Performance: networked

xcd, the par daemon, serves a machine’s GPUs to programs on others (see par). Here it runs on a Windows machine with an RTX 5080, on the local network, and two machines use it: an Apple MacBook Pro with an M4 Max, whose own GPU is fast, and a Linux machine with an AMD Ryzen 9 9955HX, whose only GPU is integrated. Each figure is measured with the released 0.86 xcc.

Each program runs four ways: on the machine alone (XC_PAR_REMOTE=none), on the remote GPU alone (XC_PAR_WORKERS=remote), on every GPU, the machine’s own and the remote one (XC_PAR=gpus), and as auto chooses with the remote GPU among its choices. Each run starts from empty settings, so auto measures afresh, and a remote-only figure counts only if every block ran on the remote GPU. Every mode’s checksum must agree.

Apple M4 Max (Metal), using the remote RTX 5080 (ms; best of eight runs)

benchmark	this machine	the remote GPU	every GPU	`auto`
`mandel_deep`	44.0	238	112	44.1
`mandelbrot`	2.1	2.0	1.8	1.8
`md5`	5.8	203	105	6.7
`saxpy`	502 µs	273	120	901 µs

AMD Ryzen 9 9955HX with its integrated Radeon (Vulkan), using the remote RTX 5080 (ms; best of eight runs)

benchmark	this machine	the remote GPU	every GPU	`auto`
`mandel_deep`	942	309	251	302
`mandelbrot`	2.6	654 µs	2.9	792 µs
`md5`	16.0	71.4	52.5	48.7
`saxpy`	2.9	372	136	3.7

What the tables show:

A remote GPU pays where a block computes much from little data. The hashcat MD5 workload sends almost nothing and hashes 64 million candidates: from the Linux machine it runs on the remote RTX 5080 many times faster than on its own devices, and auto picks the remote card.
It does not pay where the machine’s own GPU is already fast, as on the M4 Max, or where a block moves more than it computes (saxpy): auto keeps those local.
Every GPU together is not always faster. Beside a remote card many times faster, a slow integrated GPU’s share holds the fast one back; auto measures the remote card alone too and keeps whichever is fastest.

Reproducing

python3 benchmark/par/remote.py
python3 benchmark/page.py

remote.py starts xcd on the configured GPU machine and measures from this machine and the configured Linux host.