GPUs

Newer, more efficient and architecturally improved — with 50% less memory and 49% less bandwidth than the card it follows. For local AI, a generation of progress moved the wrong way.

The only new card near this price with 24 GB, and it beats both 16 GB NVIDIA options on the specifications that govern language models. What it costs you is specific, and it is not text generation.

The same memory bandwidth as an RTX 5090 with three times the capacity — so you are not buying speed. You are buying 70B at Q8, ECC and MIG, none of which exist below it.

The slowest card reviewed here, replaced by one with 56% more bandwidth for 15 W more — and the single card on this site where an old PCIe slot genuinely matters.

Eleven articles on this site lean on this card. Twenty-four gigabytes for the least money is still unbeaten — and the risks around it have been growing while the recommendation stayed the same.

Exactly half an RTX 5070 Ti's bandwidth at exactly the same capacity — and the only way to get 16 GB of CUDA memory into 180 W. A card with one virtue, which is sometimes the only one that matters.

The card this site keeps calling the value pick, tested against that claim — including the used RTX 3090 that beats it on all three numbers language models care about.

Twenty-four gigabytes and a 56% shader advantage over a used 3090 — almost none of which language model generation uses. An excellent card whose case rests on what else you run.

Excellent for image generation and held back for language models by 16 GB — with the RTX 5070 Ti offering 93% of the same job for less. A very good card in an awkward position.

The fastest consumer card for local AI, and the only one that runs 32B-class models properly. It still does not reach 70B — and it costs 575 W to find out.