v0.6.9.6 — AMD: concurrent per-card tier streaming (fixes startup reject burst, #9)
Follow-up to v0.6.9.5. That release streamed each card's PoM tier copy from the GGUF on disk (removing the multi-GB host-RAM cache), but streamed the cards one at a time — on many-card rigs the hashrate ramped up card by card for minutes after a (re)start, the pool's vardiff kept chasing the climbing rate, and shares in flight across each difficulty change were rejected low-diff. That produced the 15–20% reject burst at startup reported in issue #9, decaying once the rig reached full speed.
Changes
- Concurrent per-card streaming: the one-time possession-index build stays serialized, but every card now streams its VRAM copy concurrently. All cards come online nearly simultaneously (as in ≤0.6.9.3), so vardiff settles once. The OS page cache serves the follower reads, so total disk cost stays ~one pass of the GGUF, and the v0.6.9.5 RAM savings are unchanged.
- Staging window per card reduced 256 → 128 MiB (N cards now hold N windows transiently).
- Same-card double-install guard for the concurrent path.
Consensus safety
No consensus code touched (driver-side only; pom.rs folds/layout unchanged). Verified byte-exact: H3 salt test, pinned Gemma root R_T 846caa40…, real-tier GPU end-to-end (pinned nonce) all pass.
Validation (live pool, 3-GPU rig)
- All cards PoM-resident within ~1 s of each other (previously staggered)
- Hashrate parity, 0 rejected in the startup window
- Miner RSS unchanged vs v0.6.9.5 (~0.7 GB)
AMD assets
| file | sha256 |
|---|---|
keryx-miner-supr-amd-0.6.9.6.tar.gz (HiveOS)
| d7b3439570bcb76e7f21b7321548ac4a44b2518ee91c3974184b45131e6a71dc
|
keryx-miner-supr-amd-mmpos_0.6.9.6.tar.gz (mmpOS)
| 9a0a3da5313e3c40cd2a34d0527d89db2679fca41a608e05cfca4096607aa03a
|
keryx-miner-supr-amd-0.6.9.6.zip (Linux runnable)
| 659d87d934175a13a8655f3ba33b4eaea370b24dd139e507e1bc42a3804e7e0d
|
Windows AMD build follows via CI.
If you saw elevated rejects on v0.6.9.5: upgrade, or as a general workaround pin static difficulty via the pool password (e.g. d=16).
NVIDIA Linux assets (modern / legacy / pascal)
The 12 NVIDIA Linux tarballs + per-line SHA256SUMS are attached (walk PTX per line: modern sm_75, legacy sm_70, pascal sm_60; banner 0.6.9, BUILD 24). The v0.6.9.6 fix is AMD/OpenCL-only — NVIDIA binaries are functionally identical to v0.6.9.5, rebuilt here so both platforms ship on one release.
Windows assets
Attached by CI (click "Show all assets" if the list is collapsed): keryx-miner-supr-windows-nvidia-pom.zip (NVIDIA), keryx-miner-supr-windows-amd.zip (AMD — includes this release's concurrent-streaming fix), keryxcuda-windows.zip (plugin DLL).