Samsung has reportedly started HBM4 development with the company rumored to offer customized HBM4 memory solutions to tech giants Meta and Microsoft.
The company has started developing "custom HBM4" memory that will be "customized" for Microsoft and Meta according to Korean outlet MK. According to a semiconductor company official, Microsoft "has its own artificial intelligence (AI) chips called Maia 100 and Meta Artemis". Samsung Electronics is the best partner for these big tech companies because it has a memory division and an LSI division that can design computational chips directly".
Most of the big tech companies have their own AI data centers and AI supercomputers, so there is a huge need for the reduction of chips purchased (which saves lots of money). This is why these companies design and use AI accelerator chips separately, while buying AI chips made by the likes of NVIDIA and AMD.
Samsung is expected to begin mass production immediately after completing the development of HBM4, which is expected by the end of 2025. The company hasn't been liberal with specs just yet, but Samsung did unveil its HBM4 specifications through the semiconductor conference "ISSCC 2024" back in February 2024.
- Read more: Samsung's next-gen HBM4 to enter mass production by the end of 2025
- Read more: Samsung to manufacture logic dies for HBM4 AI memory using 4nm node
- Read more: TSMC + Samsung co-developing bufferless HBM4 memory chip, its first partnership in AI
The new HBM4 memory from Samsung is expected to have 2TB/sec of memory bandwidth, which is an increase of 66% over HBM3E, and up to 48GB capacities with HBM4, up 33% from the 36GB with HBM3E, by stacking 16 DRAM stages.
NVIDIA's next-gen Rubin R100 AI GPU will be debuting with next-gen HBM4 memory, with TSMC and SK hynix partnering for the new HBM4 memory development. Samsung is behind, but with customized HBM4 solutions offered to companies like Meta and Microsoft, things could get interesting (but also remember, Samsung has been stumbling over itself for a while now, especially in the semiconductor business).