Skip to content

NOID-GPU 1.1.0

Choose a tag to compare

View all tags
@lucasan123 lucasan123 released this 24 Aug 22:32
· 2 commits to main since this release
v1.1.0
62153f6
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

NOID-GPU 1.1.0 release notes

Release date: 24 August 2026

Highlights

  • 18.99% higher measured wall-rate on RTX 5070 Ti in a controlled B-A-B
    comparison.
  • 70.492 MH/s at 280.14 W on RTX 4090.
  • 104.526 MH/s at 453.69 W on RTX 5090.
  • RTX 4090 median board power remained 169.86 W below its 450 W limit.
  • RTX 5090 median board power remained 46.31 W below its 500 W limit.
  • Public miner developer fee reduced from 10% to 7.5%.
  • Parano1d Pool fee changed from 10% to 7.5%.
  • Compact English startup panel and English user-visible errors.
  • Windows, Linux, and HiveOS packages; Linux/HiveOS remain compatible with
    the Ubuntu 20.04 GLIBC 2.31 baseline.

Measured performance and power

Stabilized RTX 4090 and RTX 5090 runs

GPU NOID-GPU wall rate Board power MH/s/W Configured power limit Unused headroom
RTX 4090 70.492 MH/s 280.14 W 0.25163 450 W 169.86 W (37.75%)
RTX 5090 104.526 MH/s 453.69 W 0.23039 500 W 46.31 W (9.26%)

The measured hashrate therefore did not require either card to sit at its
configured maximum board power throughout the stabilized run.

The normal 16,777,216-nonce batch produced:

GPU Median kernel rate Median wall rate Short live dry-run active rate
RTX 4090 70.433 MH/s 70.314 MH/s 70.23 MH/s
RTX 5090 104.710 MH/s 104.298 MH/s 103.78 MH/s

Power data came from the longer 268,435,456-nonce stabilization run, using
two discarded warm-ups, five measured samples, and NVIDIA telemetry sampled
every 100 ms. The normal batch used five warm-ups and fifteen samples.

These were separate dedicated cards. Results are measured evidence from one
card of each model, not a promise that every rig will reproduce the same
watts.

RTX 5070 Ti controlled speed comparison

Build Median kernel rate Median wall rate
Previous frozen kernel 34.679 MH/s 34.566 MH/s
NOID-GPU 1.1.0 41.290 MH/s 41.130 MH/s
Improvement +19.06% +18.99%

Power was not resampled during this controlled B-A-B. Accordingly, this
release claims a measured speed increase on RTX 5070 Ti, but does not
claim a controlled watt reduction versus the previous kernel.

A separate Windows field run with three RTX 5070 Ti cards produced about
120 MH/s combined while instantaneous NVIDIA-SMI readings were roughly
132-180 W per card, despite configured 300 W limits. This shows that the
observed hashrate did not require the cards to reach the power cap. It is a
field observation rather than a controlled before/after power benchmark.

CUDA optimization

  • Reduced the full Poseidon2b path from 502 to 448 GF(2^128)
    multiplications per permutation    , a 10.76% reduction in arithmetic
    work    .
  • Added a lazy most-significant-bit-first comparison directly from the flat
    GPU state, avoiding full digest conversion on the normal reject path.
  • Replaced the hot MDS calculation with algebraically equivalent factorized
    full and partial circuits.
  • Shortened the Karatsuba XOR dependency chain through clmad.lo accumulation.
  • Added a full-grid, one-nonce-per-thread kernel for pool work and the offline
    benchmark.
  • Retained the persistent kernel for the rare block-target path where early
    exit is valuable.
  • Kept the final sm_120 hot kernels at 64 registers per thread with no static
    stack, local-memory spill, or shared-memory allocation.

The performance gain comes from doing less work per hash and improving the
GPU execution path. It is not presented as a result of simply raising the
power limit.

User-visible changes

  • Replaced the verbose normal GPU startup with a compact NOID-GPU-WORKER
    panel showing payout, selected GPU indices, hardware-test result, and
    mining status.
  • Removed the obsolete allow-custom-coinbase startup advice.
  • Public CLI names, help, status, benchmark labels, warnings, CUDA errors,
    launchers, and HiveOS documentation are now in English.
  • Operating-system I/O failures no longer inherit localized Windows text;
    they use an English category and retain the numeric OS error code.
  • Detailed CPU, batch, and arithmetic diagnostics remain available through
    explicit self-test and benchmark modes.

Correctness and release gates

  • Added a deterministic offline benchmark of the real GPU PoW kernel with
    separate CUDA-event and wall-clock timing.
  • Compared 16,384,000 optimized GF(2^128) products against emulated and
    independent schoolbook paths.
  • Compared 40,960 fast permutations with the generic matrix path.
  • Verified 64 complete CPU/GPU digests, 101,029 target comparisons, both
    positive search paths, and exact attempted-nonce counts.
  • Every nonce returned by CUDA is still verified by the CPU; a mismatch stops
    the miner instead of submitting unverified work.
  • Preserved stale-job detection and bounded concurrent share submission.

The miner/worker and Poseidon2b suites pass 113 tests with zero failures
(90 miner/worker tests and 23 arithmetic tests). Exact packaged executables
also passed offline GPU correctness and benchmark gates.

Packages and compatibility

Public assets:

  • noid-gpu-windows-1.1.0.zip
  • noid-gpu-linux-1.1.0.tar.gz
  • noid-gpu-hiveos-1.1.0.tar.gz

Requirements:

  • NVIDIA Ampere or newer: sm_80, sm_86, sm_89, sm_90, or sm_120.
  • NVIDIA driver 580 or newer.
  • CUDA runtime included; no CUDA toolkit required on the mining rig.
  • Linux/HiveOS maximum required GLIBC symbol: GLIBC_2.30, compatible with
    HiveOS Focal/Ubuntu 20.04 GLIBC 2.31.

The public noid-gpu build supports its built-in pools and does not accept an
arbitrary node. The unrestricted open-pool, no-developer-fee noid-worker is a
separate private operator/test build, not a public 1.1.0 asset.

Developer fee

The public miner fee is 7.5% of scheduled mining time, implemented as a
45-second window in every 600-second cycle. Switching occurs only between
jobs, so a short sample can vary slightly at a job boundary. The full fee
address is printed at startup, and the miner never changes a block coinbase.
Pool fees remain separate.

Power-limit guidance

A configured power limit is a ceiling, not a target. Raising it alone does
not make a GPU consume more power or hash faster when the card is already
below the cap. Tune core clock and power one change at a time and compare both
MH/s and MH/s/W. Reject settings that introduce verification errors, rejected
shares, instability, or thermal throttling.

Pool fee

The Parano1d Pool fee is now 7.5%, reduced from 10%. The miner developer
fee is separate and is also 7.5% in this release.

Assets 6
Loading