Which GPUs meet Unsloth's fine-tuning VRAM minimums

An RTX 4090 (24 GB) meets Unsloth’s minimums for QLoRA up to 27B and LoRA up to 9B. An RTX 5090 (32 GB) reaches QLoRA 40B and LoRA 11B. An A100 or H100 80GB reaches QLoRA 90B and LoRA 32B. The table below runs the same check for every discrete GPU in the catalog.

How the check works: a GPU meets a row when its memory is at least the number Unsloth printed. Card memory is the marketed size (24 GB is counted as 24 GiB, the catalog’s convention for discrete VRAM), so 24 GB meets a 24 GB minimum but not 26 GB. The Unsloth figures were checked on 2026-10-22 and are their published minimums, not a measurement by vramcalculator.io. The full source table is in QLoRA and LoRA VRAM by model size.

Quick checks

• RTX 4090 24GB: meets 9B LoRA at 24 GB. Does not meet 32B QLoRA at 26 GB or 11B LoRA at 29 GB.
• RTX 5090 32GB: meets 40B QLoRA at 30GB. Does not meet 14B LoRA at 33 GB.
• A100 80GB SXM: meets 32B LoRA at 76 GB and 90B QLoRA at 53GB. Does not meet 40B LoRA at 96GB or 405B QLoRA at 237 GB.

Which GPUs meet each minimum

“All 28 discrete GPUs” means every card in the list further down, since the smallest is 8 GiB. “None” means even the 80 GiB cards (A100 and H100) fall short.

The 28 discrete GPUs

Sorted by memory, then name.

Unified memory, not scored

These machines have plenty of memory, but it is unified memory shared with the CPU and operating system, not discrete VRAM, so they are listed without a verdict. For how a Spark compares with a discrete card, see DGX Spark vs RTX 5090, and for the general rule, two GPUs and unified memory.

Calculator version not recorded for this article. model data 2026-10-02.1. Sources last checked 2026-10-22.