FuriosaAI RNGD (Renegade)

Added in the Phase-A gap-fill for the Oct-14 research ingest. RNGD is FuriosaAI's second-generation, LLM-oriented NPU (the part that actually targets transformer serving, unlike the vision-NPU Warboy). TSMC 5nm, 180W single PCIe card, mass production January 2026. Vendor ships a furiosa-llm pip package that serves OpenAI-compatible inference (furiosa-llm serve <model>). UNOBTAINIUM for a hobbyist: hardware access is request/eval only (furiosa.ai/signup, bare-metal or API-endpoint eval for enterprise), no retail channel, and a public tok/s table was still absent as of 2026-10-15. On 2026-10-16 the vendor blog had a Hot Chips range (2000-3000 tok/s aggregate on a roughly 10B model, single card), an early GPT-J 6B figure (about 12 queries/s, single card), an LG measurement (EXAONE 3.5 32B, four cards, batch 1: 60 tok/s at 4K and 50 tok/s at 32K), and approximate four-card Qwen3-32B aggregates (about 1500 tok/s at batch 64, about 1100 tok/s at batch 32). See the furiosa benchmark records. None of those is a single-card batch-1 decode number for a current open model. Real serving stack, gated hardware - not a buyable cheap-inference card. Product page (2026-10-16) also states 256 MB on-chip SRAM and 512 TFLOPS FP8 peak.