Overview & Capability
Custom silicon AI chip provider delivering world-record throughput and sub-100ms time-to-first-token on Llama 3 70B and 405B models via SN40L RDUs.
Key Highlights
- Over 400+ tokens/second on Llama 3.3 70B
- Full precision 16-bit inference without quantization degradation
- OpenAI-compatible developer API
Robbie's Independent Take
Verified: 2026-10-11
Excellent candidate for local execution with zero cloud lock-in. For 7B-8B parameter models, budget at least 8 GB of unified memory or VRAM (Q4_K_M quantization). For 32B-70B models, aim for 24 GB+ to maintain interactive throughput without CPU offload penalties.
Alternatives & Replaces
- Standard cloud GPU inference