HP’s ZGX Fury AI position is now disposable to order, and HP paired the readiness news pinch a collaboration pinch Red Hat and NVIDIA to put Red Hat AI Factory pinch NVIDIA connected apical of it. The ZGX Fury is HP’s return connected NVIDIA’s DGX Station design, built astir the GB300 Grace Blackwell Ultra Desktop Superchip pinch 748GB of unified representation and up to 20 petaFLOPS of FP4 compute, and HP is positioning it little arsenic a individual workstation than arsenic a shared conclusion container that a department, a mill floor, aliases a branch agency tin tally without a information halfway down it. We person 1 successful the laboratory now, truthful a afloat reappraisal is coming; this is what HP has said truthful far.
HP ZGX Fury Hardware: One GB300 Superchip, 748GB of Memory, Tower aliases 5U
The halfway of the strategy is the aforesaid silicon we tested successful the MSI XpertStation WS300: 1 Blackwell Ultra GPU pinch 252GB of HBM3e astatine 7.1TB/s, tied to a 72-core Grace CPU pinch 496GB of LPDDR5X complete NVLink-C2C. HP’s spec expanse fills successful specifications the level announcements skipped. The CPU representation is 4 128GB SOCAMM modules delivering 396GB/s, and the Grace CPU is soldered to the big processor module alternatively than socketed. The 2 pools adhd up to the 748GB coherent abstraction that lets the GPU reside CPU representation directly, which is what makes trillion-parameter conclusion and fine-tuning of models successful the 100 cardinal parameter people imaginable connected a azygous box. HP’s footnote connected those exemplary sizes is that the harness quantizes astatine FP4.
Two embedded M.2 slots bent disconnected the Grace CPU connected PCIe 5.0 and clasp the operating strategy successful a package RAID 1 mirror. Two much M.2 slots travel disconnected the PCIe move wrong the ConnectX-8 SuperNIC and service arsenic a RAID 0 information volume, pinch 2TB aliases 4TB of self-encrypting NVMe chosen astatine purchase.
Networking is the ConnectX-8 pinch 2 QSFP112 ports astatine 400Gbps each, which tin nexus 2 ZGX Fury systems together, positive a 10GbE RJ-45 for the big and a abstracted 1GbE RJ-45, Mini-DP, and micro-USB for the BMC. The remainder of the I/O is workstation-normal: 2 USB-A and 2 USB-C ports up front, 4 much USB ports astatine the rear, audio jacks, a Kensington slot, and a C20 inlet for the powerfulness cord. There is nary show output from the GB300 itself; HP offers an optional NVIDIA RTX PRO GPU to thrust monitors truthful the Blackwell Ultra GPU stays dedicated to inference. The chassis is simply a building that besides ships pinch rails for a 5U rack slot, and HP uses liquid cooling pinch optimized airflow, which matches what we recovered connected the MSI unit, wherever a 1,400W-rated loop kept the GPU astatine 71C nether afloat load.
HP ZGX Fury Software: Ubuntu, Z Runtime, and Red Hat Certification
HP ships the ZGX Fury pinch Ubuntu 24.04 LTS and NVIDIA’s AI developer tools, an NVIDIA-approved partner BIOS and BMC firmware, and 2 HP-specific layers. HP Z Runtime is simply a pre-installed command-line instrumentality for pulling, serving, and managing models locally, and HP Z Toolkit adds open-source frameworks, MLflow research tracking, and Ollama testing pinch find and sync crossed ZGX systems. The thought is that a squad prototypes connected a ZGX Nano and moves the aforesaid workflow to a ZGX Fury erstwhile it needs much memory, much throughput, aliases much concurrent users.
The caller portion is Red Hat. HP says the ZGX Fury is certified for Red Hat Enterprise Linux and listed successful the Red Hat Ecosystem Catalog today, and the 2 companies are processing what HP calls an open, enterprise-grade AI level that runs Red Hat AI Factory pinch NVIDIA connected the ZGX Fury. Red Hat AI Factory pinch NVIDIA is Red Hat’s packaging of RHEL, OpenShift, and Red Hat AI Enterprise pinch NVIDIA AI Enterprise for deploying models, agents, and applications crossed hybrid cloud. On the ZGX Fury, HP says the operation is meant to trim situation setup clip and deployment risk, amended GPU utilization done optimized CUDA libraries, scheduling, and multi-GPU workload orchestration, and fto developers offload compute to the container without changing their existing workflows. The level is besides being designed to tally aggregate AI workloads connected 1 strategy pinch workload isolation and governance, which is really HP gets from a deskside instrumentality to thing IT tin negociate arsenic separator infrastructure.
“The early of AI is moving person to wherever group work, machines run and captious decisions are made,” said Jim Nottingham, Senior Vice President and Division President of Advanced Compute and Solutions astatine HP. “Together pinch Red Hat and NVIDIA, HP is extending endeavor AI from the information halfway to the separator pinch an open, enterprise-grade conclusion level designed to springiness customers greater choice, power and consistency arsenic they deploy section AI factories.” Chris Marriott, Vice President of Enterprise Platforms and Solutions astatine NVIDIA, framed it the aforesaid way: moving “powerful AI locally while maintaining the security, scalability, and consistency enterprises demand.”
The ZGX Fury is orderable now done HP; pricing was not disclosed successful the announcement. The Red Hat AI Factory integration is simply a planned solution alternatively than a shipping SKU, and HP says customers will beryllium capable to measure it successful a sandboxed situation connected HP devices earlier moving to production, pinch timing, eligibility, and supported configurations still to come. The competitory image is filling successful quickly: MSI’s WS300 is shipping connected the aforesaid superchip, and AMD’s Threadripper Halo Station is aimed astatine the aforesaid workloads. Our ZGX Fury reappraisal will put HP’s type of the level done the aforesaid exemplary and testing we ran connected the MSI.
English (US) · 
Indonesian (ID) · 
