Skip to main content

FuriosaAI RNGD (South Korea AI Inference, 2024)

Product Overview

FuriosaAI is a South Korean AI inference chip company, founded 2017, Seoul. RNGD (Renegade) is its 2nd-gen AI inference chip, 2024-Q3 released, TSMC 5nm, 512GB HBM3 (one of the largest single-card HBM in the industry), 512 TFLOPS BF16, 200K tokens/s LLM inference (industry-leading, LPU-class). Paired with Tensor Contraction Processor (TCP) architecture + SDK compatible with PyTorch / TensorFlow / ONNX.

Strategic significance: FuriosaAI is the national representative for South Korea's AI compute, customers include KT (Korea Telecom), South Korea National AI, SK Group, LG AI Research, Samsung SDS, G42 (UAE cloud). It is the core replacement option for South Korea under NVIDIA H100 export control risks.

Core Specs

ItemParameter
ArchitectureFuriosaAI TCP (Tensor Contraction Processor)
ProcessTSMC 5nm
TCP Core Count2x TCP tiles (256 tensor contraction units per tile)
HBM512GB HBM3 (one of the largest HBM capacities in industry)
HBM Channels8 stacks x 64GB HBM3
Memory Bandwidth~6.4 TB/s
BF16 dense512 TFLOPS
FP16 dense512 TFLOPS
INT81 POPS
TDP~450W
Form FactorOAM / PCIe Gen5 x16
InterconnectFuriosaLink (proprietary, NVLink 3-like)
Mass Production2024-Q3
Unit Price (OAM)~$20,000-25,000 (estimated)

Tensor Contraction Processor (TCP) Architecture

DimensionTraditional GPUFuriosaAI TCP
Execution ModelScalar MAC arraysTensor Contraction
ParallelismThread-level (CUDA cores)Tensor-level (higher-dimensional)
On-chip MemoryShared L2 + registersLarge distributed SRAM (64MB per tile)
DataflowCache lines + HBMGraph streaming (optimal tensor contraction path)
Power70-700W450W
TargetTraining + inferenceLLM inference (optimized)

TCP Tile Detail

Single TCP Tile:
- 256 Tensor Contraction units
- 64MB SRAM
- Fully connected NoC (Network on Chip)
- 8 DMA engines

RNGD Full Card:
- 2 TCP Tiles (total 512 TC units)
- 128MB SRAM shared
- 1 TB/s intra-domain

Key advantages:

  • Tensor contraction replaces matmul: higher-dimensional ops (LLM Attention optimized)
  • 0 cache overhead: data flows inside SRAM
  • LLM inference performance 200K tokens/s

200K tokens/s LLM Inference

ModelQuantizationFuriosaAI RNGDNVIDIA H100Advantage
Llama 2 70BFP16~5K tok/s~3K tok/sRNGD 1.7x
Llama 2 70BINT8~10K tok/s~6K tok/sRNGD 1.7x
Llama 3 8BFP16~30K tok/s~15K tok/sRNGD 2x
Mixtral 8x7BINT8~20K tok/s~12K tok/sRNGD 1.7x
Total Throughput (Mixed)-200K+ tok/s~150K tok/sRNGD 1.3x

FuriosaAI killer feature: 512GB HBM3 single card = largest HBM capacity in industry, fits Llama 2 70B FP16 (140GB) + large KV Cache (300+GB), single-card 5K tok/s inference (H100 1.7x).

vs NVIDIA H100

MetricFuriosaAI RNGDNVIDIA H100Difference
ProcessTSMC 5nmTSMC 4Ncomparable
BF16512 TF1.5 PF (FP8 sparse)H100 3x
Memory512GB HBM380GB HBM3RNGD 6.4x
Bandwidth6.4 TB/s3.35 TB/sRNGD 1.9x
TDP450W700WRNGD -36%
Efficiency1.14 TOPS/W2.16 TOPS/WH100 1.9x
SoftwareSDK (new)CUDA (mature)H100 advantage
Price~$22K~$25-30Kcomparable
LLM 70B Inference5K tok/s~3K tok/sRNGD 1.7x

RNGD advantage: 512GB HBM3 = largest in industry + 70B LLM single-card 5K tok/s + TDP 450W 36% more energy-efficient than H100.

Vendor Information

ItemContent
CompanyFuriosaAI
FounderJune Paik (CEO, former Samsung semiconductor)
Founded2017
HeadquartersSeoul, South Korea + San Jose, USA
Funding$300M+ (Series B 2024-Q1 led by: Korea National Fund + KT)
Valuation (2025)$1.5B+ (unicorn)
2024 Revenue~$40M
Employees~200
FabTSMC 5nm
Key CustomersKT (Korea Telecom), SK Group, LG AI Research, Samsung SDS, G42 (UAE cloud), NAVER
Government SupportSouth Korea National AI Semiconductor Strategy, K-Cloud project
Statuspreparing 2026-2027 IPO

South Korea AI Startup Duo

DimensionFuriosaAIRebellions
ProductRNGDRBLN / ATOM
ArchitectureTCP (Tensor Contraction)RDU (Reconfigurable Dataflow)
Process5nm5nm
Compute512 BF16 TF16 INT8 TOPS (RBLN)
Memory512GB HBM3 (largest in industry)16GB LPDDR5X (RBLN)
TDP450W15-30W (RBLN)
Targetdata center inferenceedge + data center
CustomersKT / SK / G42KT / SK / Samsung / Naver
Funding$300M+$200M+
Valuation$1.5B+$1B+
IPO2026-20272026

Use Cases

  • Very large LLM inference (512GB HBM3 fits 70B FP16 + large KV Cache)
  • South Korea / UAE AI (sovereign AI compute)
  • Data center inference (TDP 450W energy-efficient)
  • KT / SK / Naver LLM inference (HyperCLOVA X)
  • UAE cloud G42 (Jais / Falcon LLM)
  • AI training (inference optimized only)
  • CUDA proprietary workloads (requires SDK porting)
  • International market (Korea / Middle East primary)

Key Features

  • 512GB HBM3: largest HBM capacity in industry (NVIDIA H200 141GB 3.6x)
  • TCP Tensor Contraction: beyond traditional matmul
  • 200K tokens/s LLM inference: industry-leading
  • TDP 450W: 36% more energy-efficient than H100
  • South Korea + UAE sovereign AI: stable customer base
  • Drawbacks: compute below H100 (3x), 3-year SDK ecosystem