DeepSpeed ZeRO-Offload is a technique for training models whose total memory footprint exceeds GPU VRAM capacity. During training, the states that are not needed for the current computation step — the optimiser states between parameter updates, the gradients between the backward pass and the parameter update — are moved from GPU VRAM to system RAM. The GPU is then freed to hold only the parameters and activations needed for the current step. This allows training of models whose full state footprint significantly exceeds GPU VRAM, as long as system RAM is large enough to hold the offloaded states and the CPU is fast enough to manage the data movement without becoming the bottleneck. You need ZeRO-Offload when your model's combined parameter, gradient, and optimiser state footprint exceeds your GPU VRAM — typically when training 30B+ parameter models in fp16 on a single GPU. If your model fits within 32GB VRAM, you do not need ZeRO-Offload, and the G6 Cobalt with 64GB is the correct machine.

The answer depends on your model size and which FSDP sharding strategy you use. With FSDP full sharding and CPU offloading enabled, the optimiser states for a 30B parameter model in fp16 with an Adam optimiser occupy approximately 120GB in system RAM under ZeRO stage 2 offloading. The operating system, Python environment, data loader, and other processes add additional RAM consumption. 128GB is the practical minimum to hold the offloaded states for a 30B parameter model alongside normal system processes. For models significantly above 30B, or for complex offloading configurations where more states are offloaded simultaneously, RAM requirements increase further. Call Kevin and describe your model size, parameter count, and offloading configuration before ordering — he will confirm whether 128GB covers your specific setup.

Both machines use the same RTX 5090 32GB GPU and Ryzen 9 9950X processor. The single difference is system RAM: the G6 Cobalt has 64GB DDR5, the Cobalt Max has 128GB DDR5. 128GB is required when your training pipeline uses DeepSpeed ZeRO-Offload or PyTorch FSDP CPU offloading to train models whose full state footprint exceeds 32GB VRAM — 64GB system RAM is insufficient for the offloading process at 30B+ scale. If your model fits within 32GB VRAM and you do not use CPU offloading, the Cobalt with 64GB system RAM is the correct machine and the Cobalt Max adds cost without benefit for your workflow. The case is also different: the Cobalt uses a Corsair 5000D, the Cobalt Max uses an APNX C1.

Yes, CPU offloading adds latency compared to keeping the full training state in VRAM, because data movement between GPU and system RAM takes time that would otherwise be used for computation. The throughput reduction depends on the offloading configuration — ZeRO stage 2 with only optimiser state offloading has less overhead than full parameter offloading — and on the CPU and memory bandwidth available on the host side. The Ryzen 9 9950X's DDR5 memory bandwidth minimises the host-side latency in this data movement, but some throughput reduction compared to a pure-VRAM training configuration is unavoidable. The trade-off is that CPU offloading makes it possible to train models that would otherwise require multiple high-end GPUs or cloud instances with HBM-based VRAM. For researchers whose model genuinely exceeds 32GB VRAM, the throughput trade-off of offloading on a single RTX 5090 is typically preferable to the cost and complexity of the alternatives.

Build time is 3 to 5 working days from order confirmation, including the 24-hour stress test and the 128GB DDR5 memory profile verification applied before dispatch. 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, particularly if you have a research project start date in mind.