The NVIDIA GPU with the most VRAM is the Rubin GPU, with 288 GB of HBM4, followed by the GB300 at 279 GB and the B300 at 270 GB. Those come in racks and eight-GPU servers. The largest card you can put in a workstation is the 96 GB RTX PRO 6000 Blackwell, and the most on a consumer card is the 32 GB RTX 5090. On QuantaCloud the RTX PRO 6000 Blackwell rents for $2.16/GPU-hr. The 141 GB H200 NVL is on demand when the catalog lists it, and everything from the B200 up is built to order.
NVIDIA GPUs ranked by memory#
The table runs from 288 GB down to 24 GB. Memory sizes and types are NVIDIA's, read on 2026-11-12 or taken from its datasheets.
| GPU | Memory | Type | Where it ships | On QuantaCloud |
|---|---|---|---|---|
| Rubin | 288 GB | HBM4 | Vera Rubin NVL72 racks, HGX Rubin NVL8 servers | Built to order |
| GB300 | 279 GB | HBM3E | GB300 NVL72 racks | Built to order |
| B300 | 270 GB | HBM3E | Eight-GPU HGX B300 servers | Built to order |
| GB300 Desktop Superchip | 252 GB | HBM3e | NVIDIA's desk-side DGX Station | Not offered |
| GB200 | 186 GB | HBM3E | GB200 NVL72 racks | Built to order |
| B200 | 180 GB | HBM3E | Eight-GPU HGX B200 servers | Built to order |
| H200 | 141 GB | HBM3e | Servers | H200 NVL at the console price when the catalog lists it, or built to order |
| RTX PRO 6000 Blackwell | 96 GB | GDDR7 | Workstations and servers | RTX PRO 6000 at $2.16/GPU-hr |
| H100 NVL | 94 GB | HBM3 | Servers | Built to order |
| RTX PRO 5500 Blackwell | 84 GB | GDDR7 | Workstations | Not offered |
| A100 80GB and H100 | 80 GB | HBM2e, or HBM3 on the H100 SXM | Servers | A100 at $1.48/GPU-hr. H100 PCIe at the console price when the catalog lists it |
| RTX PRO 5000 Blackwell | 72 GB or 48 GB | GDDR7 | Workstations | Not offered |
| RTX A6000, RTX 6000 Ada, L40, L40S | 48 GB | GDDR6 | Workstations and servers | From $0.48/GPU-hr for the RTX A6000 |
| GeForce RTX 5090 and RTX PRO 4500 Blackwell | 32 GB | GDDR7 | Consumer PCs and workstations | Not offered |
| GeForce RTX 4090 and RTX 3090, RTX PRO 4000 Blackwell | 24 GB | GDDR6X, or GDDR7 on the RTX PRO 4000 | Consumer PCs from earlier generations, and workstations | Not offered |
Prices checked 14 Nov 2026, 09:56 UTC
Three gaps in that table are worth knowing. NVIDIA's current GeForce line jumps from 16 GB on the RTX 5080 to 32 GB on the RTX 5090, so the 24 GB consumer cards are the RTX 4090 and RTX 3090 from earlier generations. NVIDIA's current lineup has no 64 GB GPU: the steps around it are 48, 72 and 80. It has no single GPU with 128 GB either. The 128 GB in a DGX Spark is system memory shared between the CPU and the GPU, and the next real step up is the 141 GB H200.
What each size really holds#
The number on the box is not what a model gets. A GPU usually reports a little less than its nominal size, a 48 GB GDDR6 card with ECC on gives up one sixteenth of it, and an inference engine such as vLLM takes 92% of what is reported by default. The budgets below are that 92%, in GiB, and count weights and KV cache only (our calculation from public nvidia-smi output).
| Card memory | Reported | Engine budget | What fits |
|---|---|---|---|
| 24 GB (RTX 4090) | 24,564 MiB | 22.1 GiB | Qwen3-8B in BF16 (15.3 GiB) with room for about 49,000 tokens of KV cache |
| 32 GB (RTX 5090) | 32,607 MiB | 29.3 GiB | The same 8B model with about 102,000 tokens. Qwen3-32B in FP8 (32.0 GiB) does not fit |
| 48 GB (RTX A6000, RTX 6000 Ada, L40, L40S) | 46,068 MiB with ECC on | 41.4 GiB | Qwen3-32B in FP8 with 9.4 GiB left, or Llama 3.3 70B in 4-bit AWQ (37.0 GiB) with 4.35 GiB left |
| 80 GB (A100, H100 PCIe) | 81,920 and 81,559 MiB | 73.6 and 73.3 GiB | Llama 3.3 70B in FP8 (67.7 GiB) with 5.9 or 5.6 GiB left, about 19,400 or 18,300 tokens |
| 96 GB (RTX PRO 6000 Blackwell) | 97,887 MiB | 87.95 GiB | Llama 3.3 70B in FP8 with 20.3 GiB left, about 66,400 tokens |
| 141 GB (H200) | 143,771 MiB | 129.2 GiB | Llama 3.3 70B in FP8 with 61.5 GiB left, about 201,000 tokens. Its BF16 weights (131.4 GiB) do not fit |
| 180 GB (B200) | 183,359 MiB | 164.7 GiB | Llama 3.3 70B in BF16 with 33.3 GiB left, about 109,000 tokens |
| 270 GB (B300) | 275,040 MiB | 247.1 GiB | Llama 3.3 70B in BF16 with 115.7 GiB left, about 379,000 tokens |
Two notes. The RTX A6000 and RTX 6000 Ada report 49,140 MiB with ECC off, a budget of 44.15 GiB, and I size against the lower figure. I have found no public nvidia-smi output for a Rubin GPU, so it has no row yet, and its lead may be smaller than the first table shows: NVIDIA's MIG guide calls the B300 a 288 GB product too, with a largest profile of 269 GB. For your own model, the VRAM guide walks through the sum and the LLM VRAM calculator does it for you.
More memory on one card, or several cards#
Several smaller cards can stand in for one large card only when the software splits the work, and that depends on the job.
| Job | Does memory pool across GPUs | How |
|---|---|---|
| LLM serving with vLLM | Yes | Tensor parallelism splits every layer across the GPUs in one VM. Without NVLink, vLLM's docs recommend pipeline parallelism instead |
| LLM serving with Ollama or llama.cpp | Yes | Ollama spreads a model across all GPUs when it does not fit one, and llama.cpp splits by layer by default |
| Training with FSDP | Yes | Weights, gradients and optimizer state are sharded, so each GPU holds a fraction |
| Training with DDP | No | Every GPU holds a full copy. DDP adds throughput, not room |
| One diffusion model in ComfyUI | No | Its weights are not split across GPUs. ComfyUI's multi-GPU nodes clone the model onto a second card to share the work, or move the text encoder and VAE there |
The ComfyUI row is the one I would check first. FLUX.2 [dev] in FP8 is a 35.46 GB model file with an 18.03 GB text encoder. Two 48 GB cards let you put the text encoder on the second GPU, but the diffusion model itself still has to live on one card. If it is larger than that card, ComfyUI offloads weights to system memory, which costs speed, and a second GPU does not change that. The ComfyUI multi-GPU guide covers those nodes.
For LLMs the trade is price against simplicity. Two 48 GB cards with ECC on give vLLM 82.8 GiB against 87.95 GiB on one RTX PRO 6000 Blackwell, and four give 165.6 GiB, enough for Llama 3.3 70B in BF16 (our calculation: 2 x 41.39 and 4 x 41.39). On 2026-09-27 two RTX A6000s in one VM cost from $0.96 an hour against $2.39 for one RTX PRO 6000 Blackwell, 40% of the price for the same 96 GB on the box (our calculation: 0.96 / 2.39). What you give up is real, though: the RTX A6000 has no FP8 or FP4 hardware, the model's layers now exchange data between two cards on every token, and the launch command gains a parallelism flag. The tensor parallel guide and the DDP and FSDP guide cover the setup. These are today's prices:
| GPU | Memory | From | Available now |
|---|---|---|---|
| RTX A6000 | 48 GB | $0.48/GPU-hr | Yes |
| RTX 6000 Ada | 48 GB | $0.79/GPU-hr | Yes |
| L40 | 48 GB | $0.94/GPU-hr | Yes |
| L40S | 48 GB | $1.09/GPU-hr | Yes |
| A100 PCIe 80GB | 80 GB | $1.48/GPU-hr | Yes |
| A100 SXM4 80GB | 80 GB | $1.50/GPU-hr | Yes |
| H100 PCIe | - | Not listed | No |
| RTX PRO 6000 Blackwell | 96 GB | $2.16/GPU-hr | Yes |
| H200 NVL | - | Not listed | No |
Prices checked 14 Nov 2026, 09:56 UTC
Launch RTX PRO 6000 BlackwellOn-demand instances are Ubuntu VMs, and stopping one deletes its disk, so copy results off first. Four-GPU and eight-GPU VMs, the H100 and the H200 NVL come and go, so check what the catalog lists before you plan around them, and ask for a build when you need them for a fixed term. The deploy guide covers the console steps.
When one GPU is not enough#
Above 141 GB per GPU the answer is a build, not a rental. A B200 has 180 GB and a B300 270 GB, eight to a server, which makes 1.44 TB and 2.16 TB in one NVLink domain (our calculation: 8 x 180 and 8 x 270). QuantaCloud builds those to order as dedicated GPU servers: you send a brief, and the configuration, lead time and terms come back in writing.
Plan a B300 buildVRAM questions#
Which GPU has the most VRAM?
NVIDIA's Rubin GPU, with 288 GB of HBM4. It ships in Vera Rubin NVL72 racks, which NVIDIA lists as available. In an eight-GPU server that has shipped for longer, the B300 has 270 GB.
What is the highest VRAM consumer GPU?
The GeForce RTX 5090, with 32 GB of GDDR7. The next GeForce down, the RTX 5080, has 16 GB. Above 32 GB you are in workstation cards: 48 or 72 GB on the RTX PRO 5000 Blackwell, 84 GB on the RTX PRO 5500 and 96 GB on the RTX PRO 6000.
Is there a 64 GB GPU?
Not in NVIDIA's current lineup. The nearest sizes are 48 GB, 72 GB on the larger RTX PRO 5000 Blackwell, and 80 GB on the A100 and H100. A DGX Spark is sold with 64 GB or 128 GB, but that is system memory shared with the CPU.
Is there a GPU with 128 GB of VRAM?
No GPU in NVIDIA's current lineup has 128 GB of its own memory. The 128 GB of a DGX Spark is shared with its CPU, and the Rubin CPX that NVIDIA announced in 2025 with 128 GB of GDDR7 is not on NVIDIA's current Vera Rubin pages. The practical choices are one 141 GB H200, or two 80 GB or 96 GB cards with software that splits the model.
Does VRAM add up across two GPUs?
Only if the software splits the work. vLLM, Ollama and llama.cpp spread an LLM across GPUs, and FSDP shards a training job. DDP and a single ComfyUI model do not gain room from a second card.
The rule I follow: buy memory on one card when the tool runs on one GPU or the model is close to the limit, and split across cheaper cards when the engine supports it and the saving is worth the extra setup. Size against the engine's budget, not the box figure, and move up one tier only when the numbers say so. The GPU catalog lists every card with its live price.