Samsung brings AI performance to its HBM memory to deliver breakthrough performance and efficiency - HBM PIM
Even memory has compute now!
Published: 17th February 2021 | Source: Samsung |
Samsung brings AI performance to its HBM memory to deliver breakthrough performance and efficiency
Samsung's latest HBM memory promises to deliver tremendous performance and efficiency improvements by bringing the power of AI processing to its latest chips, enabling higher performance levels than today's HBM chips by non-traditional means.
It is claimed that these new HBM-PIM (High BandWidth Memory-Processing In Memory) chips do "not require any hardware or software changes", allowing these new modules to be integrated into existing products to deliver "over twice the system performance while reducing energy consumption by more than 70%" when compared to their HBM2 Aquabolt modules.
Right now, Samsung has not been clear about how their HBM-PIM technology works. Right now it looks like this technology many not require specific software/hardware support to enable these increased performance and efficiency targets, though things will be clarified on February 22nd as Samsung released a detailed paper on the technology at the International Solid-State Circuits Virtual Conference (ISSCC).
Even memory has compute - HBM-PIM
Samsung has claimed that its HBM-PIM technology enables parallel data processing and minimises data movement within its HBM memory modules. By using a "DRAM-optimised AI engine" Samsung can increase the effective performance of their HBM DRAM technology, enabling transformative performance increases.
HBM-PIM is designed to eliminate one of the largest bottlenecks for memory performance, integrating processing power into memory to eliminate unnecessary sequential data transfers and make the best possible use of their DRAM-to-processor interconnect.
Most of today’s computing systems are based on the von Neumann architecture, which uses separate processor and memory units to carry out millions of intricate data processing tasks. This sequential processing approach requires data to constantly move back and forth, resulting in a system-slowing bottleneck especially when handling ever-increasing volumes of data.
Right now, it remains to be seen how long it will take for HBM-PIM memory to be utilised by the industry, though you can bet that Samsung's promise if a 2x memory bandwidth increase will be of interested to a lot of potential customers. Samsung sees this new memory type as being useful for AI computing and other HPC applications, which makes sense given the high bandwidth requirements of datacenters.
We will need more information about this new HBM-PIM memory from Samsung before reaching any firm conclusions on it, but if it works as intended, Samsung will have created the fastest and most efficient HBM memory on the market.
You can join the discussion on Samsung's HBM-PIM memory on the OC3D Forums.