saying the
> way they've positioned it is wrong, because you pay per vCPU, so it makes sense to compare it that way
> - but I point this out here so no one thinks Neoverse uarch somehow passed SKX in performance.
Your math is off, you have to factor in the speedup due to hyperthreading. Assuming it gives a 30% boost on average, it's 1.3/2 * 1.44 = 94%. So per-core performance is on par with a 3.1GHz Skylake-SP (AWS m5 instances use Platinum 8175 which has 3.1GHz all-core turbo). Enabling hyperthreading lowers per-thread performance, hence the significant vCPU difference.
If you're interested in throughput then the Graviton2 provides roughly double the throughput of the 24-core Skylake-SP with hyperthreading enabled. Or 2.5 times with hyperthreading disabled.
Take your pick, but either way, this counts as beating Skylake by a huge margin. Then there is power and cost (24-core Xeon Platinum is ~$5000...).
Wilco