FuriosaAI RNGD
RNGD (pronounced "renegade") is FuriosaAI's second-generation, inference-only accelerator: a passively cooled, dual-slot PCIe Gen5 card built on TSMC 5nm, with 48 GB of HBM3 at 1.5 TB/s, 256 MB of on-chip SRAM and a 180 W TDP. It is not a GPU. Its Tensor Contraction Processor (TCP) architecture treats tensor contraction, not 2D matrix multiply, as the primitive, and the Furiosa compiler maps whole models onto it, so it runs only what the Furiosa SDK supports. FuriosaAI sells it as a single card and as the NXT RNGD Server, an 8-card, 4U, roughly 3 kW air-cooled appliance. Where it fits in a build: power-capped, air-cooled racks running mainstream open-weight LLM inference, where tokens per watt matter more than per-card peak throughput or CUDA compatibility.
- Model coverage is the gating question, not FLOPS. RNGD runs only what the Furiosa compiler and Furiosa-LLM support; check that your exact model, quantization and context length are on the supported list (precompiled artifacts are published up to 32K context) before you buy, because there is no CUDA fallback.
- 48 GB per card is small for large models. Anything past roughly a 30B model at FP8 needs tensor parallelism across cards, and that traffic rides PCIe Gen5 (about 128 GB/s per card), not an NVLink-class fabric. Furiosa says it overlaps inter-chip DMA in the compiler, but large-batch scaling data has not been published.
- The efficiency case is real but vendor- and customer-reported. LG AI Research reported 2.25x better LLM performance per watt than a GPU setup (the GPU is unnamed in Furiosa's post; The Register reports it was A100) and about 50 to 60 tokens/s at batch 1 for EXAONE 3.5 32B in FP16 on 4 cards. Re-run your own workload before you accept the ratio against current GPUs.
- The TDP moved from 150 W (Hot Chips 2024, v2025.1 docs) to 180 W in the current docs. Budget power and airflow at 180 W per card; the NXT RNGD Server is specified at 3 kW for 8 cards plus a dual-EPYC host.
- The fit is a power-capped, air-cooled room: five 3 kW servers per 15 kW rack (vendor claim) without liquid cooling. If the site has liquid cooling and power headroom, per-rack throughput on HBM GPUs is likely higher; RNGD's argument is tokens per watt and per rack under an air-cooling limit.
Is FuriosaAI RNGD a GPU?
No. It is an inference-only accelerator built on FuriosaAI's Tensor Contraction Processor architecture: 8 processing elements of 64 slices each, 256 MB of on-chip SRAM, and 48 GB of HBM3. It does not run CUDA; models are compiled with the Furiosa SDK and served with Furiosa-LLM, which is vLLM-compatible and exposes an OpenAI-compatible API.
What are the RNGD card's key specs?
512 TFLOPS FP8, 256 TFLOPS BF16, 512 TOPS INT8 and 1,024 TOPS INT4; 48 GB HBM3 at 1.5 TB/s; 256 MB SRAM; PCIe Gen5 x16; 180 W TDP; passively cooled, dual-slot, full-height, 3/4-length card with a 12VHPWR connector; up to 8 instances via SR-IOV.
What is the NXT RNGD Server?
FuriosaAI's turnkey 4U, air-cooled appliance with 8 RNGD cards (384 GB HBM3, 12 TB/s aggregate, 4,096 TFLOPS FP8), two AMD EPYC 9354 CPUs, 1 TB DDR5, NVMe storage and 2x 25 GbE dual-port NICs, rated at 3,000 W. FuriosaAI says five fit in a 15 kW air-cooled rack.
How much does RNGD cost?
FuriosaAI does not publish a price for the card or the server; both are quote-only. An informal figure of around $10,000 per card was reported from side conversations at Hot Chips 2024, which is not an official price.