DESILO's THOR: Encrypted AI Benchmark for FHE Suite

DESILO Inc. announces that its framework, THOR, is now the reference implementation for encrypted language-model inference in the global FHE Benchmarking Suite. This means a common standard is established for comparing encrypted AI inference performance. THOR is the first framework to run an entire large language model inference process directly on encrypted data. It allows powerful AI models to be used on sensitive data without exposure risk, benefiting sectors like healthcare, finance, and the public sector. What's interesting is that THOR runs the BERT AI language model on encrypted data, completing the benchmark inference workload on a single GPU. It maintains accuracy similar to an unencrypted model, within about one percentage point. The bottom line is that THOR has drastically improved inference time, cutting it from about 10 minutes to about 2 minutes. It also accelerates matrix multiplication, a core computation, by up to 9.7 times. This allows for objective comparison and accelerates progress in secure AI.