MTT S5000 product page (Moore Threads)

PRIMARY VENDOR SOURCE. PH100 die on the 'Pinghu' (平湖) architecture. OAM module form factor, air-cooled and liquid-cooled variants, 8-GPU MGX platform, SGX5000 server. Publishes NO absolute TFLOPS for the card. Publishes inference throughput instead: >=4000 tok/s prefill and >=1000 tok/s decode per card, under a specific and reproducible-looking methodology (PD-disaggregated 2P4D across 6x 8-card S5000 servers, ISL/OSL 1K/2K for decode, mean TTFT <=4 s, MTP on, native FP8, third-party serving engine). ALSO the source of this repo's best vendor-baseline datapoint: the unnamed 'international flagship' comparison parts are footnoted as dense-FP16 989 TFLOPS / 3.35 TB/s, 148 TFLOPS / 4.0 TB/s, and 312 TFLOPS / 2.0 TB/s.