Moore Threads MTT S5000

What the card is

The MTT S5000 is a full-function GPU card from Moore Threads aimed at large-model training, inference and high-performance computing. It is built on the PH100 chip and on the company's Pinghu architecture.^[1] The phrase "full-function" describes the intended role: one card covers both training and inference work, instead of the vendor splitting its line into separate training cards and inference cards.

Compute, memory and interconnect

Moore Threads' published parameters give the card a maximum dense AI compute of 1,000 TFLOPS, together with 80 GB of memory.^[2][3] Memory bandwidth is 1.6 TB/s, and the bandwidth of the card-to-card link is 784 GB/s.^[2][3] These figures describe different bottlenecks. Dense compute sets how quickly matrix arithmetic can be executed, counting every multiply-and-add rather than assuming that many weights are zero. Memory bandwidth sets how quickly data can be fed to that arithmetic, and interconnect bandwidth sets how quickly cards can exchange the gradients and activations that move between them when a single model is spread over many devices. In large-model work the second and third numbers often decide whether a cluster can come close to using its peak arithmetic at all.

Precision range

The card supports the full range of precisions from FP8 to FP64.^[1][2] Moore Threads says it has native hardware FP8 support, and that its FP8 engine lifts training performance by more than 30% on architectures such as DeepSeek and Qwen.^[1] Precision formats differ in how many bits they spend on each number, so lower-precision formats move less data and can run faster, at the cost of accuracy that has to be managed by the training recipe. "Native" here distinguishes arithmetic executed by dedicated hardware from arithmetic emulated on top of a wider format. FP64 at the other end is the traditional format for scientific and engineering computation, which is why the card is also positioned for high-performance computing rather than for AI alone.

Reported training results

In Llama3-70B training the card reached a model FLOPs utilisation above 60%, and in DeepSeek-236B training above 40%, according to Moore Threads.^[1] Model FLOPs utilisation compares the arithmetic a model actually needs with the arithmetic the hardware could deliver at peak, so it measures how much of the machine the training loop is able to keep busy.

Scaling to large clusters

Moore Threads reported 95% scaling linearity when moving from a thousand-card cluster to a ten-thousand-card cluster.^[1] Scaling linearity describes how close a larger cluster comes to multiplying throughput in proportion to the number of cards; the shortfall is the communication and synchronisation overhead that grows as more devices have to agree on the same model state.

Video and packaging

The card's multimedia engine can decode up to 160 streams of 1080p30 video, or 40 streams of 4K30 video.^[1] Moore Threads ships the card as OAM modules in both air-cooled and liquid-cooled versions, in an eight-GPU MGX platform, and in the MTT SGX5000 server.^[1] Offering the same silicon in several physical forms lets system builders choose between standard rack modules and pre-integrated machines.

Where it sits

Moore Threads said the MTT S5000 had entered volume production, and that clusters built on it supported training of trillion-parameter models.^[2] In its 2025 results announcement the company said it had launched the flagship MTT S5000 training-and-inference GPU card during the reporting period and that the card had reached volume production.^[2] The company reported 2025 revenue of 1.505 billion yuan, up 243.37%, with the MTT S5000 as its main shipping product.^[2] That makes the card the centre of Moore Threads' current data-centre business rather than a demonstration part.