NVIDIA GeForce RTX 5070 Ti 16GB

The GeForce RTX 5070 Ti is the second GB203 part, a 70 SM cut of the same die that gives the RTX 5080 its 84. It pairs 8,960 Blackwell CUDA cores and 280 fifth-generation Tensor cores with 16 GB of GDDR7 on a 256-bit bus, good for 896 GB/s, inside a 300 W board. For AI work the number that matters is 351.5 TFLOPS of dense FP8, or 703 TFLOPS of dense FP4, both of which NVIDIA publishes directly in the RTX Blackwell architecture whitepaper rather than leaving to be inferred from the marketing AI TOPS figure. Like every GeForce Blackwell card it runs TF32 at the plain FP32 shader rate rather than double it, a restriction that professional and datacenter Blackwell parts do not have, so treat the TF32 row as a floor rather than a headline. The 16 GB capacity puts 7B and 8B models comfortably in memory at 16-bit and much larger ones once quantised.