A data center aisle with black server racks on both sides under bright linear ceiling lights.

Furiosa NXT RNGD Server

The most power efficient data center appliance for agentic systems

Name: Furiosa NXT RNGD Server
Brand: FuriosaAI
Availability: InStock
Rating: 5 (1 reviews)

Furiosa NXT RNGD Server

Most data centers have more demand for AI than power to run it. Furiosa NXT RNGD Server adds performant AI capability without a new facility or a new cooling system.

Peak FP8

4 petaFLOPS

Memory

384 GB HBM3

memory BANDWIDTH

12 TB/s

power

3 kW

Furiosa NXT RNGD AI server system featuring multiple red accelerator modules installed in a high-performance rack chassis.

Specifications

Accelerators

8× RNGD Tensor Contraction Processors (TCP)

Accelerator memory

384 GB HBM3 total, 12 TB/s

Performance

4,096 TFLOPS FP8
2,048 TFLOPS BF16
4,096 TOPS INT8
8,192 TOPS INT4

System Power

3,000 W

CPU

2 × AMD EPYC 9354 (Genoa), 64 cores total, 3.25 GHz base / 3.75 GHz boost

System Memory

1 TB DDR5

Networking

2 × 25 GbE dual-port NIC (configurable)

Storage

OS: 960 GB NVMe M.2
Data: 2 × 3.84 TB NVMe U.2

Ubuntu 24.04 LTS

Software

Furiosa SDK

Form Factor

4U rackmount, air-cooled

Up to 2× higher inference efficiency

RNGD delivers up to 2× higher token output per kilowatt across interactive serving thresholds for Qwen3-32B (FP8), maximizing concurrent user capacity within fixed power limits.

Token Throughput per Kilowatt

Short Context (1K ISL / 1K OSL)

Tokens per second per kW based on Qwen3-32B (FP8) using 1K/1K ISL/OSL comparing Furiosa RNGD (SDK 2026.4) and NVIDIA RTX PRO 6000 (vLLM 0.28.0). Measured on 4-card instances and scaled to standard 8-card server power envelopes (Furiosa NXT RNGD: 3.0 kW, RTX PRO 6000: 6.5 kW). LLM inference performance subject to change.

Token Throughput per Kilowatt

Long Context (31K ISL / 1K OSL)

Tokens per second per kW based on Qwen3-32B (FP8) using 31K/1K ISL/OSL comparing Furiosa RNGD (SDK 2026.4) and NVIDIA RTX PRO 6000 (vLLM 0.28.0). Measured on 4-card instances and scaled to standard 8-card server power envelopes (Furiosa NXT RNGD: 3.0 kW, RTX PRO 6000: 6.5 kW). Curves are shown up to 48 TPS per user, the highest per-user speed measured on both systems. LLM inference performance subject to change.

Download full benchmark paper

At 37 tokens/s per user

tok/s/kW

users/kW

Furiosa RNGD

779

21.3

Furiosa RNGD ÷ RTX PRO 6000

Throughput per kW

×1.66

CONCURRENCY

Shown when a system doesn’t reach the speed under the pointer.

Customer Stories

Scaling EXAONE Across Generations with One Inference Platform

POWER EFFICIENCY VS. A100

2.3×

EXAONE MODELS VALIDATED

5

Largest model validated, on 4 RNGD cards

236B

For a global enterprise like LG, foundation models have the potential to support AI applications across a wide range of businesses and use cases. As EXAONE evolved, LG AI Research needed inference infrastructure that could deliver power-efficient performance across successive model generations and support the organization at corporate scale.

Abstract grayscale digital art depicting flowing, curved lines and scattered small squares creating wave-like patterns on a black background.

From Solar Pro 3 API to AI-Powered Search at Nationwide Scale

Foundation Model Builders

Korea's First Commercial NPU-as-a-Service, Built on Furiosa RNGD

Cloud Service Provider (CSP)

CSP & IT Services

"Furiosa RNGD provides a compelling combination of benefits: excellent real-world performance, a dramatic reduction in our total cost of ownership, and a surprisingly straightforward integration.”

Kijeong Jeon, Product Unit Lead

"The combination of flexgrid.cloud’s modular infrastructure and FuriosaAI’s RNGD allows us to offer the last mile of high-performance AI compute where businesses need it most. This isn't just a performance win; it’s the blueprint for the future of Sovereign AI."

Sami Kibria, flexgrid.cloud

Built for every layer of the stack

Whether you develop frontier models, operate cloud infrastructure, or build vertical AI applications, Furiosa integrates natively into your operational architecture to deliver sustainable inference economics.

Dimly lit data center aisle with rows of black server racks on both sides and linear ceiling lights.

Foundation Model Provider

Co-optimize quantization and compiler graph execution with our engineering team to deliver sustained model serving across expansive context windows.

Modern cubic building with a patterned facade stands in a landscaped area with rows of trees and a cloudy sky in the background.

Cloud Service Provider

Cloud Service Provider (CSP)

We help regional cloud providers and data center operators expand AI cloud capacity and maximize compute density per megawatt using standard 180W air cooling.

Eight red GPU units mounted inside a server chassis with cables and circuit boards visible.

Enterprise & Sovereign AI

Keep your proprietary models and inference strictly behind your corporate firewall. We work with financial, industrial, and public institutions to deploy AI.

View of Earth's night side from space showing illuminated city clusters and roads across a large land area under a starry sky.

AI Application ISVs

Scale your active user base. Deliver real-time agents, search summarization, and RAG at a lower TCO while protecting SLO.

A single large, dark cloud in a moody, dimly lit sky with subtle textured shades of gray.

Neoclouds

Diversify your compute catalog beyond GPU allocations. We integrate RNGD into your virtualized infrastructure.

Person with brown hair, viewed from behind, sitting in front of a large computer screen filled with dense programming code and file directories.

MLOps/LLMOps Platforms

Accelerate your customers' models without altering their orchestrators. RNGD integrates seamlessly into existing Kubernetes clusters and serving pipelines.

Start Your RNGD Evaluation

Test RNGD against your own production SLOs. Get bare-metal access to validate your full stack on real hardware, or an OpenAI-compatible API endpoint to validate your application with nothing to install. Available worldwide.