8-GPU SXM systems
HGX-class servers with eight GPUs linked over NVLink. The usual choice for multi-GPU training. They are heavy, draw several kilowatts, and need suitable rack power and cooling.
Home Hardware
Hardware
Data center GPUs, servers, nodes, and racks.
The hardware families Sandman Ventures buys and resells, with the specifications buyers ask about first. This page is a reference and not a stock list. Availability is confirmed by inquiry.
GPUs
NVIDIA H100, A100, V100, L40S, RTX A6000, T4. AMD Instinct MI250X and MI300X.
Systems
8-GPU SXM servers, PCIe GPU servers, single nodes, and whole racks.
Testing
Detailed analytics and a written report with every offer.
Delivery
Vienna to the DACH region and international destinations.
GPU families
What each GPU is used for.
Memory is usually the first limit for large models, and the interconnect decides how well several GPUs work together. Both are listed for each family.
| Family | Memory | Form factor and interconnect | Typical use |
|---|---|---|---|
| NVIDIA H100 | 80 GB HBM3 | SXM5 with NVLink up to 900 GB/s, or PCIe Gen5 | LLM training, large-scale inference |
| NVIDIA A100 | 40 GB or 80 GB HBM2e | SXM4 with NVLink up to 600 GB/s, or PCIe Gen4 | Training, fine-tuning, inference, HPC |
| NVIDIA V100 | 16 GB or 32 GB HBM2 | SXM2 with NVLink up to 300 GB/s, or PCIe Gen3 | Research, fine-tuning, budget inference |
| NVIDIA L40S | 48 GB GDDR6 | PCIe Gen4 | Inference, generative AI, rendering |
| NVIDIA RTX A6000 and A40 | 48 GB GDDR6 | PCIe Gen4 | Rendering, simulation, virtual workstations |
| NVIDIA T4 | 16 GB GDDR6 | PCIe Gen3, 70 W, low profile | Inference, video transcoding, virtual desktops |
| AMD Instinct MI250X | 128 GB HBM2e | OAM with Infinity Fabric | HPC and large-model training on ROCm |
| AMD Instinct MI300X | 192 GB HBM3 | OAM with Infinity Fabric | Large-memory LLM inference on ROCm |
Memory and interconnect figures are the manufacturers' published specifications for each family. They do not describe the condition of a specific unit. Condition is recorded in the quality report. Need a model that is not listed? Ask in your inquiry.
Systems
From a single node to a whole rack.
The same GPU can arrive in very different machines. The system decides power draw, cooling needs, and how many GPUs work as one.
PCIe GPU servers
Rackmount servers from vendors such as Dell, HPE, Supermicro, Lenovo, and Gigabyte, from 1U to 4U, with two to eight PCIe GPUs. Common for inference, fine-tuning, rendering, and virtualization.
GPU nodes
Single compute nodes that extend an existing rack or cluster. Fit depends on your network fabric, power, and software stack, so the inquiry asks for those details.
Whole racks
Complete racks of GPU servers, handled as one lot and one shipment. The report covers what is in the lot and states anything that was not checked.
Before you buy
Four things to check first.
Most failed deployments come from the room and the software, not from the GPU. These checks save a second round of inquiries.
Power and cooling
A single 8-GPU server can draw several kilowatts. Check circuit capacity, PDU connectors, the rack power budget, and whether your room can remove that much heat.
Form factor and interconnect
SXM modules only work in SXM servers and cannot move to a PCIe machine. PCIe cards can be linked in pairs with NVLink bridges on models that support them.
Software support
Each GPU generation supports specific CUDA, driver, and framework versions. Confirm that your stack supports the generation before you buy.
Space and logistics
Servers are deep and heavy. Check rack depth and rail type, and plan how the hardware reaches the rack, including a dock or a lift where needed.
Selling hardware
Retiring GPU servers or racks? We buy them.
Send us what you have. We review the details and reply by email with an offer, or tell you plainly if we cannot take the hardware.
- HardwareGPU model and count, and the server make and model.
- ConditionAge, known faults, and whether the system can still be powered on.
- PaperworkSerial numbers, purchase date, and photos if you have them.
- LocationWhere the hardware is and how it can be reached.
- TimelineWhen the hardware has to be gone.
Inquire
Not sure which GPU or system fits? Describe the workload.
Tell us the model size or application, the quantity you need, and where the hardware has to go. We reply with what we can offer, or say so if we cannot match it.