Samsung has constructed the world’s first massive scale computing system utilizing GPUs with inbuilt processing-in-memory (PIM) chips. These reminiscence modules, which have been loaded onto 96 AMD Intuition MI100 GPUs, elevated AI coaching efficiency by 2.5x, in line with a report by Enterprise Korea.
Samsung initially demonstrated the PIM-modified GPUs in October, however solely lately mixed 96 PIM-modified GPUs in a cluster. In comparison with regular video reminiscence, these modified MI100 chips not solely carried out 2.5x higher, in addition they reduce energy consumption by 2.67x, drastically growing the GPUs’ effectivity at working AI algorithms.
PIM is a brand new technology of pc reminiscence that may pace up computationally advanced workflows dealt with by processors reminiscent of CPUs and GPUs. Because the title suggests, every reminiscence module is able to processing knowledge by itself, lowering the quantity of information wanted to journey between the reminiscence and the processor.
Samsung has been growing PIM for a while now. The corporate demoed a number of implementations in 2021, involving a number of totally different reminiscence sorts together with DDR4, LPDDR5X, GDDR6, and HBM2. In LPDDR5 type, Samsung noticed a 1.8x enhance in efficiency with a 42.6% discount in energy consumption and a 70% discount in latency on a take a look at program involving a Meta AI workload. Much more spectacular, these outcomes have been from a typical server system with no modifications to the motherboard or CPU (all that modified was a swap to PIM-enabled LPDDR5 DIMMs).
Samsung is not the one firm growing PIM chips — SK hynix launched its personal PIM modules earlier this yr. In keeping with SK hynix’s preliminary testing, its GDDR6-AiM (Accelerator in Reminiscence) software sped up AI processing by 16x and decreased energy consumption by 80%. That is loads faster than Samsung’s modified MI100s, however we do not know what SK hynix used for testing so it is not a direct comparability.
Regardless, PIM appears like a potent answer to rushing up AI-accelerated workflows. “As the pinnacle of the AI analysis middle, I wish to make Samsung a semiconductor firm that makes use of AI higher than some other firm,” Choi Chang-kyu, vp and head of the AI Analysis Middle at Samsung Electronics Superior Institute of Expertise, informed Enterprise Korea.