SparseDSP

Certified sparse outputs

A portfolio of certified sparse-output engines across GPU and FPGA. Vector computes only the k significant components on the GPU and Relay certifies the result; Discovery reduces a corpus to a verified candidate set and serves an open LLM; Core brings certified-exact sparse FFT to FPGA. One discipline throughout: certified results where the sparse path is provable, and a safe, authoritative fallback everywhere else.

The portfolio

Four products, one certification discipline

Every product emits a verified result inside its stated envelope, and fails closed to a safe, authoritative path outside it — dense where the device supports it. The substrate changes — CUDA GPU, verifiable LLM serving, FPGA — the discipline does not.

CUDA operator layer

SparseDSP Vector

A CUDA sparse FFT (cuSFFT) that computes only the k components that matter.

Explore Vector →

Certification & fallback runtime

SparseDSP Relay

The runtime that certifies Vector’s output and guarantees a dense fallback.

Explore Relay →

Discovery + verifiable LLM serving

SparseDSP Discovery

Reduce a large corpus to a small, verified candidate set — then serve an open LLM behind provenance-carrying, fail-closed contracts.

Explore Discovery →

FPGA · certified sparse FFT

SparseDSP Core

Certified exactness on FPGA — sparse when provable, with a certified dense fallback.

Explore Core →

CUDA sparse-FFT lane · Vector + Relay

cuSFFT · device-resident

4–5×

faster end-to-end

vs a fair pinned dense baseline

16–128×

fewer input bytes

sparse producer, structural

iso-correct

vs dense cuFFT + top-k

same seeds, same top-k target

k ≪ N

less data over the interconnect

certified payload, sparse-aware consumers

RTX + GB10

measured on NVIDIA GPUs

NVIDIA RTX 5070 Ti and GB10 / Spark

Measured against a dense cuFFT + top-k baseline, in the supported range. Outside it, Relay routes to the dense path and dense cuFFT stays authoritative. Discovery and Core carry their own measured numbers on their pages.

Measured on NVIDIA

Built and benchmarked on the NVIDIA platform

The SparseDSP GPU lanes are measured on NVIDIA RTX 5070 Ti and GB10 / Spark — Vector and Relay head-to-head against a fair pinned dense cuFFT baseline, Discovery on a single-GPU vLLM serving lane. Core is the FPGA lane. SparseTech is a member of the NVIDIA Inception program.

Request an evaluation View Research

NVIDIA, CUDA, cuFFT, GB10, RTX, and Inception are trademarks and/or registered trademarks of NVIDIA Corporation. NVIDIA Inception is a startup-enablement program; membership does not constitute NVIDIA endorsement of these results or claims.