RTX 5090 AI Workstation

Yes, the 32GB of VRAM runs quantized models up to roughly 70B parameters and handles full-precision inference on smaller models. For full-precision work on the largest models, we would recommend a dual-GPU or server configuration. As a sizing rule, weights need roughly two bytes per parameter at BF16 and roughly half a byte at 4-bit, and you then need headroom on top of that for the KV cache, which grows with both context length and batch size. Tell us the specific models and context lengths you intend to run and we will confirm what fits before you commit.

Yes. We pre-install the operating system, GPU drivers, CUDA toolchain, and your preferred ML frameworks, and document the exact versions so your environment is reproducible. The unit arrives assembled and imaged, and is tested under load before delivery rather than shipped as a parts list for someone on your side to build. The version record matters more than it sounds: it is what lets you rebuild the same environment in eighteen months, or match it on a second machine.

Not for one developer working on one model at a time. A second card earns its cost when two people need the machine at once, when you want to serve a model while training another, or when a model will not fit in 32GB even quantized. It does not make a single training run on a model that already fits meaningfully faster per step in most frameworks, so buy the second card for capacity and concurrency rather than as a general speed upgrade. If either is likely within the year, start at the Dual RTX 5090 AI Workstation instead of upgrading later.

Under a desk is workable for a single-GPU tower, with two caveats. It draws enough current that it should not share a circuit with other heavy equipment, and under sustained training load the fans are clearly audible in a quiet room. If several of these go into one open-plan office, treat the combined heat as you would any other kilowatt-class equipment and check what the room's cooling can actually shift. We will ask about your intended location during the quote for exactly this reason.

Workstation platforms do not universally include a baseboard management controller, so out-of-band access is not something to assume on a tower. In practice most teams reach these machines over SSH or remote desktop, which is adequate for a system in the same building as its user. If you need power control and console access that works independently of the operating system, say so at quote stage and we will confirm what the platform supports, or point you at a rack server where IPMI is standard.

That point usually announces itself in one of three ways: models that no longer fit even at 4-bit, more than one person queuing for the GPU, or a production service that has to stay up while somebody else experiments. A workstation is the wrong shape for all three. At that stage a rack GPU server gives you more cards, redundant power, room-scale cooling and out-of-band management: see /products/gpu-servers. We would rather move you across than sell a second workstation that meets the same ceiling.

Submit a quote request with your intended workloads. A specialist will confirm the configuration, sourcing timeline, and a detailed itemized quote, typically within two business days. Have five things ready and the call is usually enough to settle the configuration: the model names and sizes you intend to run, whether the work is inference or training, roughly how large your datasets are, when you need the machine, and any facility constraints such as available circuit, desk versus machine room, and who supports it.