GPU miner #12
Replies: 3 comments 8 replies
|
I built one. It works, and here are the numbers. On the proof: you're right, and that's exactly why it needs a pool. A GPU can't build the HistoryStep, so it can't mine Parano1d on its own — it would sit idle waiting for a CPU between every block. Behind a pool that stops mattering: the pool's node proves once, and every GPU attached to it searches nonces on that same template in parallel. The GPU never touches the proof. On the arithmetic: you were right to be unsure, and it turned out in the GPU's favour. Poseidon2b over GF(2^128) does need carryless multiplication that CPUs have in silicon and GPUs must emulate. Emulating it costs a lot per operation — but there are enough operations in flight to pay for it several times over. Measured per card, from the pool's own share accounting (not from what the miner claims about itself): Measured hashrate Binaries for Windows, Linux and HiveOS: https://github.com/lucasan123/noid-gpu-miner It carries a disclosed 10% dev fee, announced on every run with the address in full, and it stops rather than mine if it can't deliver it. Only binaries are published — if you'd rather run something you can read, the official parano1d-miner is the right choice and I won't argue with it. |
|
If the pool credits valid shares proportionally, 20 MH/s of Poseidon2b nonce work has the same expected value whether it comes from a CPU or GPU. |
|
Yes, the question was addressed to the pool owner. |
|
Yes — 20 MH/s is 20 MH/s, whatever produced it. The pool never learns what hardware sent a share: it gets a nonce, recomputes the digest itself, and credits it if it beats the target. Every accepted share weighs the same, and a block's reward is split as prize × your weight / total weight, minus the pool fee. An EPYC and a 5090 delivering the same accepted shares per second get paid the same. But the hardware that actually decides your earnings isn't yours — it's the pool's CPU. After every block the pool's node has to prove a new template before anyone can mine at all. On our pool that takes 5.9 s on average, so out of a 20-second block only about 13 s (67%) are mineable. Those seconds are lost by every miner attached, at once, no matter what card you own. A pool one second slower at proving costs you ~5% of your hashrate; three seconds slower costs 15%. So when you compare pools, ask for that number — not their advertised hashrate. Ours is on the dashboard as template time (proving), next to the per-block proving time for every block we've won: https://parano1d-pool.fun/ |
|
Thx friend |
|
Thanks for publishing. Please create a dedicated thread in the Mining category named [Miner] noid-gpu-miner for releases, setup, benchmarks and user support. Include the developer fee, supported hardware, source availability and download links. also |
|
What are the roadmap and the next stage ? |
|
I would like to share another independent implementation: Hashborn NOID Miner. Hashborn is a closed-source CUDA miner for the nonce-search phase of Parano1d mining. It works with a pool: HistoryStep construction remains on the pool's CPU, while the GPU searches the resulting template. Current release: v0.1.1
Measured CUDA throughput:
During live RTX 5090 mining, we have also observed more than 100 MH/s on the pool dashboard, not only in the miner's internal CUDA counter. Effective wall and pool hashrate still depend on template preparation time, pool availability and job transitions. For example, the GPU may sustain approximately 104.5 MH/s while the interval wall rate is lower whenever the pool is preparing the next HistoryStep. RTX 30-series support has passed compilation, native-instruction and mathematical verification, but we have not yet measured it on a physical RTX 30 card. Only binaries and documentation are published. The miner is not open source, and every user must decide independently whether to trust and run a closed-source binary. The release provides published SHA-256 checksums and a signed built-in Release and documentation: https://github.com/hashborn72/noid_miner/releases/tag/0.1.1 We welcome independent pool-side measurements and comparisons under the same pool and template conditions. |
|
Thanks for publishing. Please create a dedicated thread in the Mining category named [Miner] Hashborn NOID Miner for releases, setup, benchmarks and user support. Include the developer fee, supported hardware, source availability and download links. |
Uh oh!
There was an error while loading. Please reload this page.
Could somebody write a private GPU miner? For nonce search, yes.
But nonce search is only the second phase of Parano1d mining. First, the node must build a new HistoryStep proof for every parent. The current prover is CPU-only and optimized for binary-field arithmetic through AVX2/AVX-512 VPCLMUL on x86 and PMULL on ARM. There is no CUDA or OpenCL prover. A GPU nonce kernel cannot build this proof. It must wait for the CPU, and if another block arrives first, the prepared work becomes obsolete.
A GPU may still gain something during nonce search because it can run many independent attempts. How much is unknown. Poseidon2b over GF(2^128) is not SHA-256. Modern CPUs have native carryless multiplication instructions for this arithmetic, while GPUs would need to emulate it. Their parallelism may compensate for that, but only a real implementation and benchmark can answer the question.
All reactions