According to Samsung, SKhynix started to the integration of computing accelerators in memory chips. The logic of actions is simple – von Neumann’s computer architecture is hopelessly outdated for solving a number of problems. Placing computing accelerators directly on memory chips saves energy and multiplies the performance of certain operations – this is the future for many artificial intelligence applications.
Exactly one year ago, Samsung presented the first memory with PIM blocks (Processing In Memory). That honor went to the HBM3 memory chips. As you know, HBM memory is stacked, so the accelerator crystal did not increase the chip footprint and practically did not affect the density of the memory chip. Everything is a little different with SK hynix PIM storage, but it has its own undeniable advantages.
SK Hynix was the first to integrate an accelerator block into GDDR6 memory. By the way, she called this block “Accelerator in Memory” or Accelerator in Memory (AIM), which does not change the essence. Apparently the accelerator took some of the real estate of the crystal from the DRAM cell array – this is not stacked memory. But on the other hand, with the SK hynix solution, you can equip graphics cards and accelerators with GDDR6 AIM memory, which have proven themselves so well in AI and machine learning tasks. In addition, it brings PIM storage closer to computers and even smartphones, which will push the AI capabilities of mobile technology to new heights.
According to SK hynix, the use of GDDR6 AIM memory in a range of tasks speeds up calculations by up to 16 times while reducing consumption by up to 80%. The processing speed will increase to 16 Gb/s, which the company promises to talk about in detail at the ISSCC 2022 conference starting on February 20 in San Francisco. In addition, the GDDR6 AIM supply voltage has become even smaller and is 1.25 V instead of 1.35 V for conventional GDDR6 memory.
Since in-memory computing has not yet become an industry standard, SK hynix has partnered with Korea’s SAPEON to promote GDDR6 AIM memory solutions. SAPEON was recently spun off from SK Telecom Corporation as an independent developer of AI chips. Together they present a practical solution based on accelerated memory. It is believed that modern operating systems can easily support such memory, since the main load will fall on the controllers integrated in the memory, and not on the CPU or GPU.