Aircloud Compute is now Lightyard. Aircloud continues as a private inference platformAircloud continues for inference

Hardware · Next generation

NVIDIA HGX Rubin NVL8

NVIDIA's eight-GPU Rubin system, on a waitlist for first allocation. The site can be planned before it ships.

Sheet 01 · HGX Rubin NVL88 GPUs per node
GPUs
8 × Rubin
GPU memory
288 GB HBM4 each
NVLink
3.6 TB/s per GPU
Power
Not final
Status
Waitlist
Drawing is illustrative · Figures are NVIDIA's
Specifications

NVIDIA HGX Rubin NVL8 specifications.

What NVIDIA has published so far.

Generation
Rubin
GPUs
8 GPUs per nodeOne NVLink domain per system, as on HGX B300.
GPU memory
288 GB HBM4 per GPU
NVLink
3.6 TB/s per GPU, sixth generationAgainst 1.8 TB/s per GPU on HGX B300.
Networking
1.8 TB/s per systemBi-directional, as NVIDIA states it.
Host CPUs
NVIDIA Vera or x86Paired with Vera, NVIDIA calls it HGX Vera Rubin NVL8.
Coolant inlet
45 °C
System power
Not finalNVIDIA hasn't published a figure yet.
Availability
Waitlist for first allocation

HGX Rubin NVL8 figures are from NVIDIA's HGX platform page and may change before volume shipments.

Site requirements

Planning a site before it ships.

NVIDIA hasn't published a rack layout or a power figure yet. This is what's known and how we plan around it.

  1. 1

    Liquid to every position

    NVIDIA lists a 45 °C coolant inlet, and the systems announced so far are liquid-cooled. The hall needs facility water at each rack position.

  2. 2

    Power with headroom

    With no published system figure, we size feeds above what HGX B300 needs and confirm the design against the figure in the order.

  3. 3

    Network sized per system

    NVIDIA gives 1.8 TB/s per system, against 1.6 TB/s per system for HGX B300. We size the fabric when the order and NVIDIA's reference design are set.

  4. 4

    A Vera or an x86 host

    The choice changes the host board, not the GPU side. It's settled in the design, with your software stack in mind.

Choosing a platform

HGX Rubin NVL8 against HGX B300.

Both are eight-GPU systems. Rubin moves to HBM4 and doubles NVLink bandwidth per GPU; B300 can be built now and runs on air.

HGX Rubin NVL8 against HGX B300.
HGX B300HGX Rubin NVL8
GenerationBlackwell UltraRubin
GPU memory288 GB HBM3e per GPU288 GB HBM4 per GPU
NVLink1.8 TB/s per GPU3.6 TB/s per GPU
Networking1.6 TB/s per system1.8 TB/s per system
CoolingAir, or liquid where the hall has itLiquid
PowerAbout 14 kW per systemNot final
AvailabilityAbout 10 to 14 weeks from signingWaitlist for first allocation

Figures from NVIDIA. Rubin figures may change before volume shipments.

Other platforms

Other NVIDIA platforms we build on.

The same four stages on each: design, source, build and operate. All hardware.

Questions

HGX Rubin NVL8 questions.

We build HGX Rubin NVL8 clusters as a deployment your team runs, a managed cluster you own and we operate, or a reserved cluster we own. Other platforms.

Get started

Tell us where you're starting.

Capital, hardware, a site or a team: tell us what you have and when you need it running. We usually reply within a few hours and come back with a written plan.

sales@lightyard.ai