Chinese Moore Threads Unveils Chunxiao GPU: 4,096 SPs, GDDR6, PCIe Gen 5
Moore Threads Intelligent Technology is a leading graphics processor developer in China. announced Next-generation GPUs for gaming, artificial intelligence, and data center workloads. The MTT S80 gaming graphics card and MTT S3000 server board promise computing performance comparable to Nvidia’s GeForce RTX 3060 Ti, but real-world performance in gaming and professional applications has yet to be tested.
Chunxiao GPU
Moore Threads’ new graphics processor is based on the company’s Chunxiao architecture, which supports FP32, FP16, and INT8 precision, and is compatible with the company’s MUSA computing platform and application programming interface (also works with standard API ). The MTT Chunxiao GPU runs at 1.80 GHz to 1.90 GHz and packs 4,096 stream processors, 128 tensor cores, 256 texture units, and 256 rendering outputs. The GPU has a 256-bit memory interface and with a data transfer rate of 14 GT/s he can use GDDR6 memory. As for the host bus, it features 16 PCIe Gen5 lanes and fully supports the SR-IOV specification for PCIe virtualization in server environments. A GPU that supports up to 32-way partitioning, useful for rendering Android games.
The chip also comes with a fairly powerful video engine that supports AV1, H.264 and H.265 codecs for up to 8K video and can decode up to 32 streams at 1080p30 resolution. In terms of output, the GPU has 4 display engines supporting resolutions up to 8Kp30, along with 3 DisplayPort 1.4 and 1 HDMI 2.1 interfaces.
Moore Threads states that the Chunxiao GPU could deliver 14.4 FP32 TFLOPS or 15.2 FP32 TFLOPS depending on the clocks. This matches the single-precision compute performance of Nvidia’s GeForce RTX 3060 Ti GPU. On paper, this puts you in the list of the best gaming graphics cards available today, but how the Chunxiao GPU performs in the real world remains to be tested.
The Moore Threads Chunxiao graphics processor, on the other hand, is fairly complex with 22 billion transistors as it is targeted at AI, gaming and data center workloads. To put the numbers into context, Nvidia’s GA104 has 17.4 billion transistors, while AMD’s Navi 21 has 26.8 billion transistors. Moore Threads does not disclose which process technology it uses to manufacture its Chunxiao graphics processors.
Introducing two products
For now, Moore Threads plans to offer two products based on the Chunxiao GPU. MTT S80 graphics card With 14.4 FP32 TFLOPS throughput and 16GB memory, MTT S3000 Server Card It has 15.2 FP32 TFLOPS of computing performance and 32 GB of memory.
Moore Threads hasn’t revealed the power consumption of the MTT S80 and MTT S3000 products, but the former comes with a fairly sophisticated cooling system with three fans.
One of the main advantages offered by Moore Threads Chunxiao is its wide compatibility. Works with a variety of client and data center hardware platforms (Arm, Ampere, Intel, etc.) and operating systems. Compatible with Microsoft’s DirectX (so it comes with the appropriate drivers for Windows), Khronos Group’s OpenGL/OpenGL ES, his own MUSA, and multiple specialized APIs. Additionally, Chunxiao can work with PyTorch, TensorFlow, PaddlePaddle, Jittor, and other mainstream deep learning frameworks and popular AI models.
Moore Threads works with Unreal Engine and Unity developers as well as popular titles (e.g. call of duty, Cross-fire, counter strike, diablo 3, league of legendsSuch). On the other hand, Moore Threads admits that Chunxiao only properly supports about 20 DirectX titles at the moment and doesn’t make any performance promises.
The GPU developer plans to start selling MTT S80 graphics cards on JD.com on November 11, 2022. Pricing has not been announced.