Chipmaker Advanced Micro Devices Inc. today threw down the gauntlet, announcing that it’s ready to take on Nvidia Corp. with a new generation of silicon chips that are geared toward artificial intelligence workloads.

During a special “Advancing AI” event today, the chipmaker announced the launch of multiple new chips and software services, including its highly anticipated AMD Instinct MI300X series data center AI accelerators (pictured). The MI300X was first revealed in June. The company said it will provide an alternative to Nvidia’s graphics processing units, which power the vast majority of AI workloads today.

Alongside the MI300X, the company revealed a new ROCm 6 open software stack, which comes with significant optimizations and functionality to support advanced large language models like OpenAI’s GPT-4 and Google LLC’s Gemini.

It also announced the availability of its new Ryzen 8040 Series chips for smaller devices such as personal computers, laptops and notebooks, complete with the Ryzen AI software for running large language models and building AI apps that can run locally on those machines.

“At AMD, when we think about it, we actually view AI as the single most transformational technology over the last 50 years,” AMD Chief Executive Lisa Su said in a keynote presentation today. “AI is absolutely the No. 1 priority at AMD. We are so well-positioned to power that end-to-end infrastructure that defines this new AI era.”

MI300X series chips

The AMD Instinct MI300X accelerators are targeted specifically at generative AI workloads, possessing industry-leading memory bandwidth for training and inference tasks. They’re being made available today alongside a second new chip, the Instinct MI300A accelerated processing unit (below), which combines the latest AMD CDNA 3 architecture with Zen 4 central processing units to deliver higher performance for AI and high-performance computing workloads.

The MI300A APUs are said to be the first data center APU that’s specifically designed for HPC and AI, leveraging 3D packaging and the 4th generation AMD Infinity architecture to deliver 1.9 times the performance per watt of the previous-generation AMD Instinct M1250X5. They’re focused on energy efficiency, which is of great importance for AI and HPC workloads as they can be extremely resource-intensive. According to AMD, the MI300A APUs benefit from integrated CPU and GPU cores within a single package to reduce drastically the energy requirements of these workloads, while maintaining the required high performance.

The APUs feature unified memory and cache resources giving customers an easily programmable GPU platform, the company said, along with “highly performant compute, fast AI training and impressive energy efficiency.”

Multiple launch partners

The chipmaker was joined on stage at the event by a host of partners that intend to offer the power of the MI300 series chips to their customers. For instance, Microsoft Corp. turned up to showcase how it’s deploying the MI300X accelerators within its new Azure ND MI300x V5 virtual machines, which are cloud-hosted servers optimized for AI workloads.

Oracle Corp. was also there to detail how it will offer bare-metal server solutions based on the MI300X accelerators, as well as a new generative AI service that’s powered by the same chip. Meanwhile, experts from the Lawrence Livermore National Laboratory dropped by to talk about how the MI300X chips will be deployed in El Capitan, a new supercomputer that’s currently under construction. It will deliver more than 2 exaflops of double-precision performance once it’s up and running, they said.

Not to be left out, Meta Platforms Inc. said it’s adding thousands of MI300X accelerators to its data centers, where they will be paired with the new ROCm 6 software to power AI inference workloads. The open-source ROCm 6 software stack is optimized to work with AMD’s chips and comes with support for new data types, advanced graph and kernel optimizations, optimized libraries and state-of-the-art attention algorithms. Using the software, Meta said it was able to deliver an eight-times performance increase for overall latency for text generation tasks based on the open-source Llama 2 model, compared with the older ROCm 5 software running on older GPUs.

Server makers including Dell Technologies Inc., Hewlett Packard Enterprise Co. and Lenovo Group Ltd. all joined AMD onstage to discuss their new data center server offerings powered by the MI300X chips. Dell will integrate the chips within its next generation PowerEdge XE9680 servers, while Lenovo said it also plans to bring the chips to its latest ThinkSystem platforms. Other server makers, including Supermicro Computer Inc., are also planning to launch new hardware featuring AMD’s accelerators.

Elsewhere, a number of specialized AI cloud providers, including Aligned AI, Arkon Energy, Denvr Dataworks Inc. and Tensorwaves Inc. all said they’re planning to offer new services that will further expand access to the MI300X GPUs.

Finally, OpenAI said it will also add support for the AMD Instinct accelerators and APUs to Triton 3.0, an open-source Python-like programming language that enables researchers with no experience with Nvidia’s CUDA software framework to write highly efficient GPU code for AI applications.

ADM Ryzen 8040 CPUs

Alongside the data center-focused GPUs, AMD also intends to become a major power in PC-based AI that runs directly on the local hardware. The new AMD Ryzen 8040 CPUs are said to deliver the most powerful on-CPU AI capabilities thus far, supporting 64% faster video editing and 37% faster 3D rendering than the previous generation of Ryzen chips.

“We believe AI should be everywhere,” Su said.

In addition, PC gamers can expect up to 77% faster gaming compared to alternative chips. The Ryzen 8040 CPUs are based on the Zen 4 processor architecture and come with up to eight cores that can deliver 16 threads of processing power. They’re designed to work with AMD’s Ryzen AI 1.0 software, which is a new offering that will enable developers to deploy applications built with pretrained AI models using frameworks such as PyTorch or TensorFlow and run them on select laptops.

AMD said the Ryzen 8040 CPUs are not just about advancing performance, but also extending battery life, which is made possible through some innovative new power management features. They also come with advanced LPDDR5 memory capabilities, which makes them suitable for immersive virtual reality experiences and high-performance gaming and streaming.

The company also offered PC makers a taste of things to come, unveiling its next-generation “Strix Point” CPUs, which will begin shipping in late 2024. They’re based on a XDNA 2 architecture that will provide a threefold increase in AI compute performance compared with the prior generation.

At the event, AMD welcomed Microsoft officials back on stage to discuss how they’re working to integrate AI experiences in Windows PCs based on the Ryzen 8040 CPUs. The company explained that the chips are ready to leverage the entire range of the Windows 11 ecosystem to deliver optimized performance, including full support for its most advanced security features. Select systems running the Ryzen 8040 processors will also be able to access out-of-the-box AI with Windows Studio Effects Pack, ensuring privacy at home or on the go with background blur, eye-gaze tracking and noise cancellation.

Photos: AMD