Hashwarp v2.0.0

Hashwarp 2.0 is a rewrite of nearly everything around the XHash kernels: setup, output, monitoring, safety and packaging, with fixes for crashes and wrong results, overclocking, and a dashboard for one rig or many. Every option of 1.x still works.

Download

• Linux: hashwarp-v2.0.0-linux.tar.gz, for glibc 2.28 or newer (Ubuntu 20.04, Debian 10, RHEL 8 and later). Everything else is built in: no CUDA toolkit or OpenSSL needed.
• Windows: hashwarp-v2.0.0-windows.zip, for Windows 10 or newer. The Visual C++ runtime is in the archive.
• HiveOS: hashwarp-2.0.0.tar.gz (HiveOS needs this name). In the flight sheet choose the Custom miner and paste this file's link as the installation URL.
• mmpOS: hashwarp-2.0.0-mmpos.tgz.

One build per system mines on NVIDIA GPUs with CUDA (driver 525 or newer, every GPU from Pascal to Blackwell) and on AMD and other GPUs with OpenCL. There are no separate CUDA and OpenCL builds anymore.

Highlights

• Easy to start. Run hashwarp without options: a short wizard asks whether you mine on a pool or solo and saves hashwarp.json. Or use --pool, --wallet and --worker, and keep options in a JSON configuration file (--save-config writes one).
• Clear output. Each line has its time, level and source (gpu0, pool, share ...); a line per share with its difficulty and latency, or the pool's reason when rejected; the hashrate with its 1 and 15 minute averages and the effective (pool side) hashrate; temperatures, fan, power and efficiency; a table per GPU every minute. Keys: h for help, s statistics, p/r pause/resume, 0-9 one GPU. --log-format json writes one JSON object per line.
• Dashboard and API. --api-bind 127.0.0.1:21550 serves a web dashboard, a JSON API and Prometheus metrics. --rig adds your other rigs to the dashboard, and --monitor serves that view without mining.
• Overclocking. Memory and core clocks, power limits and fans per GPU (NVIDIA, and AMD on Linux), restored on exit. --oc-tune finds each GPU's best memory clock and the most efficient core clock; mine with --oc-profile tuned or efficient.
• Safety. Temperature limits per GPU (core, memory or hotspot), --shutdown-temp, and a watchdog that restarts a failed GPU and exits with its own code when a GPU hangs. Every solution is checked on the CPU before it is submitted. --selftest checks the GPUs offline.
• Good for the network. Hashwarp warns when your pool finds most of Parallax's blocks, which lets it reverse payments, and names the smaller pools. The wizard lists such a pool last.
• Solo mining. --solo mines on your own node. With --solo ws://127.0.0.1:8546 a node that supports it pushes each new block's work at once instead of being polled every second.
• Rig operating systems. Packages for HiveOS and mmpOS.

Performance

115.5 MH/s on a stock RTX 4090, about 94% of its memory bandwidth and as fast as SRBMiner-Multi on the same card, without a dev fee: Hashwarp has none. A second CUDA kernel is timed against the standard one on each GPU, and the faster one is used.

Fixes

• The epoch now comes from the seed hash, and pool difficulty is converted to an exact target: shares on some pools were rejected or the wrong DAG was built.
• Multi-GPU rigs on getwork and EthProxy no longer hash the same nonces twice.
• A malformed message from a pool no longer crashes Hashwarp.
• A GPU no longer stays paused for good after a failed DAG build, and out of bounds DAG writes are fixed on CUDA and OpenCL.
• GPU memory is reported per GPU (it was the first GPU's on mixed rigs).
• TLS sends the server name (SNI) and recognizes certificate errors; reconnects back off instead of hammering a pool, and a single pool is never dropped.
• Stratum: autodetection of the protocol, and pools such as rplant that answer in their own way.

Upgrading from 1.x

• Command lines and configuration from 1.x keep working: the ethminer style options (-P, -U, --cu-devices ...) are still accepted.
• The output is different. If a script parses Hashwarp's log, use --log-format json or the API's /api/v1/summary instead. Hashrates are written MH/s and difficulties GH.
• The API listens only on 127.0.0.1 unless --api-bind says otherwise, and is read only unless --api-allow-control is given (with --api-token on other addresses). --api-port still works, read only.
• The Windows archive no longer has fix-opencl-icd.vbs: NVIDIA GPUs mine with CUDA, which needs no OpenCL registration. See Troubleshooting Hashwarp on Windows if an AMD GPU isn't found.

Documentation

• Mining with Hashwarp: pools, solo mining, GPUs, tuning, overclocking, logs, HiveOS and mmpOS.
• API: dashboard, rigs, JSON API, Prometheus.
• Parallax mining guides