Measured on PC 1 (docs/bench-log.md, "the 9070 XT on the eGPU"): the hash runs at 18.0 MH/s against a dependent
random-read ceiling of 2.42 to 2.68 G loads/s at 1 GiB (--memprobe), 128 loads per hash, so 92% of what the card does
for this access pattern; group-warps 1/2/4/8 and the old platform all give 18.0; wave32, no spills; the stream probe
reads 635 GB/s of the rated 640, random 64 B lines run at the same count as random 4 B loads. Not the eGPU link: the
link only carried the 16 MiB per-job read-back (7.3 ms per 2^21-nonce job), which the select pass removes
(124.2 to 117.3 ms per job, 16.88 to 17.87 MH/s inside jobs, kernel 116.0 ms on both).
host.c: the same card on an older platform of the same vendor is marked and hidden from --list (PC 1 ran two workers
on one 9070 XT at 8.9 + 9.4 MH/s); the default pick skips it; printKernelInfo on every path (preferred multiple,
private memory, sub-group size); --readback select|full with the transfer and time budget in the stats line;
--memprobe (chase, indep x8, 64 B lines, stream, ALU) with a fresh seed per repetition; test_host.c covers the fold.
detect.rs: parse_opencl_list with three tests (the PC 1 listing, a CPU device, a repeated index).
engine.rs: EXPORT_LOCK serialises igneum-miner export-pack across the per-card threads (PC 1's packs\devnet held
one epoch's program.h with another's seeds.txt after an epoch change and every OpenCL worker start refused it).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
23810dfe58