A cluster node rather than a standalone machine. Eight NVIDIA RTX PRO 6000 Blackwell Server Edition cards deliver 768 GB of VRAM, and a ConnectX-8 SuperNIC switch sits on the board itself: traffic goes outward to neighbouring nodes at 400 Gbps. Memory runs at the full 6400 MT/s because all 24 modules sit one per channel.
Configuration
- Platform: GIGABYTE XL43-ZX2-AAS1, 4U rack chassis, 8 GPUs
- Processors: 2× AMD EPYC 9555, 64 cores each
- Memory: 1.5 TB DDR5-6400 ECC Registered (24× 64 GB)
- Graphics: 8× NVIDIA RTX PRO 6000 Blackwell Server Edition 96 GB
- Networking: onboard ConnectX-8 SuperNIC, NVIDIA BlueField-3 DPU
- Storage: 8 hot-swap 2.5" Gen5 NVMe bays, total capacity selectable below
- Power: 4× 3200 W 80 PLUS Titanium redundant supplies
Room to grow
The machine is designed to be repeated: a second and third node stand beside it and stitch together over the same ConnectX-8. One limit is worth knowing up front: the operating range tops out at 30 °C, so the room has to hold cooling more steadily than for an ordinary server.
How to order
We match node count, memory and storage to the task. A manager quotes the configuration price and lead time. Send a request and we will reply.
Choosing between a desktop unit and a cluster?
Tell us which models you plan to run and how many people will work with them. We will calculate whether one node is enough or a GPU server makes more sense.
We supply as an official PNY partner, with the manufacturer warranty.
Message us on Telegram or leave a request.
Common questions
Why is the ConnectX-8 on the board itself?
This machine computes together with neighbouring nodes, and the network between them becomes the bottleneck. The 400 Gbps SuperNIC switch sits on the board, so it takes no PCIe slot and adds no separate card per node.
Why is the operating range only up to 30 °C?
That is the vendor figure: eight 600 W cards in an 800 mm deep 4U chassis run hotter per litre than in a deeper one. The room has to hold cooling steadily or the system will throttle. Worth checking before ordering, not after.
Does this platform support H200 NVL and NVLink?
No. The platform specification lists only eight RTX PRO 6000 Server Edition cards; H200 NVL and NVLink bridges are not mentioned at all. If the task needs a shared memory pool over NVLink, look at the G494 model with eight H200 NVL cards.
Does the memory really run at the full 6400?
Yes, because all 24 modules sit one per channel. That is the difference from the 48-slot machine: it has a higher capacity ceiling, but speed drops when every slot is filled. Here 1.5 TB and full speed do not conflict.
How many nodes make a cluster?
It depends on the model and the training window, so we name the number after discussing the task. The machine is designed to be repeated: nodes stand side by side and stitch together over the same ConnectX-8, with no separate fabric required.
