FuriosaAI: LG AI Research EXAONE 3.5 on four RNGD cards

Vendor announcement that LG AI Research adopted RNGD. The tok/s sentence is also quoted by GamesBeat the same day: one server, four RNGD cards, batch size one, EXAONE 3.5 32B, 60 tok/s at a 4K context window and 50 tok/s at a 32K context window. Tensor parallel across cards. The comparison GPU is not named. The 2.25x performance-per-watt and 3.75x tokens-per-rack ratios are marketing and are not stored as benchmark values.