Connection to DriversCloud Create a DriversCloud.com account Reset your DriversCloud.com password Account migration
Instinct MI455X: AMD Brings Out the Heavy Artillery in the GPU World
High-performance GPUs to keep pace with NVIDIA’s lead in the world of AI—and, why not, even surpass it?
AMD’s annual Advancing AI event is, of course, an opportunity to discuss artificial intelligence, to highlight just how crucial this technology is today for tech companies, and, at the same time, to take stock of the technical advances made by AMD, which is competing as fiercely as possible against companies like Intel and NVIDIA. On the Intel front, we recently discussed the EPYC Venice server CPUs and their Zen 6 microarchitecture. AMD hasn’t rested on its laurels, and the GPU world has been treated to equally interesting announcements.
This time, the focus is on upgrading the Instinct GPUs, the latest version of which—the MI355X—had already allowed AMD to close the gap with NVIDIA. The Instinct MI455X is expected to perform even better, leveraging the U.S. company’s brand-new graphics architecture, CDNA 5. AMD has clearly set the tone by highlighting chips capable of delivering the power needed for large-scale artificial intelligence training and inference systems. To achieve this, AMD is once again relying on TSMC to manufacture its chips using the N2 process—2 nm—in order to accommodate the 320 billion transistors integrated by AMD. Note, however, that only the compute portions of the chip are manufactured using the N2 process. For the other components (cache, I/O, etc.), the company is using the N3P process, which is slightly less advanced.
But the cutting-edge technology isn’t limited to this monstrous graphics chip, which will be paired with a memory subsystem that’s perhaps even more mind-boggling: AMD engineers have paired the Instinct MI455X with a total of 432 GB of HBM 4 RAM. This delivers a nearly unbelievable bandwidth of 23.3 TB/s and a 192-channel data interface. More modest, the dual 192 MB L2 cache—divided into two 96 MB sections—isn’t half bad either. All in all, AMD doesn’t hesitate to highlight a peak performance of 20.13 PFLOPS in MXFP6 and MXFP8 before emphasizing the 40.26 PFLOPS in MXFP4. This raw performance should allow AMD to flex its muscles: naturally, the previous-generation Instinct MI355X is left far behind, but NVIDIA and its current Vera Rubin are also set to bite the dust… until NVIDIA’s next announcement?

