What simply occurred? At its Q2 2024 earnings name, Micron Technology CEO Sanjay Mehrotra introduced that the corporate has offered out its total provide of high-bandwidth HBM3E reminiscence for all of 2024. Mehrotra additionally claimed that “the overwhelming majority” of the corporate’s 2025 output has additionally already been allotted.

Micron is reaping the advantages of being the primary out of the gate with HBM3E reminiscence (HBM3 Gen 2 in Micron-speak), with a lot of it getting used up by Nvidia’s AI accelerators. According to the corporate, its new reminiscence expertise can be extensively utilized in Nvidia’s new H200 GPU for synthetic intelligence and high-performance computing (HPC). Mehrotra believes that HBM will assist Micron generate “a number of hundred million {dollars}” in income this 12 months alone.

According to analysts, Nvidia’s A100 and H100 chips have been broadly utilized by tech firms for his or her AI coaching fashions final 12 months, and the H200 is anticipated to comply with of their footsteps by changing into the most well-liked GPU for AI functions in 2024. It is anticipated for use extensively by tech giants Meta and Microsoft, which have already deployed lots of of hundreds of AI accelerators from Nvidia.

As famous by Tom’s Hardware, Micron’s first HBM3E merchandise are 24GB 8-Hi stacks with 1024-bit interface, 9.2 GT/s information switch fee and a 1.2 TB/s peak bandwidth. The H200 will use six of those stacks in a single module to supply 141GB of high-bandwidth reminiscence.

In addition to the 8-Hi stacks, Micron has already began sampling its 12-Hi HBM3E cubes which are mentioned to extend DRAM capability per dice by 50 p.c to 36GB. This enhance would permit OEMs like Nvidia to pack extra reminiscence per GPU, making them extra highly effective AI-training and HPC instruments. Micron says it expects to ramp up the manufacturing of its 12-layer HBM3E stacks all through 2025, that means the 8-layer design can be its bread and butter this 12 months. Micron additionally mentioned that it’s bringing elevated capability and efficiency with HBM4 within the coming years.

In addition, the corporate is engaged on its high-capacity server DIMM merchandise. It not too long ago accomplished validation of the business’s first mono-die-based 128GB server DRAM module that is mentioned to offer 20 p.c higher power effectivity and over 15 p.c improved latency efficiency “in comparison with opponents’ 3D TSV-based options.”



Source link