Post #1802917

2026-04-15 13:16 UTC

Obviously when I teach #GPGPU I point out that the main point is performance, and “how to measure” is a topic we address. Obviously, the primary metric is kernel runtime. I then introduce the effective bandwidth metric (bytes read + bytes written, divided by time taken by the kernel) which is a good way to compare some similar kernels AND to discuss hardware limits and how close we are to them (I don&#39;t always discuss the roofline model though, maybe I should). When we start to look into more advanced things, we hit the “snag” that sometimes a kernel using a more efficient technique may be _less_ effective at using a particular resource, e.g. by having having a _lower_ effective bandwidth —I do this on purpose to show how kernel runtime remains the ultimate “tell” on how good a kernel is compared to another (regardless of the additional information the effective bandwidth may tell us). At this point I introduce a metric for which I don&#39;t actually know if there is a name: number of elements processed per seconds, which is just the number of elements, divided by the kernel runtime. I call this effective throughput, but sometimes I get the nagging feeling that this may not be the correct term?

Replies (4)

@hattom@mastodon.social 2026-04-15 13:25

@giuseppebilotta@fediscience.org This reminds me of MLUPS, "Million Lattice Updates per Second". Throughput seems like an OK term to me. What's clear is that they understand what you mean by it.

Open ##2276514
@athas@freeradical.zone 2026-04-16 05:46

@giuseppebilotta@fediscience.org I think that measure is normally just called "throughput". I teach it as the counterpart to "latency" in my systems course. I am also happy I am not the only one uncomfortable with using memory bandwidth or FLOP/s as the main measure of performance.

Open ##2276516
@ProjectPhysX@mast.hpc.social 2026-04-16 05:57

@giuseppebilotta@fediscience.org that effective throughput depends on the algorithm. Usually something like particles/s or grid cells/s. For LBM, the latter is called MLUPs/s, Mega Lattice UPdates/second.

Open ##2276517
@JSMuellerRoemer@c.im 2026-04-16 07:31

@giuseppebilotta@fediscience.org I’d just call it throughput

Open ##2276518