NVIDIA GeForce RTX 40 Series

Pick your card

Per-card guides below — the two cards that matter most for LLMs are marked.

What makes this series different

• VRAM moved down-market. For the first time, 16 GB is a mid-range spec (4060 Ti 16G, 4070 Ti Super). That's a real capability tier for LLMs — 13B at Q6/Q8 and MoE ~30B models single-card — at prices a 3090 never touched.
• No NVLink, ever. Ada dropped the bridge entirely. 4090 pairs scale over PCIe 4.0 with a ~25–35% split tax. If multi-GPU is the plan, the RTX 3090 remains the tile to buy — this is the single biggest gotcha in the series.
• GDDR6X, but narrower buses below the top. The 12 GB cards use a 192-bit bus (504–672 GB/s); only the 4080/4090 get 256/384-bit. The 4060 Ti stays at 128-bit, which caps its t/s despite the 16 GB.