The RTX 5060 Ti 16GB bandwidth ceiling gives way to a practical range
Decode is weight-bound, so the useful buyer number for the RTX 5060 Ti 16GB is the practical range, not the 448 GB/s bus ceiling.
| RTX 5060 Ti 16GB | |
|---|---|
| theoretical bandwidth | 448 GB/s |
| practical sustained bandwidth | 380-420 GB/s |
| pin speed | 28 Gbps |
| bus width | 128-bit |
| 7B FP16 decode | 32 t/s theoretical, ~25 t/s practical |
| INT8/FP8 decode | ~65 t/s theoretical, 50-55 t/s practical |
| INT4 decode | ~130 t/s theoretical, 95 t/s practical |
| bandwidth vs 4060 Ti | +55% at +15 W TDP |
GIGAGPU's page derives the 448 GB/s number from 28 Gbps per pin on a 128-bit bus. By access pattern, the practical sustained range sits at 380 to 420 GB/s.
A 7B FP16 model lands near 25 tokens per second, INT8 or FP8 near 50 to 55, and INT4 near 95.
The +55% bandwidth over the 4060 Ti at +15 W TDP is the architectural point, but it assumes the model fits in memory and the runtime avoids stalls.
Power draw, thermals, driver overhead, and the exact quantization path are not measured.
The buyer should price the card on the practical range, not the bus ceiling.
Use the practical range as the buyer number, keep the peak as a bus calculation, and flag the missing power and runtime data.
After nvidia.com. We did not report this. The pictures, if any, are theirs.
