At the annual Computex technology conference, GPU manufacturers showcased their latest technological achievements. Among them, NVIDIA andAMDThe performance is particularly outstanding.
Nvidia showcased the mass-produced Blackwell chip at the conference and announced its product roadmap for the next three years, including the next-generation Rubin AI platform. AMD CEO Lisa Su showcased AMD's CPU, GPU products and future roadmap, including the desktop Ryzen 9000 series CPUs, AI PC chips, data center chips and GPUs based on the new Zen 5 architecture.
In her speech, Su Zifeng particularly emphasized the Ryzen CPUs based on the Zen5 architecture, which are optimized with neural network processors to accelerate AI workloads. The Zen5 architecture has a wider CPU engine instruction window, capable of running more instructions in parallel, thereby achieving leading computing throughput and efficiency. Compared with the previous generation Zen4, Zen5 has doubled the instruction bandwidth, doubled the data bandwidth between the cache and floating point unit, doubled the AI performance, and supports full AVX512 throughput.
In the product display session, AMD showed the Ryzen 99950X processor with Zen5 architecture for the first time. This processor has 16 cores and 32 threads, with an acceleration frequency of up to 5.67GHz, a TDP of 170W, and a total of 80MB of L2 and L3 cache. The internal game test data provided by AMD shows that the performance improvement of Ryzen 99950X varies in different games. For example, it is 4% higher than Intel Core i9-14900K in Borderlands 3, 13% higher in Cyberpunk 2077, and 23% higher in Horizon Zero Dawn.
Su Zifeng also demonstrated AMD's next-generation AI PC chip, the Ryzen AI300 series APU (third generation). This series of chips is designed to provide the next-generation AI PC experience, requiring the NPU, CPU, and GPU to achieve optimal performance. The Ryzen AI300 series uses the XDNA AI NPU, which is claimed to be the strongest NPU on the mobile side, with a computing power of up to 50TOPS. It not only meets Microsoft's standard for the new generation of AI PCs "NPU computing power of more than 40TOPS", but also exceeds Qualcomm Snapdragon X Elite's 45TOPS and Intel Lunar Lake's 40-45TOPS.
Finally, Su Zifeng announced AMD's Instinct GPU roadmap for 2024-2026, including the MI325X launched this year, the MI350 in 2025, and the MI400 in 2026. The MI325X will have a larger computing power scale and use HBM3E memory. The basic performance specifications show that the memory of the MI325X has increased by 2 times, up to 288GB; the bandwidth has increased by 30% to 6TB/s. This marks AMD'sAI ChipsThe determination and strength to challenge NVIDIA in the field.