48 GB GPUs you can deploy now

The 48 GB tier is where DeployGPU offers the most choice: four different cards, all with the same memory size and none of them requiring an H100-sized budget. The table below covers every one of them at once, with live prices.

How the four cards differ

Specifications from NVIDIA's datasheets for the RTX A6000, RTX 6000 Ada, L40 and L40S. The RTX A6000 is the odd one out twice over: it is the only Ampere-generation card here, so it has no FP8 Tensor Core, and it is the only one with any GPU-to-GPU link at all, an optional bridge that joins 2 GPUs.

The RTX 6000 Ada has no page of its own

It deserves a closer look here. NVIDIA lists the RTX 6000 Ada with the same 18,176 CUDA cores, 568 Tensor Cores and 142 RT Cores as the L40S, and its FP8 Tensor Core performance is rated at 1,457 TFLOPS with sparsity, close to the L40S's 1,466. What sets it apart is memory bandwidth: 960 GB/s, the highest of the four 48 GB cards, at the same 300 W as the A6000 and L40. For AI work that leans on bandwidth as much as capacity, it is worth comparing against the L40S directly rather than defaulting to whichever offer is cheapest. Compare RTX 6000 Ada prices on GpuPerHour.

What fits in 48 GB

These are weights-only sizes of published checkpoints, calculated rather than measured, and they apply to all four cards since the memory size is the same.

FP8 runs natively on the three Ada cards but not on the Ampere-based RTX A6000. MXFP4, the format gpt-oss ships in, has native hardware support only on Blackwell cards such as the RTX PRO 6000 and B200, so on all four 48 GB cards the serving engine handles it in software. The models hub has the same math for other current open models.

Fine-tuning fits differently than serving

The weights-only numbers above are for serving, not training. Full fine-tuning with mixed-precision Adam takes roughly 18 bytes per parameter once gradients and optimizer state are counted, which puts a 13B model well over 200 GB, far past what any single 48 GB card holds. Parameter-efficient methods change that math: the QLoRA paper reports fine-tuning a 65B parameter model on a single 48 GB GPU, and LoRA-style methods generally run at roughly 2 times the parameter count in gigabytes rather than 18 times. On this tier, whether a fine-tuning job fits on one card almost always comes down to which of those two methods it uses.

Moving up or down a tier

Once a job regularly needs more than one 48 GB card, the 80 GB tier usually simplifies things by fitting the same model on a single A100 or H100 instead. There is no dedicated tier below 48 GB on DeployGPU: the RTX 4090 and RTX 5090 exist only on Vast.ai, labelled testing only, and are not part of this comparison.