P40 is about 70% of P100 performance in almost everything.
P100 does not have power states - as its a hack - relies on nvlink to regulate p states tho it doesn't have it to regulate power states on pcie.
P40 has more vram, and normal pstates you would expect.
P100 = 2070 sometimes in games (i'd say around + ~10% difference sometimes)
P100 = Runs typically hotter than P40 at full load.
P100 = Is slower by around 15% in fps vs radeon vii
P100 = doesn't have h265
P100 = is great for sr-iov 2x vgpu, it feels like perfect split in half.
P100 = suffers from bigger memory latency vs P40 (all hbm cards suffer from that, but have much better bandwidth)
P100 = when tested with vgpu, it has problems with certain games. Though it runs 2x cyberpunk fine on medium with 50-60fps @ 1080p.
P40 = is around 1080 in most cases.
Inside primegrid
openclatiGFN15 (short task)
6800xt took 29sec
rad7 took 50sec
2080 took 62sec
p40 took 59sec
p100 took 48sec
ResNet-50 v1.5
A100 2887 images/sec
V100 1467 images/sec
P100 1100 images/sec