Rising component costs may adjust prices on selected items without prior notice. Supply constraints may add shipping time — we’ll notify you if affected. OEM costs and supply tightening — contact us to secure stock and pricing. For projects or bulk orders, choose “Request for Quotation” on any item.
Call our specialists(65) 6978 3502 Email our sales teamsales@sourceit.com.sg

Servers & Power

GB200 NVL4: The Rack-Scale System Nobody Talks About (and Scientists Should)

Ask which NVIDIA rack packs the most GPUs and most people will say the NVL72. They would be wrong. HPE's NVIDIA GB200 NVL4 holds up to 136 GPUs in a single rack, nearly twice as many.

It gets almost no attention, because it was not built for the chatbot economy. It was built for laboratories. And for a university, a weather agency or a research institute, it may be a more sensible purchase than the famous one.

NVL4 vs NVL72 in pictures

Compare GB200 NVL4 by HPE GB300 NVL72 by HPE
🧠 GPUs per rack Up to 136 Blackwell
72 Blackwell Ultra
🔗 GPUs linked by NVLink 4 per tray
All 72
🧱 Building block 1U tray: 4 GPUs + 2 Grace CPUs, up to 34 trays One integrated 72-GPU system
🔢 Double-precision (FP64) maths Strong
A design goal
Weaker
Reported as much weaker on Blackwell Ultra
🌐 Network between nodes Open choice
Your choice of Ethernet or InfiniBand
Proprietary
NVLink inside the rack
🎯 Built for 🔬 Simulation plus AI 💬 Very large AI models

Sources: HPE GB200 NVL4 QuickSpecs, HPE AI rack-scale systems, The Register on GB200 NVL4, Glenn K. Lockwood's B300 notes.

What it is

HPE describes the GB200 NVL4 as a dense rack-scale system for converged HPC and AI workloads. Each 1U tray carries four NVIDIA Blackwell GPUs and two Grace CPUs on a single board, with the four GPUs joined into a small NVLink domain. A 48U rack takes up to 34 trays, which is where the 136-GPU figure comes from.

Three reasons scientists should look twice

1. Science still runs on double precision

AI models happily compute in low-precision formats. Climate models, fluid dynamics and molecular simulation generally do not. They need double-precision (FP64) arithmetic. HPE positions the NVL4 for exactly these buyers, describing it as combining GPU acceleration with double-precision performance.

This is where the newest hardware is not automatically the best. As we noted in GB200 vs GB300 NVL72, independent analysis reports that the Blackwell Ultra GPU in the GB300 gives up much of its double-precision performance in exchange for AI speed. For a simulation-heavy site, the older Blackwell GPU in the NVL4 is the feature, not the compromise.

2. Most scientific jobs do not need 72 GPUs acting as one

The NVL72's defining trick is a single 72-GPU NVLink domain. That is valuable for one enormous AI model. A typical research cluster runs hundreds of jobs from dozens of groups, most of which use a handful of GPUs each. For that pattern, many small NVLink islands joined by a good cluster network is the natural shape, and it is what the NVL4 provides.

3. You are not locked into one vendor's network

On the NVL4, NVLink stops at the edge of each board. Everything between trays travels over ordinary Ethernet or InfiniBand cards. As The Register observed when the board was announced, that frees system builders such as HPE from NVIDIA's own interconnects for scaling. In HPE's case, that opens the door to its own Slingshot interconnect.

The catch

Density has a price. The Register reported a thermal design power of 5.4 kW for each four-GPU board. Multiply by 34 trays and the compute alone is above 180 kW, before networking or anything else. That is our arithmetic, not an HPE rack specification, and a real configuration may populate fewer trays. But the direction is clear: a full NVL4 rack is a direct-liquid-cooled, very-high-density system that is at least as demanding on a building as an NVL72. Read the facility checklist in Can Your Singapore Colocation Hall Take a 140 kW Rack? and assume the answer needs to be "more".

It is also a generation behind on AI-specific performance. A site whose future is mostly large language models should not choose it over a GB300.

Who should consider which

Your workload mix Lean towards
🌦️ Mostly simulation, some AI GB200 NVL4
🎓 Many users, many small and medium GPU jobs GB200 NVL4
🧬 AI models trained on simulation output GB200 NVL4 or a mix of both
💬 Mostly very large language models GB300 NVL72
🔌 Building cannot support liquid cooling Neither Look at air-cooled HPC nodes and 8-GPU servers.

The broader point is a slightly uncomfortable one for the industry: the most hyped system is tuned for one kind of customer. Research computing is a different customer, and it is allowed to buy a different machine.

Related on SourceIT

SourceIT Approved

SourceIT Approved

Researched and verified by the SourceIT team for Singapore buyers.

Share This Article
Facebook

More from our blog


← Older Post Newer Post →