AMD Challenges NVIDIA with Ambitious Rack-Scale AI Accelerators for 2026

AMD preps rack-scale Instinct MI450X IF128 with 128 GPUs to challenge Nvidia's VR200 NVL144 in 2026

AMD's venture carries significant complexity and deployment risks. AMD plans to launch its first two rack-scale Instinct accelerators in 2026 to compete with Nvidia's VR300 NVL144, reports SemiAnalysis. They are named AMD's Instinct MI450X IF64 and Instinct MI450X IF128, and both are designed for AI deployment. If they prove to be a success, it could change the landscape of AI hardware over time. While AMD's Instinct MI300-series AI and HPC GPUs are very powerful on paper, they cannot compete against Nvidia's GB200 NVL72 rack-scale solution in terms of performance scalability, as their maximum scale-up world size is eight processors. But next year, things are going to change as AMD plans to release its Instinct MI450X IF64, and Instinct MI450X IF128 solutions, replete with 64 and 128 GPU packages, to compete against Nvidia's VR200 NVL144 (with 72 GPU packages). Theoretically, AMD's MI450X IF128 can have an edge over Nvidia's VR200 NVL144. However, its complexity and technical challenges may limit its initial success. The Instinct MI450X IF128 will be AMD's first system to support multiple AI processors across two racks using Infinity Fabric, extended over Ethernet. The machine will rely on 16 1U servers running one AMD EPYC 'Venice' CPU with four Instinct MI450X GPUs equipped with their own LPDDR memory pool and a PCIe x4 SSD. Each of the 128 GPUs will have over 1.8 TB/s of unidirectional internal bandwidth for inter-GPU communication within the same scaling domain, thus enabling significantly larger compute clusters than AMD has supported so far. For scale-out communication outside the local group of GPUs (i.e., MI450X IF128 machines), the system will include up to three 800GbE Pensando network cards for each GPU. This provides a total outbound network bandwidth of 2.4 Tb/s per device (via PCIe). A secondary configuration will also be available, allowing each GPU to use two 800GbE network cards connected using a PCIe interface. However, this version will not be able to use the full bandwidth of the interfaces, as the PCIe 5.0 links are insufficient to fully support two high-speed network cards. Unlike Nvidia's GB200-series systems, which use active optical cables with embedded components to connect racks, AMD will employ a simpler passive copper wiring approach. This strategy may help reduce system cost and power consumption, but could be limited by signal integrity or cable length constraints. Also, due to the system's complexity, manufacturing and deployment may face delays or technical issues. To address this risk, AMD is preparing a smaller version of the same architecture called MI450X IF64. This variant will be confined to a single rack and use a simplified interconnect design, which promises to enable a more predictable rollout. If AMD manages to execute this architecture successfully, it could improve its position in the AI compute market, particularly AI inference systems. Whether it will be able to challenge Nvidia is something that remains to be seen, though.

AMD Set To Unveil Its First-Ever Rack-Scale Architecture With Instinct MI400 Lineup; Could Potentially Tip The Balances Away From NVIDIA

AMD seems to plan to aggressively enter the AI market through its Instinct MI450 AI clusters, as it would be the firm's first rack-scale product. When compared in terms of scalability, NVIDIA leads the frontier with its AI clusters, notably the GB200 Blackwell configurations, which have seen massive adoption by the market. Not just because of their performance, but there isn't a rack-scale solution in the industry that is comparable to NVIDIA's GB200/GB300 lineup, but that looks to change in the future, as according to SemiAnalysis, AMD plans to introduce its first rack, the MI450 IF128 cluster, by H2 2026, and it is said to overthrow Team Green's dominance in the segment, giving Team Red a chance in the AI race. It is claimed that AMD plans to drop 128-GPU and 64-GPU solutions as its first rack-scale products, and they are set to directly rival NVIDIA's "Vera Rubin" VR200 NVL144 architecture, well, at least on paper. One of the momentous technologies Team Red is expected to adopt with its rack-scale solution is the use of "Infinity Fabric" over Ethernet as the interconnect mechanism, which will allow the firm to deliver over 1.8TByte/s unidirectional bandwidth per GPU. Regarding scale-out communications, AMD is set to include 3x Pensando 800GbE network cards per GPU with the MI450 rack-scale solution, allowing them to provide a network bandwidth of 2.4 Tbit/s, 1.5x higher than NVIDIA's VR200 NVL144. There will also be a secondary configuration available for scale-out communication, which will involve the use of 2x custom Ethernet network cards per GPU, but with a PCIe interface. While on paper, AMD plans to take a lead with rack-scale solutions, there are some downsides, which we'll cover ahead. The MI450 IF128 cluster is said to be one of the most complex designs, which will make high-volume production a big concern. Moreover, being AMD's first rack-scale venture, the firm would be met with competition that has already managed to dominate the segment, so Team Red needs to come up with a more solid option. For that, AMD plans to focus more on the "smaller" MI450 IF64 rack-scale solution, which will feature a simpler yet effective design. It is clear that AMD wants to scale up its position in the AI industry with the MI400 lineup, and it could very easily turn out to be a tipping point for the firm if it manages to deliver on performance and efficiency.