272 bytes
Work claim
TNet v1: nonce, row index, piece index and a 256-byte piece. This is not the size of a complete block.
Read the sourceThe evidence behind TNet
Open research.
Exact results.
Specifications, measured results and the questions that remain open. Every number comes with its scope.
11.2 ms
CPU verification
AMD Ryzen 7 8745HS, eight threads, portable Rust build with runtime AVX2 dispatch; winning-row recomputation with epoch weights already prepared.
Read the source86.7 %
Tensor share on RTX 3090
One RTX 3090 board, CUDA 12.8, cuBLAS, frozen TNet v1 parameters, median of seven attempts. This is the share of attempt time spent on int8 GEMM.
Read the source159.5 ns/ticket
RTX 3090 ticket cost
Same RTX 3090 experiment; full batch of 65,536 rows. Hardware, power limit and batch size affect performance.
Read the source88.2 %
Tensor share on CMP 50HX
CMP 50HX, Turing, CUDA 13.3, cuBLAS, frozen TNet v1 parameters. Three measured attempts after warm-up.
Read the source512 MiB
Weights per epoch
Eight 8192 by 8192 int8 weight matrices. A node can hold both current and next epoch weights; this is not its total memory requirement.
Read the sourceWhy exact requantization matters
Exact integer rounding between matrix layers prevents collapsing the whole computation into one linear product.
Why verify one row
Winning pieces are verified by recomputing their row, avoiding a separate succinct proof system.
Parity across implementations
Recorded GPU tickets match Rust, and Python checks the small-instance reference vectors. Turing and Ampere experiments agree byte for byte on checked tickets.
What remains open
External review, a measured LUT kernel, the advantage of dedicated int8 GEMM hardware, additional accelerator architectures and light-client design remain open.
Research boundaries
What is still open.
- The mining computation is not an external customer's AI inference or training task.
- ASIC resistance is not established.
- Precomputation comparisons include analytical bounds, not a measured optimized LUT implementation.
- Performance measurements describe the documented devices and configurations, not every miner or verifier.