# FP8 (E4M3) tensor-core issue rate on the unlocked modified 48 GB RTX 4090
# harness: 1056 blocks x 8 warps x 8 independent accumulator chains,
#          mma.sync.aligned.m16n8k32.row.col.f32.e4m3.e4m3.f32, 8192 FLOP/instruction
# instruction validated first: all-ones E4M3 inputs give exactly k=32 in every output
# session: unlocked boot, lightdm inactive, no compute process on /dev/nvidia*
# gpu: NVIDIA GeForce RTX 4090, 610.57.04, 10.35 W, 35, 210 MHz, 3105 MHz

# instruction check: d = 32.0 32.0 32.0 32.0 (expect 32.0 each)
  rep 0: 18.374 ms  602.7 TF/s
  rep 1: 18.362 ms  603.0 TF/s
  rep 2: 18.360 ms  603.1 TF/s
  rep 3: 18.360 ms  603.1 TF/s
  rep 4: 18.359 ms  603.1 TF/s
e4m3/e4m3->f32  best 603.1 TF/s   (reference: f16/f32 328.9, f16/f16 330.8, s8/s32 657.7)

# same-session reference (mma_bench 20000 5):
# # NVIDIA GeForce RTX 4090, 128 SMs, sm_89, iters=20000 reps=5
# mode       best TF/s   all runs
# f16/f16        330.8 309.9/309.9/315.0/330.8/330.7
# f16/f32        330.6 330.6/330.6/330.6/330.6/330.6
# bf16/f32       329.5 329.5/327.8/328.8/327.1/328.7
# s8/s32         657.6 657.6/657.6/657.6/657.6/657.6
# dp4a           163.5 163.5/163.5/163.5/163.5/163.5
# ffma            78.7 78.7/78.7/78.7/78.7/78.7
