RESIDENT COMPUTE // ONE REAL PROBLEM PER DEMO
See it work.
Voxell builds GPU-native AI: the model and its state live on the GPU, so decisions happen in microseconds, on the box, with nothing leaving the building. Each demo is the proof, on a real problem.
Measured on real problems
Financial comprehension, fraud detection, cache throughput, retrieval sorting, and state synchronization.
Bullish or Bearish? One Word Decides
A resident 0.8B reads a headline correctly - one word apart from buy or sell - in 84 ms, before the market reacts.
PROVES // Resident inference
Every Fraud Tool Has a Blind Spot
Catch coordinated copycats with no index to rebuild: 8 of 8 versus an indexed SIEM's 4 of 8.
PROVES // Coherence
Your Cache Still Runs on a CPU
Redis verbs served by a resident GPU megakernel: 217M ops/s at 8.8µs p50, with the CPU out of the data path.
PROVES // ARC
I Beat NVIDIA's GPU Sort by up to 18×
Every number measured live and verified against std::sort: up to 18× faster than NVIDIA CUB, never slower.
PROVES // MASH
RAG's Hidden Bottleneck
Every retrieval ends in a top-k sort over millions of scores. MASH sorts quantized scores 3.6× faster than CUB.
PROVES // MASH
Real-Time Cross-Device Sync - npm i @voxell/lux
Add real-time cross-device state to your app in one npm install - local-first, in your own infra, no Kafka.
PROVES // Voxell LuxHave a problem that needs decisions at machine speed?
Bring the problem and the acceptance test. We will show the result on the hardware.