How much GPU energy does fp16 save? Measured on a Tesla T4

How much GPU energy does fp16 actually save over fp32? Spec sheets give TDP and peak FLOPS, not joules per job, so I measured it on a Tesla T4 using the GPU's own NVML energy counter.

The job

import sys, torch
dtype = {"fp16": torch.float16, "fp32": torch.float32}[sys.argv[1]]
a = torch.randn(8192, 8192, device="cuda", dtype=dtype)
for _ in range(100):
    b = a @ a
torch.cuda.synchronize()

fp16 and fp32 runs alternate (so temperature and clock drift hit both sides), 5 measured runs each, medians reported, Mann-Whitney test between the two.

Results: the whole job

Setup	fp16	fp32	GPU energy saved
No cooling between runs	475 J, 8.6 s	2.53 kJ, 39.7 s	81.2%
Every run starts at ≤65 °C	426 J, 7.8 s	2.21 kJ, 33.8 s	80.8%
≤65 °C + 3 warm-up runs	413 J, 8.4 s	2.50 kJ, 40.1 s	83.5%
Same, a second session	392 J, 7.9 s	2.29 kJ, 36.1 s	82.9%

p = 0.008 every time: every fp16 run beat every fp32 run.

Results: just the matmuls

Measuring inside the process, without Python start-up and CUDA set-up, 20 matmuls took about 73 J and 1.1 s in fp16 against about 486 J and 7.4 s in fp32: roughly 85% less GPU energy and 6.6× faster.

import torch, joules
with joules.Session(gpus=[0]) as s:
    with s.phase("fp16") as p:
        ...  # your code
print(p.energy_j, p.duration_s)

Why the gap is so large

The T4 is Turing: fp16 matmuls run on the tensor cores, but there is no TF32, so fp32 runs on the ordinary CUDA cores. Energy is power × time, so most of the saving comes from finishing sooner. fp32 also sat throttled 93% of the time at the T4's 70 W cap, against about 60% for fp16.

On an A100, H100 or L4, fp32 matmuls can use TF32, so expect a much smaller gap there. Measure it rather than assume it.

Run it yourself

pip install -U landauer-gap
joules compare --names fp16,fp32 --baseline fp32 --runs 5 --warmup 3 --cool-to 65 --gpus 0 \
  "python3 bench.py fp16" "python3 bench.py fp32"