Specifications
Model
KRS8500V3
Platform
NVIDIA GB300 NVL72 accelerated by Blackwell Ultra
GPU Count
72x NVIDIA B300 per rack
Compute Tray
2x Grace CPUs + 4x Blackwell Ultra GPUs
Compute Tray Count
18 per rack
GPU Memory
288 GB HBM3e per GPU
Switch Tray
2x NVLink N5600_LD per tray, 9 per rack
Switch Bandwidth
14.4 TB/s
Cooling
CPU, GPU and ConnectX-8 liquid-cooled; remainder air-cooled
Storage per Tray
8x E1.S NVMe SSD + 1x M.2 NVMe SSD
North-South Networking
1x FHFL PCIe Gen5 x16 (BlueField-3)
East-West Networking
2x ConnectX-7 mezzanine carrying 2x ConnectX-8 each, 4x 800G OSFP
Fans
CPU region 8x 12 V 4056 hot-swap with N+1 redundancy
Power Shelves
(6+2)x 33 kW
Bus Bar
1400 A
NVLink Domain
72x 1 configuration
NV OOB Switch
Option of 2x, 3x or 4x SN2201_M
NVL Cartridge
4 per rack
Rack Manifold
44RU, bottom-feed or top-feed
Dimensions
600 x 2285 x 1200 mm
Weight
1,360-1,590 kg
Overview
The Aivres KRS8500V3 packages the NVIDIA GB300 NVL72 reference architecture into a compact 600 mm-wide Open Rack v3 cabinet. Each of the eighteen 1U compute trays carries two Grace CPUs and four Blackwell Ultra GPUs at 288 GB of HBM3e per GPU, giving the rack roughly 20 TB of pooled high-bandwidth memory inside a single NVLink domain.
Scale-up switching is handled by nine trays of dual NVLink N5600_LD silicon delivering 14.4 TB/s, while scale-out networking is split between two ConnectX-7 mezzanines - each carrying two ConnectX-8 devices, fanned out to four 800G OSFP ports - and a single FHFL PCIe Gen5 x16 slot for a BlueField-3 DPU.
Every compute tray also carries eight E1.S NVMe SSDs and an M.2 boot device, so local scratch capacity scales with the GPU count instead of relying on a separate storage tier. Eight 12 V 4056 hot-swap fans with N+1 redundancy cool everything outside the cold plate loop.
Key Benefits
Full GB300 NVLink domain in a narrow cabinet: 600 mm width fits denser row layouts than the wider reference racks. Liquid cooling on the heat-producing silicon only: CPU, GPU and ConnectX-8 sit on cold plates while the rest stays air-cooled, trimming CDU sizing. Per-tray E1.S storage: local NVMe scales with compute, which matters for checkpointing at trillion-parameter scale.
Applications
Trillion-parameter model training, large-scale mixture-of-experts inference, checkpointing-heavy HPC, national and sovereign AI programmes, and AI factory builds based on the NVIDIA GB300 NVL72 reference design.
Request a Quote — Aivres KRS8500V3 — NVIDIA GB300 NVL72 Exascale AI Rack
QS Compute — global B2B supply of AI computing hardware, edge AI systems and accelerators. Volume pricing, 15-day sample lead time.
Get Your Quote →