Which GPU has the most VRAM

The highest VRAM GPU in the table below is AMD's Instinct MI455X, with 432 GB of HBM4 per GPU, and the highest NVIDIA figure in it is 288 GB of HBM4 on the Rubin GPU. The largest VRAM GPU DeployGPU rents as a VM is the B300, at 270 GB per GPU on its HGX board, then the B200 at 180 GB and the H200 at 141 GB.

Highest VRAM GPUs ranked by memory per GPU

The table ranks the GPUs this page covers by the memory on one GPU, largest first, from AMD's and NVIDIA's own pages checked 12 November 2026. It is not every part either company sells.

Figures are for the whole GPU, not a MIG slice. The H100 HBM types come from NVIDIA's Hopper whitepaper, its H100 PCIe product brief and its H100 NVL product brief. NVIDIA's Blackwell Ultra datasheet gives 270 GB for the GPU on the HGX B300 board and 279 GB for the GPU in GB300 NVL72, and the B300 row uses the HGX figure. The B200 row is NVIDIA's 180 GB. AMD lists the MI455X with a 23 July 2026 launch date, and its MI400 page expects volume deployments of the Helios rack in the second half of 2026.

Memory per GPU against memory in one VM

A model that fits on one card needs no link between GPUs. A model that needs several cards shares its layers across them, and how fast that runs depends on the link, as the multi-GPU page explains. DeployGPU VMs come in fixed sizes of 1, 2, 4 or 8 GPUs, so the figure that sizes a job is the memory in the whole VM, and 92% of it is what the weights and the KV cache can share (calculated).

The B300 figure is from NVIDIA's HGX page and the B200 figure is from its DGX B200 page. The others are the per-GPU figure above times 8. For more than 8 GPUs, request a cluster quote.

48 GB: RTX A6000, RTX 6000 Ada, L40 and L40S

Four cards share 48 GB of GDDR6, and bandwidth runs from 768 GB/s on the RTX A6000 to 960 GB/s on the RTX 6000 Ada. One card holds gpt-oss-20b, 13.8 GB of MXFP4 weights, with about 30 GB left for cache, while Llama 3.3 70B in BF16, 141.1 GB, needs 4 cards (calculated). The 48 GB page compares them and the RTX 6000 Ada page covers the one with the highest bandwidth.

80 GB: A100 and H100

The A100 and the H100 SXM and PCIe cards carry 80 GB, and the H100 NVL carries 94 GB. gpt-oss-120b is 65.2 GB in MXFP4, so one 80 GB card holds it with about 8 GB left for cache (calculated). The 80 GB page compares the two families.

96 GB and 141 GB GPUs: RTX PRO 6000 and H200

The RTX PRO 6000 has 96 GB of GDDR7 and the H200 has 141 GB of HBM3e. Llama 3.3 70B in BF16 needs 2 cards of either, which leaves about 36 GB on 2 RTX PRO 6000 and about 118 GB on 2 H200 (calculated). The 96 GB and larger page compares them and the H200 page lists its offers.

180 GB and more: B200 and B300

One B200 holds Llama 3.3 70B in BF16 with about 24 GB left (calculated), and an 8x B300 VM is the one DeployGPU shape that holds the 1,560.94 GB Kimi K3 checkpoint. B300 stock comes and goes, and Auto Provision can deploy a machine when one appears. The Memory column in these tables comes from the offers feed, which can differ from NVIDIA's figures, such as 192 GB for a B200 against NVIDIA's 180 GB. See the B200 page, the B300 page and the frontier models page.

Consumer cards and MIG slices

The GeForce RTX 5090, with 32 GB, and the RTX 4090, with 24 GB, exist on DeployGPU only as Vast.ai containers in the dashboard, labelled Testing Only, and never as VMs. The smallest GPU DeployGPU rents as a VM is 48 GB. NVIDIA's pages also list MIG partitions, for example up to 7 instances of 10 GB on an H100 SXM, and DeployGPU does not offer MIG slices as VMs, so every offer is whole GPUs.