AMD’s Next-Gen MI400 AI GPU Spotted in Linux Kernel Patches
▼ Summary
– AMD has begun upstreaming Linux kernel patches for its next-generation GPU IP blocks, including GFXHUB 12.1, which may indicate enablement for the CDNA5 architecture.
– The GFX 12.1 IP could be used for next-generation Instinct MI430 and MI450-series accelerators for HPC and AI workloads, or possibly for integrated GPUs.
– Leaks suggest GFX1250 monikers are reserved for CDNA5 accelerators like the MI450-series, while GFX1300 monikers are for RDNA5 GPUs, though AMD has not confirmed this.
– The enablement of the Instinct MI400-series highlights AMD’s steady execution of its CDNA roadmap and suggests a refresh of its accelerator family next year.
– A key question is whether AMD can launch and ramp up production of its MI450 and Helios solution before Nvidia’s Vera Rubin platform in Q3 2026.
Recent updates to the Linux kernel reveal that AMD is laying the groundwork for its next-generation GPU architecture. The appearance of new IP blocks, specifically GFXHUB 12.1, strongly suggests that development is underway for upcoming high-performance computing and artificial intelligence accelerators. While the exact product family remains unconfirmed, this move signals AMD’s continued push to compete in the rapidly growing AI hardware market.
Among the hardware enablement patches submitted by AMD are several key IP blocks: PSP 15.0.8 for platform security, IH 7.1 as the interrupt handler, MMHUB 4.2 for graphics IP, GFXHUB 12.1 for GPU IP, and GMC 12.1 for the graphics and memory controller. These components are essential for integrating new silicon into the Linux ecosystem, a step that typically occurs well ahead of a product’s official launch.
Although AMD has not formally linked GFX 12.1 to any specific product, its latest Radeon RX 9000-series graphics cards, based on the RDNA 4 architecture, already use the GFX 12 IP. The newer GFX 12.1 version could represent an enhanced subset for integrated GPUs, or it may belong to next-generation discrete offerings built on either RDNA 5 or CDNA 5 architectures. Earlier leaks indicate that GFX1250 identifiers are reserved for CDNA 5-based accelerators, while GFX1300 appears intended for RDNA 5 GPUs.
One possibility is that GFX 12.1 refers to integrated RDNA 4 GPUs with meaningful architectural differences from current designs. However, industry observers consider this less likely, since all Zen 5-based APUs are expected to feature RDNA 3.5 graphics. A more plausible interpretation is that GFX 12.1 corresponds to AMD’s forthcoming Instinct MI400-series accelerators, which will be based on the CDNA 5 architecture.
According to hardware leaker HXL, the GFX1250 index is tied to AI-focused Instinct MI450-series GPUs, while GFX1251 refers to the HPC-oriented Instinct MI430-series. AMD has not verified these claims, but the pattern aligns with the company’s established naming conventions.
The initiation of enablement for the Instinct MI400 family underscores AMD’s steady execution of its CDNA roadmap. It also indirectly confirms that the company remains on schedule to introduce a new generation of accelerators tailored for AI and high-performance computing workloads in the coming year.
A critical question now is when AMD plans to officially launch its next-generation AI offering, the Instinct MI450, and whether it can scale production in time to compete effectively. Of particular interest is the company’s ability to ramp manufacturing for its Helios rack-scale solution, which pairs EPYC Venice CPUs with Instinct MI450 GPUs, ahead of NVIDIA’s anticipated Vera Rubin platform launch in the third quarter of 2026.
(Source: Tom’s Hardware)