Suprminer opens NEB mining to Tesla V100 and brings faster methods to Turing cards
Source report: 2026-11-22 · Editorial analysis published: 2026-11-22
The November 22 releases add V100 support, test table use on each card and improve dense Kepler work on selected NVIDIA GPUs. Developer benchmarks are separate from mining income.

Older cards get a new option
Suprminer 1.9.54, published on November 22 at 12:44 UTC, lets Tesla V100 cards mine NEB after earlier versions refused them. It also brings its faster calculation path to Turing cards: the RTX 20 and GTX 16 families and CMP 30HX, 40HX and 50HX. For someone with an existing card, this is a new software option to evaluate before considering different hardware.
NEB is proof-of-work based on training small neural networks, rather than Bitcoin's SHA-256 hashing. The work changes with the model in each block. Kepler and TESS are names of these workload models, not the NVIDIA architecture named Kepler. A supported GPU is not thereby a Bitcoin ASIC, and these release notes do not add a new ASIC to the catalogue.

What was actually measured
The developer reports 131 million attempts per second on Kepler and 942 million on TESS for a 16 GB Tesla V100. Those November 22 model tests used Linux, unchanged clocks and a 300 W power allowance. That allowance is a setting, not a measurement of the whole machine at the socket. V100 support is offered in the Linux packages only.
Turing has a different evidence level: the developer has no physical Turing card and gives no speed figure. The path was compiled and exercised on an RTX 5090 through the driver's compiler, with valid shares. On the actual target card, each model is checked against the plain kernel, the simpler original calculation method. If the check fails, the miner keeps that original method. This is a compatibility safeguard, not a benchmark on an RTX 2060 or GTX 1660.
Tables have to earn their memory
The fast path selects useful attempts and stores reusable results in tables in GPU memory. A larger table is not automatically a better choice: reading it can cost more time than recalculating part of the work. Version 1.9.54 therefore times a completed large table's scan with and without the table on that particular card, then keeps the faster route.
For troubleshooting, NEB_DISABLE_NONCE_SELECTION=1 forces the plain kernel on any card. The normal mining algorithm argument remains -a neb. A useful comparison holds the card, model and settings constant and checks accepted pool work, rather than comparing one screen's headline speed with a different model seen yesterday. These are published software controls; ASIC.tools has not run the binaries or independently benchmarked the cards.

Earlier today: denser Kepler work
Version 1.9.53 arrived at 10:26 UTC. For the denser Kepler models used since November 21 it reports about 10–13% faster work on RTX 40/50, RTX PRO 6000, H100 and CMP 170HX. In its paired table the RTX 5090 rises from 368 to 405 MH/s and the RTX 5080 from 206 to 231 MH/s. RTX 3070 stays at 59 MH/s; RTX 30 cards and TESS are unchanged by that optimization.
Other paired examples are PRO 6000 from 390 to 429 MH/s, H100 from 217 to 241, RTX 4070 Ti SUPER from 139.5 to 154.5, and memory-unlocked CMP 170HX from 80.5 to 91. These are Linux tests with unchanged clocks, all shares checked and none rejected according to the developer. Their lower absolute speeds than earlier release tables reflect a different, denser model. They are not evidence that the new release slowed those same cards. Other cards in 1.9.54 retain the 1.9.53 kernels.
One package, two builds
Since 1.9.53 each platform archive includes NVIDIA and AMD OpenCL builds, with checksums in SHA256SUMS. HiveOS packages come in _u2004 and _u2204 variants matching the system image. HiveOS and mmpOS start the AMD build on a rig without NVIDIA; --amd or --nvidia explicitly selects a vendor. SMOS includes miner and miner-amd in one ZIP, while Windows provides separate start-neb.bat and start-neb-amd.bat launchers.
Docker still has separate NVIDIA and AMD image tags. The developer has not tested mining on physical Windows GPU hardware. The advertised 0% miner fee concerns the program's own deduction; pool charges, electricity and equipment costs remain separate. A faster supported route can improve useful work, but the coin price, network competition, accepted shares and actual operating cost determine whether switching makes financial sense.
The tagged guide specifies NEB 1.0.1 data: neb-datasets.bin is 16,087,037 bytes. The program checks its hash, can fetch a verified copy over HTTPS and rejects an unrecognized explicit file. A new executable with the wrong data is therefore a different problem from an unsupported card; keep the package and dataset requirements together when diagnosing startup.
Source: Suprminer ↗ · Suprminer 1.9.53 release and paired tests ↗ · Tagged NEB setup guide ↗ · NEB pool and setup ↗
Mining calculator ↗

