NVIDIA RTX 4070 Ti — Ada architecture at the price of a V100.

Both cost $0.19/hr. The V100 has 32 GB HBM2 (900 GB/s) — 2.7x the VRAM and 1.8x the bandwidth. The RTX 4070 Ti has 12 GB GDDR6X (504 GB/s) but Ada Lovelace tensor cores with FP8 support and ~2x the compute throughput. For models that fit in 12 GB, the RTX 4070 Ti is faster. For models between 12-32 GB, only the V100 works. Choose based on model size, not price.

In INT4, yes — 13B INT4 needs ~7 GB, leaving 5 GB for KV cache. In INT8, 13B needs ~13 GB — exceeds the 12 GB VRAM. In FP16, 13B needs ~26 GB — far too large. For 13B inference on the RTX 4070 Ti, INT4 quantization is your only option, and the quality tradeoff is noticeable compared to INT8 on a larger card.

The 4 GB difference matters for two reasons: model fit and KV cache. A 7B model in FP16 (~14 GB) fits on 16 GB cards but not the 12 GB RTX 4070 Ti. And KV cache space directly limits how many concurrent conversations you can serve. At 12 GB, plan for INT8 quantization of 7B models and single-digit concurrent users. At 16 GB, FP16 7B models with more headroom become practical.

Yes, NVIDIA released an RTX 4070 Ti SUPER with 16 GB GDDR6X. On spot markets, they're listed separately. The SUPER variant addresses the 12 GB limitation but costs more. If the listing says 'RTX 4070 Ti' with 12 GB, that's the original model covered on this page. The 16 GB SUPER would compete with the RTX 4080 at the same VRAM tier.