@@ -1332,7 +1332,28 @@ takes 8.02 µs against NumPy's 8.14: **parity**. float64 takes 16.3 against
1332133215.8 µs (1.03×).
13331333
13341334** Open defect: products are still up to 1.2× slower than OpenBLAS on one
1335- core** at 256² and 1024² on N1, and float32 1024² on Zen 3 is at 1.04×. The exception is float64 at 1024² on Zen 3, where we are 3% faster.
1335+ core** at 256² and 1024² on N1, and float32 1024² on Zen 3 is at 1.04×.
1336+
1337+ ** Open defect: on a busy N1 host we slow down and OpenBLAS does not**
1338+ (measured 2026-10-31). The host was at load 21 on 64 cores. Back to back on
1339+ the same pinned core, NumPy's times were the usual ones, but ours were about
1340+ a third slower than on a quiet host. So the ratio to NumPy grew to:
1341+
1342+ | N1, load 21, ours / NumPy | 64² | 256² | 1024² |
1343+ | ---| ---:| ---:| ---:|
1344+ | float64 | 1.27 | 1.40 | 1.54 |
1345+ | float32 | 1.39 | 1.25 | 1.27 |
1346+
1347+ Ruled out on that host, by A/B:
1348+
1349+ - a regression: v0.26.1, v0.27.0 and v0.29.1 are equally slow there;
1350+ - reading A in place: v0.26.1, which packs A, is no faster;
1351+ - the block sizes: OpenBLAS's ` param.h ` values (KC 320, MC 240) change
1352+ nothing, and NC = 128 or 256 (a B block well inside the 1 MiB L2) moves
1353+ float64 by 4–9% but float32 1024² by +4%.
1354+
1355+ The cause is not found. The quiet-host ratios above stand for a quiet host,
1356+ and this one for a contended one. The exception is float64 at 1024² on Zen 3, where we are 3% faster.
13361357The float32 kernels for riscv64 and loong64 are still to write.
13371358
13381359** Open: the allocating operations.** ` Add ` is still 1.5–3.2× slower than
0 commit comments