Taking place in Taiwan this week, Computex 2024 is showcasing a variety of hardware advancements as industry leaders gather. Nvidia, for instance, unveiled its new Blackwell systems and made its Spectrum-X Ethernet stack generally available for AI applications.

AMD’s standout announcement was the launch of the Instinct MI325X GPU. This new model boasts a faster processor and a substantial increase in high-bandwidth memory (HBM) compared to the MI300X and Nvidia’s Blackwell. The MI325X is equipped with 288 GB of HBM3e and delivers 6.0 TBps of memory bandwidth, while the MI300X offers 192 GB of HBM3 with 5.3 TBps bandwidth.

AMD pointed out that the highest benchmark for Nvidia’s Hopper H200 accelerator indicates a memory capacity of 141 GB HBM3e and a bandwidth of 4.8 TBps. In contrast, Blackwell, which launches this fall, offers a B200 model with up to 384 GB of HBM and 16 TBps bandwidth.

The MI325X, however, showcases impressive performance, achieving peak theoretical throughput of 2.6 petaflops for 8-bit floating point (FP8) and 1.3 petaflops for 16-bit floating point (FP16), surpassing the H200’s capabilities by 30%, according to AMD.

Additionally, the MI325X is capable of handling a complete 1-trillion-parameter model, which is twice the capacity of the H200, as reported by AMD.

Utilizing the AMD CDNA 3 architecture—similar to the MI300X—the MI325X is designed specifically for data center environments, focusing on generative AI and high-performance computing.

Looking ahead, AMD plans to release the new CDNA 4 architecture in 2025, alongside the Instinct MI350 series, promising a potential 35-fold increase in AI inference performance over the MI300 series. In 2026, the MI400 series will be introduced, built on the upcoming CDNA “Next” architecture.

Brad McCredie, AMD’s corporate VP of data center accelerated compute, emphasized the company’s commitment to innovation: “With our updated annual product cadence, we are dedicated to providing the performance that the AI sector and our customers expect for the future of data center AI training and inference.”

The MI350X will retain the 288 GB of HBM3e memory like the MI325X, and it will be produced using a 3-nanometer process, a significant advancement from the 6nm technology used for MI300 chips. This new chip will also introduce support for FP4 and FP6 floating point formats, with FP6 being unique to AMD, unlike competitors who focus on FP4 and FP8.

While AMD’s Instinct series may trail behind Nvidia’s Hopper line, it is steadily gaining traction with key partnerships. AMD highlighted growing interest from companies like Microsoft for Azure OpenAI services, Dell Technologies for enterprise AI applications, Supermicro with various AMD solutions, Lenovo incorporating Instinct in its ThinkSystem, and HPE utilizing them in HPE Cray servers.