AMD has released the MI300 series in today's hot field of artificial intelligence, which is expected to be released in the fourth quarter, challenging NVIDIA's exclusive position.
The MI300 series is AMD's next-generation accelerator family for the data center and artificial intelligence markets, featuring advanced 3D packaging technology and CDNA3 architecture. The MI300 series includes two models: MI300A and MI300X.
The MI300A is a data center APU (Accelerated Processor) with integrated CPU and GPU with 24 Zen 4 CPU cores and CDNA3 GPU cores, and 128GB of HBM3 memory. The MI300A is designed to provide efficient hybrid computing power to support a variety of AI and HPC use cases. The MI300A enables zero-copy data transfer between CPU and GPU, reducing latency and power consumption, and simplifying the programming model. The MI300A can also leverage AMD Infinity Fabric link technology to enable high-speed interconnection between multiple MI300As.
The MI300X is a generative AI-focused GPU accelerator with CDNA3 GPU cores and 192GB of HBM3 memory. The MI300X is designed to provide powerful graphics processing capabilities to support the generation, editing, and enhancement of multimedia content such as images, video, and audio. The MI300X can leverage AMD Infinity Fabric link technology to enable high-speed interconnection between multiple MI300Xs.
AMD claims that the MI300 series delivers 8x better AI performance than the MI250 series and 5x better performance/watt (using FP8 with sparsity computing). AMD also said that the MI300 series products can reduce the cost and complexity of data centers, and improve operational efficiency and reliability.
The MI300 series will be available in the fourth quarter, starting with the National Nuclear Security Administration (NNSA) supercomputer, El Capitan. El Capitan is expected to be operational by the end of 2023 as the world's fastest supercomputer, with peak performance of 2 exaflops (two quadrillion floating point operations per second), twice as much as the current fastest supercomputer, Frontier#Artificial Intelligence # #芯片#
