It is claimed that NVIDIA’s new generation Rubin and Rubin Ultra platforms face some difficulties in terms of design and technical features. It is said that this situation may create a significant opportunity for the AMD MI500 series, which is expected to be released in the second half of 2027.

The platforms in question aim to offer artificial intelligence performance beyond the Blackwell architecture. However, recent reports show that NVIDIA is struggling with five different critical obstacles in the production process.

Technical difficulties with Rubin platforms

The 288 GB HBM4 memory capacity and 22 TB/s total bandwidth planned for NVIDIA’s Rubin platform are being hampered by the poor quality of the base dies from Micron and SK Hynix. Difficulties in running these memory units at high speeds bring about design changes or possible delays in the production schedule.

On the Rubin Ultra side, 1 TB HBM4E memory capacity was initially targeted using 16-Hi stacks. It is stated that due to efficiency problems, this design was reduced to 12-Hi stacks and the total capacity was reduced to 768 GB.

It is also claimed that the Rubin Ultra platform has been reduced from a four-die structure to a two-die structure. The main reason for this change is shown to be serious yield and warping problems that occur in multi-chip packaging processes.

Despite this design change, NVIDIA is expected to maintain the performance levels promised in the initial announcement. To achieve this goal, the company is pursuing a new board-level assembly strategy that bundles Rubin Ultra GPUs in a 2+2 configuration.

Heat spreader change and competition with AMD

The heat spreader design planned for Rubin GPUs also stands out as another factor affecting the production process. It is stated that since the double heat spreader layout could not meet the bending requirements, a switch was made to the single heat spreader design.

Due to these changes, the production schedule for Rubin Ultra has been updated and mass production is planned to start in September. In standard Rubin GPUs, a switch is made to traditional graphite material due to instabilities in the existing indium-graphite thermal interface material.

NVIDIA’s Rubin Ultra platform and AMD’s MI500 series are direct competitors in the artificial intelligence market. While both companies focus on silicon photonics technology, competition is expected to intensify with AMD’s MI500 platform, which will be released in the second half of 2027.

NVIDIA had successfully managed similar design adjustments it had experienced in the Blackwell series in the past, reaching mass production on time. The company is expected to overcome these new challenges in a short time and stick to its road map.

Do you think NVIDIA will be able to continue to maintain its leadership in the artificial intelligence market with these design changes?