RTX 4070 Compute Performance Overview | Generated by AI

Home 2026.07

Question: What are the TFLOPS and overall compute performance specs of the RTX 4070?

Answer:

The RTX 4070 (non-Super) is Ada Lovelace (AD104 die, cut-down). Specs relevant to your ML workloads:

Core specs

Compute throughput

Precision TFLOPS/TOPS
FP32 (CUDA cores) ~29 TFLOPS
FP16 Tensor (dense) ~116 TFLOPS
FP16 Tensor (2:4 sparsity) ~233 TFLOPS
INT8 (dense) ~233 TOPS
INT8 (sparse) ~466 TOPS
FP8 (dense, Ada’s native tensor precision) ~233 TFLOPS

No FP64 tensor path worth mentioning — Ada cards aren’t built for that, ignore it for your use case.

What this means for your workloads

For your GPT-2 760M-scale training and nanochat experiments on this box:

Practical rule of thumb: on 4070, target models where weights + optimizer state + activations fit under ~10GB (leaving headroom for CUDA context), use bf16/fp16 autocast, and reach for gradient checkpointing or LoRA rather than full fine-tunes once you’re above ~1-2B params.


Back Donate