Yes. Training a 13B parameter model with LoRA in fp16 on the RTX 5090 32GB is achievable within 32GB VRAM at workable batch sizes. In fp16, the 13B model weights alone occupy approximately 26GB — with LoRA adapters, optimiser states, and activations, total VRAM consumption at a batch size of 1 with short sequences sits in the 28 to 32GB range. At longer sequences or larger batch sizes, you may need gradient checkpointing to fit within 32GB, but the base model weights remain in fp16 throughout rather than being quantised. Full fp16 fine-tuning without any quantisation of the base weights is what 32GB VRAM enables at 13B scale. For models above 13B in fp16, or for long context fine-tuning at 13B where activations are large, call Kevin and describe your model size and sequence length before ordering.

Both machines use the same RTX 5090 32GB GPU and the same Ryzen 9 9950X processor. The single difference is system RAM: the Cobalt has 64GB DDR5, the G6 Cobalt Max has 128GB DDR5. The reason for 128GB is CPU offloading: DeepSpeed ZeRO-Offload and PyTorch FSDP move optimiser states, gradients, or full model parameters to system RAM to train models larger than 32GB VRAM on a single GPU. This process requires large amounts of system RAM — 64GB is insufficient for offloading at scale, and 128GB provides the capacity for the offloaded states alongside the rest of the system's memory requirements. If your workflow does not use CPU offloading and your model fits within 32GB VRAM, the Cobalt is the correct machine. If your workflow requires ZeRO or FSDP offloading, the Cobalt Max is required.

The RTX 5090 has a TDP of 575W. Under sustained GPU training or rendering load, the GPU draws close to that figure continuously. The Ryzen 9 9950X under sustained data loading and preprocessing workloads adds another 125 to 200W. Combined system draw approaches 800W under sustained training conditions. An 850W PSU would operate at approximately 94% of rated capacity under this load — beyond the safe working range for long-term reliability. The 1000W Corsair RM1000e operates at around 80% of rated capacity under peak load — within the efficient operating band for an 80+ Gold unit that runs overnight training jobs — with ATX 3.1 transient headroom for the RTX 5090's power spikes. Sustained operation near or above a PSU's safe working capacity reduces component longevity and increases instability risk during long jobs.

For researchers and practitioners who train regularly, the answer is usually yes over a medium-term horizon, though the calculation depends on usage patterns. Cloud GPU instances with 80GB HBM capacity — the type needed to train 13B models in fp16 without quantisation — typically cost between £2 and £4 per GPU-hour on major cloud platforms as of 2026. A training run of 10 hours costs £20 to £40 at that rate. If you run 10 such jobs per month, cloud costs reach £200 to £400 per month — £2400 to £4800 per year. The G6 Cobalt at around £4000 reaches the equivalent cloud spend in one to two years at this usage level, and you have the machine for five to seven years. The calculation shifts in cloud's favour for very occasional training runs, or where the flexibility of scaling to multi-GPU cloud instances is necessary. Neither answer is universal — call Kevin and describe your training frequency and typical job length, and he will give you an honest assessment.

Build time is 3 to 5 working days from order confirmation, including the 24-hour stress test at sustained RTX 5090 training load. Delivery to UK mainland addresses is free and fully tracked. RTX 5090 stock is subject to availability — call Kevin on 01902 714533 before ordering to confirm current stock and the build timeline. If you have a specific project start date, Kevin will tell you whether the timeline is achievable before you place the order.