5 ms·
Something related to being able to go full speed without thermal throttling / instruction side effects? I.e. You know what the consistent performance ceiling is
by viraptor 6y ago
Something related to being able to go full speed without thermal throttling / instruction side effects? I.e. You know what the consistent performance ceiling is, and it won't drop just because you started using avx512 instructions on a nearby core?
- alkonaut 6y agoNot having some instructions be a noisy neighbor is a good idea. But it seems wasteful to cap 128 cores at the power output possible if all cores have that output (if that's what it is). Assuming the default mode of operation is having just a few cores operating at 100% with the majority running at 0%, it's wasteful to not let those active cores run as fast as possible.
- jholman 6y agoWhy would the default mode of operation for a 128-core datacenter chip be to have 120 idle cores? Obviously the default mode of operation is to run all cores, or very nearly all cores, at full tilt all of the time. This chip is designed to be garbage at few-threaded workloads.
- em500 6y agoIndeed. For cloud vendors, a large part of the business model is to resell/oversell idle CPU time as much as they can ("vCPU"), taking advantage of the fact that small/individual clients are usually extremely over-provisioned CPU-wise.
- alkonaut 6y agoThat depends on the workload? For HPC it’s easy, for other loads like DBs or VMs it’s harder, especially with fewer nodes (e.g a single node running 32 VMs with dedicated cores will probably look a lot like a desktop workload with lots of idle cores). I suppose it’s going to be rare to see these in deployments with fewer than several dozen sockets, and workloads must be balanced (so e.g not using dedicated VM cores). That’s I suppose a simplification they can make by focusing only on “datacenter” while intel needs the same design to work from 1 socket to hundreds.
- viraptor 6y agoDepends what your use case is. HPC for example wants predictable performance with every pipeline filled as much as possible - there won't be idle cores. The average performance is more important than bursts there. That's one of the reasons hyperthreading is universally disabled in HPC clusters.