10,000 MTT S4000 in an AI cluster!


From China, we learn that Moore Threads is now able to connect 10,000 cards in its data centers. The aim is simple: to power artificial intelligence.

10,000 MTT S4000 cards in a data center for AI!

Moore Threads MTT S4000 - cartes chinoises

The program features huge clusters (groupings of several servers) totalling up to 10,000 boards. These include MCCX D800 server clusters with eight MTT S4000 boards, two fourth-generation Xeon processors, 1TB RAM and 15.36TB NVMe SSD storage. And to reach 10,000 boards, we need 1,250 of these machines.

Of course, the aim is to offer data centers for artificial intelligence. This will make it possible to train various LLM-type models while managing inference-type tasks.


Now, it’s worth noting that the MTT S4000 is a board designed for servers based on the Musa architecture developed by Moore Threads… The same people behind the MTT S80 and MTT S70. Each card has 48 GB of memory and 128 tensor cores, for a total of 1,280,000 tensor cores in a cluster of 10,000 cards. Finally, for communication between the cards, we find a proprietary MTLink offering a bandwidth of 768 GB/s.

However, the future remains uncertain for the brand, as US sanctions prevent the company from having access to advanced engraving processes. Of course, the company can always turn to SMIC, but the latest rumors point to high costs and limited availability too. This is all the more unfortunate given that the company managed to raise $343.7 million in a recent funding round, while securing partnerships with major Chinese telecom companies.

AMD confirms it is already working on Zen 6