Tensor Cores are specialized GPU units for high-throughput low-precision matrix multiplication and accumulation. With FP8 inputs, they compute products rapidly but may accumulate their running sum with limited precision. Large inner dimensions can therefore introduce underflow and significant numerical error. section 5.6.3
Increasing accumulation precision addresses the architectural trade-off as follows:
- Compute FP8 products in bursts on Tensor Cores. section 5.6.3
- Use promotion to reduce accumulation errors without abandoning Tensor Core throughput for the multiplication itself. section 5.6.3



