FuriosaAI推出企业级NXT RNGD服务器，全球对节能型AI推理的需求激增，超越传统GPU系统-动脉网

FuriosaAI Unveils Enterprise-Ready NXT RNGD Server As Global Demand For Energy-Efficient AI Inference Surges Beyond Traditional GPU Systems

benzinga 等信源发布 2025-09-26 07:59



可切换为仅中文







FuriosaAI

弗瑞奥萨AI

's newly launched NXT RNGD server could change the economics of enterprise AI deployments, delivering high-performance inference while using far less energy than the market's most expensive GPU systems, according to the

根据消息，新推出的NXT RNGD服务器可能会改变企业AI部署的经济性，在提供高性能推理的同时，比市场上最昂贵的GPU系统消耗更少的能源。

company

公司

。

Global power demand from data centers is projected to climb 50% by 2027 and as much as 165% by 2030, driven largely by artificial intelligence workloads,

到2027年，数据中心的全球电力需求预计将增长50%，到2030年可能增长多达165%，主要由人工智能工作负载推动。

according

根据

to Goldman Sachs (NYSE:

高盛（纽约证券交易所：

全球系统（Global System）

）。

Furiosa NXT RNGD Cuts Power Use by 70% Against Nvidia's DGX H100

Furiosa NXT RNGD 将功耗降低70%，对比Nvidia的DGX H100

FuriosaAI

弗瑞奥萨AI

said the NXT RNGD – pronounced 'Renegade' – server consumes just 3 kilowatts of power, compared to more than 10 kilowatts for an Nvidia (NASDAQ:

NXT RNGD——发音为“Renegade”——服务器仅消耗3千瓦的电力，相比之下，英伟达（纳斯达克：

NVDA

英伟达

) DGX H100 server. A typical 15-kilowatt data center rack could hold up to five RNGD servers, compared to a single DGX H100.

）DGX H100 服务器。一个典型的 15 千瓦数据中心机架最多可以容纳五台 RNGD 服务器，而相比之下只能容纳一台 DGX H100。

Don't Miss:

不要错过：

A Battery That Lasts 100 Years?

一种能持续100年的电池？

Paladin's Solid-State Graphene System Is Changing How Homes and Businesses Store Power

圣骑士的固态石墨烯系统正在改变家庭和企业储存电力的方式

If there was a new fund backed by Jeff Bezos offering a

如果有杰夫·贝索斯支持的新基金提供一个

7-9% target yield with monthly dividends would you invest in it

7-9% 的目标收益率，并且每月分红，你会投资吗？

？

The company said this efficiency allows businesses to scale large models inside existing facilities without investing in costly infrastructure upgrades.

该公司表示，这种效率使企业能够在现有设施内扩展大型模型，而无需投资昂贵的基础设施升级。

The server delivers 4 petaFLOPS of compute performance in a standard rack-sized unit powered by eight RNGD cards. Each card is based on Furiosa's Tensor Contraction Processor architecture, designed specifically for AI inference workloads rather than graphics processing.

服务器在由八块RNGD卡供电的标准机架式单元中提供4 petaFLOPS的计算性能。每块卡基于Furiosa的张量收缩处理器架构，专为AI推理工作负载设计，而非图形处理。

LG AI Research Reports 2.25 Times Efficiency Gains Using Furiosa Hardware

LG AI Research 报告称使用 Furiosa 硬件效率提升了 2.25 倍

LG AI Research

LG人工智能研究院

adopted

收养

RNGD hardware for inference on its EXAONE family of large language models in July. According to FuriosaAI, LG found RNGD delivered '2.25x better LLM inference performance vs. GPUs' while also meeting demanding latency and throughput requirements.

RNGD硬件用于其EXAONE系列大型语言模型的推理。据FuriosaAI称，LG发现RNGD“相比GPU的LLM推理性能提升了2.25倍”，同时满足了苛刻的延迟和吞吐量要求。

The agreement with

协议与

LG AI Research

LG人工智能研究院

expands the deployment of RNGD servers across industries, including electronics, telecommunications, finance, and biotechnology. FuriosaAI said this validation demonstrates the system's ability to deliver measurable savings while keeping pace with performance benchmarks.

扩大了RNGD服务器在电子、电信、金融和生物技术等行业的部署。FuriosaAI表示，此次验证展示了该系统在保持性能基准的同时实现可衡量的节省的能力。

Trending: Backed by $300M+ in Assets and Microsoft's Climate Fund,

趋势：拥有超过 3 亿美元的资产和微软气候基金的支持，

Farmland LP

农田有限合伙公司

Opens Vital Farmland III to Accredited Investors

向合格投资者开放重要农田III

OpenAI Runs GPT-OS-120b On Just Two RNGD Cards

OpenAI 仅用两张 RNGD 显卡运行 GPT-OS-120b

Earlier this month,

本月初，

FuriosaAI partnered

FuriosaAI 合作伙伴关系建立

with OpenAI to showcase its gpt-oss-120b model running on just two RNGD cards. The demonstration used MXFP4 precision and highlighted RNGD's ability to support new open-weight large language models with fewer resources.

与OpenAI合作展示了其仅使用两张RNGD卡运行的gpt-oss-120b模型。演示采用了MXFP4精度，突显了RNGD以更少资源支持全新开放权重大型语言模型的能力。

The company said this proof-of-concept illustrates how enterprises can run advanced reasoning and generative models at scale without needing massive GPU clusters. By consuming less energy and fitting inside standard racks, FuriosaAI said the RNGD server helps customers manage both performance and cost constraints..

该公司表示，这一概念验证展示了企业如何在无需大规模 GPU 集群的情况下大规模运行高级推理和生成模型。通过消耗更少的能源并适配标准机架，FuriosaAI 表示，RNGD 服务器可帮助客户同时管理性能和成本限制。

RNGD Software Stack Expands to Meet Enterprise AI Demands

RNGD软件栈扩展以满足企业AI需求

Alongside the hardware launch, FuriosaAI released updates to its software development kit, versions 2025.3 and 2025.3.1. The company said the new SDK introduces inter-chip tensor parallelism, advanced compiler optimizations, improved runtime synchronization, and support for models such as Qwen 2 and Qwen 2.5..

随着硬件的发布，FuriosaAI 还推出了其软件开发工具包的更新，版本为 2025.3 和 2025.3.1。该公司表示，新的 SDK 引入了芯片间张量并行处理、高级编译器优化、改进的运行时同步功能，并支持 Qwen 2 和 Qwen 2.5 等模型。

See Also: The ECG Hasn't Changed in 100 Years —

另请参阅：心电图在100年里没有改变——

This AI Upgrade Could Help Detect Heart Disease Years Earlier

这种人工智能升级可能有助于提前数年检测出心脏病。

FuriosaAI added that expanded quantization formats and Hugging Face Hub integration make RNGD servers compatible with a broad range of enterprise use cases. The SDK is designed as a drop-in replacement for existing inference frameworks, offering customers flexibility while reducing vendor lock-in.

FuriosaAI 还补充道，扩展的量化格式和 Hugging Face Hub 集成使 RNGD 服务器能够兼容广泛的商业应用场景。该 SDK 被设计为现有推理框架的直接替代方案，在为客户提供灵活性的同时减少了供应商锁定的风险。

Furiosa Targets Early 2026 for RNGD Server Rollout

Furiosa 计划在 2026 年初推出 RNGD 服务器

FuriosaAI said the NXT RNGD server is currently sampling with enterprise customers worldwide. The system is expected to be available for order in early 2026.

FuriosaAI 表示，NXT RNGD 服务器目前正在全球范围内与企业客户进行采样。该系统预计将在 2026 年初接受订购。

By reducing energy costs, enabling data sovereignty, and offering compatibility with widely used inference frameworks, FuriosaAI said the RNGD server is positioned as an alternative for enterprises struggling with the costs of GPU-based AI deployments.

通过降低能源成本、实现数据主权以及提供与广泛使用的推理框架的兼容性，FuriosaAI 表示，RNGD 服务器定位为那些在基于 GPU 的 AI 部署成本上挣扎的企业提供了一种替代方案。