Nexus Compute

LLM Development Workstation

Maximum VRAM and memory bandwidth for building and serving large language models locally.

Overview

The LLM Development Workstation is specified for teams working hands-on with large language models: fine-tuning, evaluation, and local serving. Nexus Compute prioritizes GPU VRAM and memory bandwidth in this configuration so you can work with the largest models a workstation can practically hold. Everything about the specification follows from one fact: with language models, VRAM is the constraint that decides what is possible, and bandwidth is the constraint that decides how fast it feels. Weights take roughly two bytes per parameter at BF16 and roughly half a byte at 4-bit, so a 70B model is around 140GB unquantized and around 35 to 40GB at 4-bit, which is why the GPU choice here is between 64GB across two cards and 96GB on one. On top of the weights sits the KV cache, which grows with context length and with the number of concurrent requests, and which is what actually runs people out of memory during long-context or multi-user testing. When a model does not fit, the options are quantization, offloading layers to system RAM at a significant speed cost, or a bigger machine. We will tell you which of those applies to your models before you order.

Warranty & Support