A native CUDA miner for NEB (nebhash v2): every mining attempt is one integer training step of the chain's neural network (on the main network both block slots run the KEP1 network: Kepler and TESS signals). The NEB main network starts on
18 November 2026, 14:00 UTC.
What's new in 1.0.131.0.12 was withdrawn (after an auto-uv fault a healthy GPU could stay DEAD). Please use 1.0.13: it has everything from 1.0.12 plus the fixes below.
https://raw.githubusercontent.com/pepsykolya/FreeForGemNeb/main/docs/img/ffgn-1.0.13.pngFixes- GPU fault recovery gets the safe state before the new engine starts. After a fault the miner waits out the driver's reset in a separate probe process, takes a fresh handle, applies the run's core lock, memory policy and power limit, sets a safe offset (2 steps below the crashed one, never above the last confirmed one, never below the base), reads it back, and only then starts the engine. A healthy card is no longer marked DEAD; DEAD / LOST only means the device cannot be opened or reached. A card that keeps faulting is paused for 10 minutes, then gets a new engine; no reboot is asked.
- Fewer stale shares: hits of a job replaced by a new clean job are dropped at once, the CPU check runs off the pool path, after "job not found" the rest of that job is not sent. 5x CMP 50HX rig: 21.8 % stale on 1.0.12 -> 8.1 % on 1.0.13.
- Late pool warning: one WARNING per 10 minutes with the median delay when the pool's answers come seconds late (a lossy or slow network link makes shares stale).
- --debug-timing: per share scan -> CPU check -> socket (ms), per job socket -> dispatch, to tell miner queues from the network.
From 1.0.12: partial L=3 table for TESS (automatic, --tbig auto), built while the pair table mines and only when it pays off. TESS rate per card (engine benchmark, same TESS models, same power):
- RTX 4070 Ti 12 GB: 415.1 -> 488.9 MA/s, +17.5..17.8 % rate, +17.2..18.0 % per watt
- RTX 3080 Ti 12 GB: 404.0 -> 472.3 MA/s, +16.9..17.0 %
- RTX 3070 8 GB: 238.9 -> 261.7 MA/s, +9.4..9.6 % rate, +9.4..9.8 % per watt
- RTX 4070 Laptop 8 GB: +6.1..8.3 % rate, +8.3..9.3 % per watt
The picture above shows Kepler and TESS absolute MA/s, W and MA/W per card, including Turing (CMP 50HX by core lock, RTX 2060 SUPER). All values are per ONE card; some were measured on multi-card rigs.
- New options: --power-limit / --pl per card, --tbig off|auto|on, --reboot-on-lost-gpu MIN
- auto-uv: resume margin after crashes, cautious mode after a card falls off the bus, Turing cap +240
- Bit-exact. Validation per card: 4x RTX 3070 and 5x CMP 50HX rigs with real GPU faults (0 DEAD), RTX 3070, RTX 3080 Ti (Windows). Dev fee 1 %.
HiveOS dashboard: ffgn reports its 5-minute AVERAGE over both slots (Kepler + TESS), so it looks lower than a miner that shows the current block's rate. Compare per slot in the miner's log, or by accepted shares on the pool.
HiveOS: custom miner ffgn-hive, algo nebhash, Installation URL:
https://github.com/pepsykolya/FreeForGemNeb/releases/download/v1.0.13/ffgn-hive-1.0.13.tar.gz
Download 1.0.13Features- Windows 10/11 x64, Linux x86_64 and HiveOS (custom miner ffgn-hive, install from the release URL)
- NVIDIA GTX 10 … RTX 50 (compute capability 6.1+), driver R525+ (RTX 50: R570+), no CUDA toolkit needed, less than 1 GB of video memory; the NEB mainnet dataset is inside every package
- Pools Aria and Suprnova with failover and return to the primary pool; solo through your own NEB node; ffgn-gate lets several rigs mine solo through one node
- Every share and block found on the GPU is computed again on the CPU with the NEB reference code before it is sent
- --auto-uv: safe undervolt per GPU, checked bit for bit against the reference; --auto-lock: picks the core clock lock of each GPU by measurement (best MA/W within 10 % of the top speed); clock locks, --temp-limit for laptops
- If the miner cannot mine, it says so loudly: NOT MINING: <reason> in the log, in the stats line (HiveOS) and in the API
- Watchdogs for pools without accepted shares, local API with Prometheus metrics, HiveOS stats, Windows autostart, GPU settings restored after any stop, new-version notice
Speed (GPU engine 0.2.8, real dataset, stock memory clock, no undervolt)MA/s = million attempts per second; KEP1 and RET2 blocks alternate, so both are shown.
| GPU | Core clock | KEP1, MA/s | RET2, MA/s | Power, W | KEP1, MA/W |
| RTX 3070 | stock | 16.95 | 9.92 | 200 | 0.085 |
| RTX 3070 | lock 1350 MHz | 15.06 | 9.33 | 141 / 139 | 0.107 |
| RTX 4070 Ti | stock | 37.59 | 17.76 | 235 | 0.160 |
| RTX 4070 Ti | lock 2250 MHz | 34.44 | 17.58 | 190 / 212 | 0.181 |
Full tables (more clock locks, watts per model):
README. Note: a memory clock locked to 810 MHz (Pearl OC profiles) makes NEB about 10x slower; ffgn replaces such a lock while it mines.
Quick startPools: Aria
neb.ariabrain.com:3370 (solo
:3371) · Suprnova
neb.suprnova.cc:7804 (solo
:7805)
Windows: unpack the zip to a folder without spaces, set
WALLET in
start.bat, run it.
Linux: ./ffgn --pool neb.ariabrain.com:3370 --wallet neb1YOURADDRESS --worker rig1HiveOS: custom miner
ffgn-hive, algorithm
nebhash, wallet template
%WAL%.%WORKER_NAME%, pool
neb.ariabrain.com:3370, Installation URL:
https://github.com/pepsykolya/FreeForGemNeb/releases/download/v1.0.13/ffgn-hive-1.0.13.tar.gzSolo: your own
neb-node (official NEB release) +
ffgn --node 127.0.0.1:3441 — see the manual.
Checksums:
SHA256SUMS.
Dev fee1 % of the mining time: one 60-second slot in every 100 minutes, shown in the log.
Video tutorialsEnglish, part 1/2 (the miner, speed, the NEB coin):
YouTube ·
TelegramEnglish, part 2/2 (wallet, Windows, HiveOS, Linux, pools and solo, settings):
YouTube ·
TelegramRussian, чacть 1/2:
YouTube ·
TelegramRussian, чacть 2/2:
YouTube ·
TelegramAll tutorials:
YouTube @FreeForGem_MinersThe videos show 1.0.1; installation and settings are the same in 1.0.13.
LinksTelegram:
t.me/freeforgem (news) ·
t.me/freeforgem_chat (chat)
YouTube:
@FreeForGem_MinersNEB:
github.com/SilasEigen/neb ·
neuralenginebyte.org ·
NEB thread