4 ms·
I believe the difference is in the cache hierarchy (more l3 less l2) and generally high latency to dram even higher latency to hbm. This makes the prefetcher be
by MobiusHorizons 2y ago
I believe the difference is in the cache hierarchy (more l3 less l2) and generally high latency to dram even higher latency to hbm. This makes the prefetcher behave differently between the two implementations, because the l2 cache isn’t able to absorb the latency
- dagmx 2y agoThat was my initial read but they have this line which made me wonder if it was somehow more than that > I suspect Grace has a very aggressive prefetcher willing to queue up a ton of outstanding requests from a single core.
- MobiusHorizons 2y agoOh good point, maybe that is configurable as well.