You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!
In the memory race does Micron play a leading role or will Samsung with better finance rule? What types of memory will be key in the AI/ML race or is it even a key factor? Will advances in memory keep up with the advances in processing? Also will automata where memory and processing are mixed ever come into play?
The ideal AI memory is DRAM density, with SRAM energy per bit and latency. Especially for edge inferencing where batch size wants to be 1 and keep latency and power low. HBM has a ways to go on energy, and is awful latency just like other DRAM.
Tight integration like Samsung Aquabolt did reduce energy and latency but the digital logic on a memory process was awful and the datapaths inflexible.
If DRAM had a foundry ecosystem with startups able to benefit from innovation the solution would already be out there. Waiting for the oligopoly to supply the innovation? Don't hold your breath.
The ideal AI memory is DRAM density, with SRAM energy per bit and latency. Especially for edge inferencing where batch size wants to be 1 and keep latency and power low. HBM has a ways to go on energy, and is awful latency just like other DRAM.
Tight integration like Samsung Aquabolt did reduce energy and latency but the digital logic on a memory process was awful and the datapaths inflexible.
If DRAM had a foundry ecosystem with startups able to benefit from innovation the solution would already be out there. Waiting for the oligopoly to supply the innovation? Don't hold your breath.
Thanks for the response, any other thoughts on advanced memory or how it will be used or integrated would be appreciated. What are your thoughts or observations on Micron working with memory?
