The memory HBM3E and Samsung failed the quality test NVIDIA. According to Reuters, the South Korean giant’s modules did not meet the expectations of the GPU giant led by Jensen Huang. The main reasons for test failure would be the excessive heating of chips e high energy consumption.
Samsung denied the information and stated that it works closely with its partners to optimize new products. “Claims of failure due to heat and power consumption are not true, testing was proceeding smoothly and as planned.”Samsung said.
HBM, high performance for AI accelerators
Evolution of the famous high bandwidth memories (HBM), the HBM3E and HBM3 version interests NVIDIA in one of the key segments for the company, equipping its accelerators for artificial intelligence. The company’s new generation of products in this category are the GPUs Blackwell.
It is also considered that the problem with Samsung’s memory chips would not be in the issue of the development of the technology itself, but in the production quality standards, which would be below what is delivered by SK Hynix, which would have been used by NVIDIA as basis of comparison. Samsung was the first company to develop 12-stack HBM3E memories with 36 GB and a bandwidth of 1,280 GB/s.
NVIDIA has not publicly commented on possible problems with Samsung’s memories
Source: https://www.hardware.com.br/noticias/2024-05/hbm3e-da-samsung-reprovadas-em-teste-de-qualidade-nvidia.html
