FuriosaAI RNGD

FuriosaAI RNGD#

FuriosaAI’s second-generation Neural Processing Unit (NPU), RNGD, is a chip designed for deep learning inference, supporting high-performance Large Language Models (LLM), Multi-Modal LLM, Vision models, and other deep learning models.

RNGD is based the Tensor Contraction Processor (TCP) architecture which utilizes TSMC’s 5nm process node, and operates at 1.0 GHz. It offers 512 TOPS and 1024 TOPS of INT8 and INT4 performance, respectively. RNGD is configured with two HBM3 modules providing a memory bandwidth of 1.5 TB/s, and supports PCIe Gen5 x16. For multi-tenant environments like Kubernetes, a single RNGD chip can work as 2, 4, or 8 individual NPUs, each fully isolated with its own cores and memory bandwidth. RNGD supports Single Root IO Virtualization (SR-IOV) and virtualization for multi-instance NPUs.

Please refer to the followings to learn more about TCP architecture and RNGD:

RNGD Hardware Specification#
Architecture	Tensor Contraction Processor
Process Node	TSMC 5nm
Frequency	1.0 GHz
BF16	256 TFLOPS
FP8	512 TFLOPS
INT8	512 TOPS
INT4	1024 TOPS
Memory Bandwidth	HBM3 1.5TB/s
Memory Capacity	HBM3 48GB
On-Chip SRAM	256MB
Interconnect Interface	PCIe Gen5 x16
Thermal Solution	Passive
Thermal Design Power (TDP)	150W
Power Connector	12VHPWR
Form Factor	PCIe dual-slot full-height 3/4 Length
Multi-Instance Support	8
Virtualization Support	Yes
SR-IOV	8 Virtual Functions
ECC Memory Support	Yes
Secure Boot with Root of Trust	Yes