AMD’s Strategic Approach to AI: Differentiation, Generative AI, and Embracing Open Source

CXOToday has engaged in an exclusive interview with Mr. Mahesh Balasubramanian, Director, AMD Datacenter GPU Product Marketing


Q1. Can you shed some light on AMD’s AI approach and how it sets your brand apart in the rapidly evolving tech landscape?

The AMD focus is multi-faceted. With a leading goal of building energy efficient and sustainable solutions, we offer our customers and partners a broad portfolio of leading AI hardware products from the data center to the edge, which no other technology company can provide. This positions AMD well to help power this broad industry transformation to AI for whatever applications are needed. We continue to invest heavily in breakthrough innovations such as advanced packaging and 3D stacking, chiplet architectures and innovative products such as our latest compute focused AMD Instinct MI300X GPUs and MI300A APUs. We are scaling innovation and architecture from cloud to edge to end points with our Zen, CDNA, RDNA, and XDNA architectures. We are also leveraging our AI IP across our portfolio of products.


Q2. AMD has a track record of challenging established leaders like Intel through unique problem-solving approaches and successful execution, so how does AMD plan to navigate the field of generative AI?

Generative AI, and LLMs are game-changing. We are going to see an explosion in productivity across industries over the next three to four years due to generative AI models. We believe that AI requires not one, but multiple engines and GPUs are important for the kind of Generative AI workloads that are happening at hyperscalers. We have been developing and investing in R&D across GPUs, CPUs, and other accelerators for several years now, to meet the demands of our customers, today and into the future. We are also working with key market makers with a focus on first-to-market opportunities such as Ryzen AI or AI HPC workloads. Our approach has no “on-demand” pricing and doesn’t throttle the core at the expense of other applications. We also have the broadest market reach today from devices and endpoints to the edge and cloud. We have formed a new centralized group that is bringing all our AI assets under one roof, resulting in an agile roll-out of new products and support for the latest AI models. The Instinct MI300X accelerator chip announced at our Data Center and AI Tech premiere is our new power-packed GPU to handle the increasing demands of generative AI large language model workloads.


Q3. What is AMD’s specific strategy for generative AI, considering the importance of addressing aspects beyond duplicating NVIDIA’s approach?

We believe that AI requires not one, but multiple engines and GPUs are important for the kind of Generative AI workloads that are happening at hyperscalers. Our differentiation is our ability to provide AI solutions from cloud to edge to endpoints that only AMD can provide. This makes us nimble and resilient in this fast-changing world of AI.  Today, our Instinct GPU accelerators offer great performance for large inference use cases with leadership in memory density and bandwidth, which are necessary for LLMs, providing more inference jobs per GPU. Our Instinct MI250 GPU is showing higher performance on LLMs than Nvidia’s A100 because of higher memory capacity and bandwidth. Compute is no longer the bottleneck. As these LLM models mature and become simpler, AMD has our other solutions from RDNA, XDNA, and x86 solutions to right size for the AI workloads at these hyperscalers.  Also, innovation in AI needs to be unrestricted through open software, frameworks and models. We are partnering with industry leaders in open standards like PyTorch, TensorFlow, ONNX, JAX, Huggingface to tightly integrate our open software solutions with theirs that offer AI developers a choice of hardware that can help build their next great AI application like ChatGPT.


Q4. With NVIDIA’s CUDA and Intel’s OneAPI dominating the AI programming tools market, how does AMD’s ROCm compete and do you believe there is room in the market for another solution?

Software is an integral part of AMD’s AI strategy. We champion the leveraging of open standards and the open-source community. We are a strong advocate for this open approach. Our software strategy focuses on ‘open-source vs proprietary’, which is critical to what we are seeing with this fast-moving, dynamic move to AI environments. Finally, as the industry moves toward an open software ecosystem, we are working with the biggest industry leaders such as Hugging Face and the PyTorch Foundation.

During our latest “Data Center and AI Technology Premiere,” we showcased the ROCm™ software ecosystem for data center accelerators, highlighting the readiness and collaborations with industry leaders to bring together an open AI software ecosystem.

PyTorch is a de facto framework for open-source software. As a founding member of  PyTorch2.0, we are working with the foundation to enable accessibility of AMD hardware to the broad AI community for AI development and enable efficiency with optimization on our hardware and software platforms. Both teams are working together to fully upstream the ROCm software stack, providing immediate “day zero” support for PyTorch 2.0 with ROCm release 5.4.2 on all AMD Instinct accelerators. This integration empowers developers with an extensive array of AI models powered by PyTorch that are compatible and ready to use ‘out of the box’ on AMD accelerators.

Leave a Response