AI workstations and GPU servers

Refurbished workstations with the slots, power and memory for large NVIDIA graphics cards, for local LLMs, training and inference. For several data-center GPUs in a rack, GPU servers are configured and sourced to order.

What matters

The GPU's memory sets the size of the model

  • Graphics memoryVRAM firstA model that does not fit in VRAM spills to system memory and runs many times slower.
  • GPU generationNVIDIA, Maxwell or newerOllama supports NVIDIA cards with compute capability 5.0 and up; Kepler cards such as the K80 are not supported.
  • PlatformSlots, lanes and powerTwo Xeons or a Threadripper PRO give the PCIe slots, lanes and power supply for one or more large cards.
  • StorageNVMe for modelsModel files run from a few GB to over 40 GB; on NVMe they load in seconds.

Source: Ollama GPU support.

How much VRAM

Model size and the card it needs

Approximate memory for running a model at 4-bit quantization, the usual setting for local LLMs. 8-bit roughly doubles it, 16-bit doubles it again, and a long context adds more.

Model sizeVRAM, 4-bitCard
7B to 8Babout 4 to 6 GBAny 8 GB card
13B to 14Babout 8 to 11 GBA 12 to 16 GB card
30B to 34Babout 18 to 22 GBA 24 GB card
70Babout 40 to 50 GBTwo 24 GB cards or one 48 GB card
Compare by class

Workstations with room for GPUs

Two CPUs or a Threadripper PRO give the memory and PCIe lanes for one or more large GPUs. Pre-configured prices start at the cheapest build, with or without a graphics card (In stock shows how many have an RTX card), or build your own and pick every part, graphics card included.

  • 1-year warranty, 90-day returnsParts, labor and shipping; up to 5 years
  • Tested in MichiganAssembled and tested before shipping
  • Free shipping in the lower 48Most builds ship in 1–2 business days

Dual-socket Xeon Scalable

2 × Xeon Scalable, 1st or 2nd Gen Several GPUs, large memory

Max coresIn stockPre-configuredBuild your own

AMD Threadripper PRO

1 × Threadripper PRO 7000/9000 WX Many PCIe lanes on one CPU

Max coresIn stockPre-configuredBuild your own

Xeon W with DDR5

1 × Xeon W-2400/2500 Current platform, one or two GPUs

Max coresIn stockPre-configuredBuild your own
Outfitting a team or a lab? Tell us the seat count and the software.Request a quote
Pre-configured

Configured for AI, in stock today

Builds with the graphics card included, cheapest first.

By software

What the common tools need

SoftwareNeedsStart with
Ollama, llama.cppVRAM for the model you run; both split a model across several cardsDual-socket Xeon with one or two 24 GB cards
PyTorch, TensorFlowAn NVIDIA card with CUDA support, and system memory for your datasetsDual-socket Xeon or Threadripper PRO
LangChain, AnythingLLMThey run on top of a local model, so size the GPU for that modelAny class above
GPU servers

Several data-center GPUs in a rack

GPU servers such as the Dell PowerEdge R750xa and HPE ProLiant DL380a, and data-center cards such as the NVIDIA A100 or L40S, are not kept on the shelf: we configure and source them to order. See GPU server models or send us the spec for a quote.

Questions

What do I need to run a local LLM?

Mostly GPU memory. A 7B or 8B model at 4-bit fits on an 8 GB card; a 70B model needs about 40 to 50 GB, two 24 GB cards or one 48 GB card. The workstation around it needs the slots and power supply for those cards.

Do you sell A100 or H100 cards?

Not from stock. Data-center GPUs and GPU servers are configured and sourced to order; send us the spec and we quote it.

Can I change the configuration?

Ready-to-ship units are sold as listed. To choose the CPU, memory, graphics card and storage yourself, open a model in the table and pick a chassis under Build your own, or send us the spec and we quote it.

Is there a warranty?

Configured systems carry PCSP's free 1-year warranty covering parts, labor and shipping, extendable up to 5 years.

Planning a multi-GPU box or a small cluster?Tell us the models you want to run and the budget; we spec the machines and quote them.
Request a quote