@madiator The Mac row in your own latency table is worth pairing with the 100ms:

@madiator The Mac row in your own latency table is worth pairing with the 100ms: 444 ms median and 981 at p95 on an M5 Pro 64GB, against 246.7 for the Jev API. And the H100 106 ms

@madiator The Mac row in your own latency table is worth pairing with the 100ms: 444 ms median and 981 at p95 on an M5 Pro 64GB, against 246.7 for the Jev API. And the H100 106 ms row is 120 examples where every other row is 324, is that a separate run?