About Lisa Su
Lisa Su, chair and CEO of AMD, delivered the commencement address at MIT on May 28, 2026, where she reflected on her own time as a student and shared career advice. She told graduates that "the best people find ways to make their luck," describing luck as "taking the risk to work on something really hard" and "choosing problems where you may not know the answer." Su also stated that "technology itself does not decide what the future looks like," adding that "the best people do" and that AI "can't decide which problems are worth solving."
In recent earnings calls and media appearances, Su has described the AI market as being in its early stages, saying "we are only in the third inning" of a nine-inning game. She cited growing demand from agentic AI as a driver for increased CPU compute requirements, and said AMD now expects the server CPU total addressable market to grow at over 35% annually, reaching over $120 billion by 2030. During a visit to Taiwan in May 2026, Su announced a $10 billion investment in the local ecosystem, including advanced packaging and substrates, and confirmed that AMD is the first high-performance computing customer for TSMC's 2nm process technology. She also stated that the company is "ramping up capacity like crazy" across silicon, packaging, and rack-scale systems.
Source: AI-verified profile updated from Lisa Su's recent appearances.
Browse all interviews →
✨ AI-enhanced transcript with speaker attribution
L
Lisa Su0:00
CDNA 3 is our brand new architecture that uses a new compute engine, the latest data formats 5 and 6, 5nm process technology, and the most advanced chiplet packaging technologies. At CES earlier this year we previewed MI 300A. It's the world's first data center APU, and what we have is our CDNA 3 GPU architecture with 24 high-performance Zen 4 CPU cores. Now these are the exact same cores in our leadership Genoa processors that I talked about earlier in the show, and we also add with that 128 GB of HBM3 memory, all in a single package. And what we have is unified memory across the CPU and GPU, which is frankly very effective, particularly for some HPC workloads. Now this results in eight times more performance and five times better efficiency compared to the MI 250X accelerator that is in the largest supercomputers today. Now MI 300A has also been designed into supercomputers already, and it's slated for the two-plus exaflop El Capitan system at Lawrence Livermore National Labs. It's the most complex chip we've ever built, with more than 146 billion transistors across 13 chiplets. Now you guys know we've led the industry with the use of chiplets in our products, and our use of chiplets in this product is actually very, very strategic. We created a family of products, so in addition to the MI 300A product with our chiplet construction, we can actually replace the three Zen 4 CPU chiplets with two additional CDNA 3 chiplets to create a GPU-only version of MI 300 optimized for large language models in AI. We call this MI 300X.
Now for MI 300X, to address the larger memory requirements of large language models, we actually added an additional 64 GB of HBM3 memory. So with that, I am super excited to show you for the very first time MI 300X. Now for those of you who are paying attention, you might see it looks very, very similar to MI 300A. Basically we took three chiplets off and put two chiplets on, and we stacked more HBM3 memory. But what you see with MI 300X is we truly designed this product for generative AI. It combines CDNA 3 with an industry-leading 192 GB of HBM3 that delivers 5.2 terabytes per second of memory bandwidth, and it has 153 billion transistors across 12 5nm and 6nm chiplets. So I love this chip, by the way. We love this chip. So look, when you look at the world, there are many things that you need. You need great compute engines, but you also need a lot of memory for everything that's going on. So when you compare MI 300X to the competition, MI 300X offers 2.4 times more memory and 1.6 times more memory bandwidth. And with all of that additional memory capacity, we actually have an advantage for large language models because we can run larger models directly in memory. And what that does is for the largest models, it actually reduces the number of GPUs you need significantly, speeding up the performance especially for inference, as well as reducing total cost of ownership. So of course, I want to show you the chip in action for the very first time. So let's watch MI 300 in action, shall we? For this demo, we wanted to show you a large language model running real-time inference on a single GPU. And since we had Hugging Face here with us, we're actually going to run the recently released Falcon 40B foundational large language model, which is currently the most popular model on Hugging Face right now, featuring 40 billion parameters. So let's watch for the first time ever, MI 300X.
