5 ms·
> Graviton is apparently pretty closely based on an Arm Neoverse N1 CPU and the 64 vs 32 core point is comparing a hyperthreaded part vs one that isn't. How do
by kllrnohj 6y ago
> Graviton is apparently pretty closely based on an Arm Neoverse N1 CPU and the 64 vs 32 core point is comparing a hyperthreaded part vs one that isn't.
How does hyperthreading change the story here? The 32-core CPU is the one that had hyperthreading while the 64-core one didn't. Hypthreading is widely regarded as being around +20% performance for multithreading-friendly workloads. Either way, the per-core performance of the 32-core x86 CPU is nearly 2x that of the 64-core ARM one. That's not a good look for being desktop-viable.
Especially when the 32-core x86 cpu also comes in a 64-core variant. And then a 2P 64-core variant even. You can have double the CPU cores that are each 2x faster than the Graviton 2 CPU cores.
Which gets back to only Apple has managed to get ARM to have good per-core performance so far.
> Plus Graviton seems to be materially more cost effective.
The c5a.16xlarge is the same price as the m6g.16xlarge. No cost effective difference in that head-to-head.
- klelatti 6y agoI actually agree that x86 will dominate the desktop for quite a while yet. I also agree that EPYC has materially better performance than the Graviton - Rome is very impressive. Just can't agree though that only Apple has the ability to make desktop / server Arm parts that don't have 'terrible' per core performance. The real issue is who has the economic incentive to build competitive desktop parts - I don't see anyone who would see it as worthwhile.
- imtringued 6y agoThat's the fundamental problem with the Apple monopoly. I would be perfectly happy if I could use an non-Apple laptop with an outdated Apple SoC. However, since only Apple gets access to their SoCs everyone is worse off.
- BurritoAlPastor 6y agoc5a and m6g instances are the same price, but m6gs have twice as much memory. c6g instances are a better point of comparison for c5a – same vCPU count still, same memory, marginally better network at 8xlarge and up, and about 88% of the price.
- _msw_ 6y agoDisclosure: I work at AWS building cloud infrastructure > The c5a.16xlarge is the same price as the m6g.16xlarge. No cost effective difference in that head-to-head. c6g.16xlarge is more than 10% cheaper than m6g.16xlarge (and c5a.16xlarge). It also provides more EBS and network bandwidth, and provides 64 cores versus 32 cores with SMT. https://ec2instances.info/?region=us-west-2&compare_on=true&selected=c6g.16xlarge,m6g.16xlarge,c5a.16xlarge https://ec2instances.info/?region=us-west-2&compare_on=true&...