Context:: homelab
Cluster:: k3s
User:: marcos
Forge Rev:: v1.0 (C17)
K8s Rev:: v1.36.5+k3s1
CPU:: 18%
MEM:: 32%

 __  __
|  \/  | __ _ _ __ ___ ___  ___
| |\/| |/ _` | '__/ __/ _ \/ __|
| |  | | (_| | | | (_| (_) \__ \
|_|  |_|\__,_|_|  \___\___/|___/

bench(keras)[18]

Keras nightly benchmarks from the homelab, in samples per second, higher is better. Every backend trains and runs a small CNN and MLP each night on whatever Keras was at that day. The GPU node is clock capped and shares its card with other jobs, so a single dip is usually the machine. A drop that stays is worth a look.

BACKEND↑	DEVICE	MODEL	MODE	LATEST	Δ PREV	TREND	RUNS	STATUS
jax	rtx-5070-ti	cnn	predict	288.7k	-9.4%	▇▇██▇▇▇▄▄▇▃▁▇▇▅	15	Steady
jax	rtx-5070-ti	cnn	train	82.3k	-1.1%	█▂██████▇▇▇▁▇▇▇	15	Steady
jax	rtx-5070-ti	mlp	predict	748.5k	-1.5%	█▆█▇▇▇▇▆▆▇▆▁█▇▇	15	Steady
jax	rtx-5070-ti	mlp	train	492.6k	+179.1%	▆▇█▇▆▆▆▅▅▆▄▃█▁▇	15	Faster
mlx	apple-m4	cnn	predict	37.6k	-0.0%	▁▁▆▇▇▇███	9	Steady
mlx	apple-m4	cnn	train	7.5k	+0.1%	▁▁▄██████	9	Steady
mlx	apple-m4	mlp	predict	442.0k	-29.6%	▆▆▆▆▁▆▇█▂	9	Regressed
mlx	apple-m4	mlp	train	180.8k	-12.4%	▂▃▆▅▁▆█▅▃	9	Steady
mlx-main	apple-m4-cpu	cnn	predict	5.6k	-0.3%	▁▁▄███▆▆▆	9	Steady
mlx-main	apple-m4-cpu	cnn	train	1.5k	+0.6%	▁▁▄██▇▆▆▆	9	Steady
mlx-main	apple-m4-cpu	mlp	predict	734.5k	-0.2%	▁▂▆▆▇████	9	Steady
mlx-main	apple-m4-cpu	mlp	train	200.3k	-0.9%	▁▁▅▁▁▁███	9	Faster
openvino	ryzen-5800x3d	cnn	predict	7.0k		█	1	Steady
openvino	ryzen-5800x3d	mlp	predict	268.2k		█	1	Steady
torch	rtx-5070-ti	cnn	predict	83.5k	-2.2%	▇▇███▇▇▄▇▇▆▁██▇	15	Steady
torch	rtx-5070-ti	cnn	train	29.6k	-0.1%	▇██████▇██▇▁███	15	Steady
torch	rtx-5070-ti	mlp	predict	286.0k	+5.5%	▇▇███▇▇▇▇▇▆▁█▇█	15	Steady
torch	rtx-5070-ti	mlp	train	91.6k	-2.2%	▇▇███▇▇▇▇▇▆▁███	15	Steady