Overview
How much VRAM each GPU has, per the vendor spec pages read on 2026-10-13: RTX 5090 32 GB, RTX 4090 24 GB, RTX 3090 and 3090 Ti 24 GB, RTX A6000 48 GB, RTX 4080 and 4080 SUPER 16 GB, RTX 4060 Ti 8 or 16 GB, RTX 4060 8 GB, and Intel Arc Pro B70 32 GB. Each row links its source; a figure the vendor page does not state is written unknown. To turn a card size into a model size, use vram-requirements and methodology.
How much VRAM does each GPU have?
Capacity, type and bus width are quoted from the linked page. Bandwidth is shown only where the vendor page states it. GB is the vendor’s label.
| GPU | VRAM | Memory type | Bus width | Bandwidth | Source |
|---|---|---|---|---|---|
| GeForce RTX 5090 | 32 GB | GDDR7 | 512-bit | unknown | NVIDIA |
| GeForce RTX 4090 | 24 GB | GDDR6X | 384-bit | unknown | NVIDIA |
| GeForce RTX 3090 Ti | 24 GB | GDDR6X | 384-bit | unknown | NVIDIA |
| GeForce RTX 3090 | 24 GB | GDDR6X | 384-bit | unknown | NVIDIA |
| GeForce RTX 4080 SUPER | 16 GB | GDDR6X | 256-bit | unknown | NVIDIA |
| GeForce RTX 4080 | 16 GB | GDDR6X | 256-bit | unknown | NVIDIA |
| GeForce RTX 4060 Ti | 16 GB or 8 GB | GDDR6 | 128-bit | unknown | NVIDIA |
| GeForce RTX 4060 | 8 GB | GDDR6 | 128-bit | unknown | NVIDIA |
| NVIDIA RTX A6000 | 48 GB (ECC) | GDDR6 | unknown | unknown (the 112 GB/s on the page is the NVLink bridge) | NVIDIA |
| Intel Arc Pro B70 | 32 GB | GDDR6 | 256-bit | 608 GB/s | Intel |
| Radeon RX 7900 XTX | 24 GB | GDDR6 | unknown | up to 960 GB/s | AMD (read 2026-10-12) |
| Radeon RX 9070 XT | 16 GB | GDDR6 | 256-bit | up to 640 GB/s | AMD (read 2026-10-12) |
AMD’s pages returned HTTP 503 on the first fetch on 2026-10-13 and were then read with a direct request, which re-confirmed the two Radeon rows (the RX 7900 XTX page lists no interface width). The AMD Radeon AI PRO R9700 (32 GB GDDR6, 256-bit, 640 GB/s peak) is on radeon-ai-pro-r9700-llm. Intel Arc B580 (12 GB GDDR6, 192-bit, 456 GB/s), A770 16GB (16 GB GDDR6, 256-bit, 560 GB/s) and Pro B60 (24 GB GDDR6, 192-bit, 456 GB/s) are on intel-arc-for-local-llm. Not listed here: Tesla P40, V100, MI60 and MI100 (the P40 and V100 are on used-datacenter-gpus; MI60 and MI100 are unknown here). Price is unknown: check current retailer price.
Which GPUs have 24GB of VRAM?
In this list: GeForce RTX 4090, RTX 3090 Ti, RTX 3090 and Radeon RX 7900 XTX, all 24 GB. Whether 24 GB is enough depends on the model: the Qwen2.5-32B Q4_K_M figure on vram-requirements is 19.8835 GB of weights + 1.0737 GB KV at 4,096 context + an assumed 1 to 2 GB overhead = 21.9572 to 22.9572 decimal GB. See 24gb-vram-llm.
Which GPUs have 16GB of VRAM?
RTX 4080, RTX 4080 SUPER, RTX 4060 Ti (16 GB variant) and Radeon RX 9070 XT. The RTX 4060 Ti also ships at 8 GB, so check the exact variant. See 16gb-vram-llm.
How much VRAM does the RTX 4080 have?
16 GB of GDDR6X on a 256-bit interface, for both the RTX 4080 and RTX 4080 SUPER, per NVIDIA’s RTX 4080 family page.
How much VRAM does the RTX 3090 Ti have?
24 GB of GDDR6X on a 384-bit interface, the same capacity NVIDIA lists for the RTX 3090.
How much VRAM does the RTX A6000 have?
48 GB of GDDR6 with ECC per NVIDIA’s workstation page. That page states no bus width or memory bandwidth for the GPU memory.
What does memory bandwidth mean for LLM inference?
Bandwidth is how many bytes per second the GPU can read from VRAM. Capacity decides whether a model fits; bandwidth is one input to how fast it runs, and no tokens-per-second figure is given here because none is read from a vendor page naming the machine. Bandwidth is bus width in bytes times the memory data rate. Worked from the Intel Arc Pro B70 row: 256 bits ÷ 8 = 32 bytes per transfer, and Intel’s 608 GB/s ÷ 32 bytes = 19 billion transfers per second per pin (19 Gbps, derived here, not a figure Intel states). NVIDIA’s pages above do not list bandwidth, so those cells stay unknown rather than being filled from a retailer or review.
Questions
How much VRAM does each GPU have?
Per the vendor pages: RTX 5090 32 GB, RTX 4090 24 GB, RTX 3090 Ti 24 GB, RTX 3090 24 GB, RTX 4080 16 GB, RTX 4060 Ti 16 GB or 8 GB, RTX 4060 8 GB, RTX A6000 48 GB, Arc Pro B70 32 GB, RX 7900 XTX 24 GB, RX 9070 XT 16 GB. The table above links each source.
How much VRAM does the RTX 4080 have?
16 GB GDDR6X, 256-bit, for the RTX 4080 and RTX 4080 SUPER, from NVIDIA’s RTX 4080 family page.
How much VRAM does the RTX 3090 Ti have?
24 GB GDDR6X on a 384-bit interface, from NVIDIA’s RTX 3090 and 3090 Ti page.
Which GPUs have 24GB of VRAM?
RTX 4090, RTX 3090 Ti, RTX 3090 and Radeon RX 7900 XTX among the cards listed here.
Which GPUs have 16GB of VRAM?
RTX 4080, RTX 4080 SUPER, RTX 4060 Ti (16 GB variant) and Radeon RX 9070 XT.
How much VRAM does the RTX A6000 have?
48 GB GDDR6 with ECC, per NVIDIA. Bus width and memory bandwidth are not on that page, so they are unknown here.
What does memory bandwidth mean for LLM inference?
It is the rate at which the GPU reads VRAM, equal to bus width in bytes times data rate. It is not the same as capacity. Tokens per second are unknown on this page.
Sources read on 2026-10-13
- NVIDIA RTX 5090: 32 GB GDDR7, 512-bit; no bandwidth.
- NVIDIA RTX 4080 family: 16 GB GDDR6X, 256-bit for 4080 and 4080 SUPER.
- NVIDIA RTX 4090: 24 GB GDDR6X, 384-bit.
- NVIDIA RTX 3090 and 3090 Ti: 24 GB GDDR6X, 384-bit each.
- NVIDIA RTX 4060 family: 4060 Ti 16 GB or 8 GB, 4060 8 GB, GDDR6, 128-bit.
- NVIDIA RTX A6000: 48 GB GDDR6 ECC; NVLink bridge 112 GB/s.
- Intel Arc Pro B70: 32 GB GDDR6, 256-bit, 608 GB/s.
- AMD RX 7900 XTX and RX 9070 XT: read 2026-10-13: 7900 XTX 24 GB GDDR6, up to 960 GB/s, no interface width listed; 9070 XT 16 GB GDDR6, 256-bit, up to 640 GB/s.
- The Qwen2.5-32B figures come from vram-requirements, which cites the Qwen2.5-32B model card and the llama.cpp quantize README.
Related
- vram-requirements sizes models; vram-calculator is planned.
- best-gpu-for-local-llm and 24gb-vs-16gb-vs-12gb-vram help choose a card.
- how-to-check-vram reads the VRAM on a card you own.
- methodology is the citation rule; changelog logs updates.
- Mac memory: what fits my RAM.