A node built to train one large model. Eight NVIDIA H200 NVL cards deliver 1128 GB of HBM3E, arranged as two sets of four with NVLink bridges: cards inside a set exchange data directly, bypassing PCIe. A model that does not fit on a single card stays inside the machine instead of crossing the network.
Configuration
- Platform: GIGABYTE G494-ZB4-AAP2, 4U rack chassis, 8 GPUs
- Processors: 2× AMD EPYC 9655, 96 cores each
- Memory: 1.5 TB DDR5-5200 ECC Registered (24× 64 GB, one module per channel)
- Graphics: 8× NVIDIA H200 NVL 141 GB, two sets of four with 4-way NVLink
- Storage: 12 hot-swap 2.5" Gen5 NVMe bays, total capacity selectable below
- Networking: NVIDIA ConnectX-7 400GbE / NDR
- Power: 4× 3000 W 80 PLUS Titanium redundant supplies
Room to grow
The 48 memory slots hold up to 3 TB, but a second module per channel drops the speed to 4400 or 4000 MT/s: capacity is paid for with bandwidth. Eight 600 W cards require the fan kit, which adds 120 mm of depth, so plan for a metre of rack space.
How to order
We match memory, storage and networking to your workload. 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
How much rack space does the machine need?
The 4U chassis is 880 mm deep, but eight 600 W cards require the fan kit, which takes depth to 1000.5 mm. Plan for a metre in the rack or the door will not close. At 66.8 kg the machine goes in with two people or a lift.
What does 4-way NVLink give over plain PCIe?
The eight cards form two sets of four, and inside a set traffic goes over the bridge rather than through the CPU. For one large model it means a layer spread across four cards is not limited by the PCIe bus. Between the sets the exchange is ordinary PCIe.
Why 1.5 TB of memory when there are 48 slots?
The base configuration holds 24 modules of 64 GB, one per channel, running at 5200 MT/s. All 48 slots can be filled for up to 3 TB, but a second module per channel drops the speed to 4400 or 4000 MT/s. Pick by workload: capacity or bandwidth.
Can other GPUs be installed?
The vendor validates the platform for eight H200 NVL cards and for eight RTX PRO 6000 Server Edition cards, but only one model per machine: the system is only validated with a uniform GPU model. Mixing is not allowed, and the RTX PRO build is a separate listing.
How do the storage options differ?
There are twelve bays; the options differ in per-drive capacity: 1.92 TB on PCIe 5.0, 3.84 and 7.68 TB on PCIe 4.0. That gives three totals: 23.04, 46.08 and 92.16 TB. For a different mix a manager will build the configuration to fit.
