Samsung today announced that it has developed a new type of HBM2 memory equipped with integrated AI processor, able to perform operations normally associated with processors, graphics chips, ASICs or FPGAs: these new chips are called HBM-PIM (processing-in-memory) and are the first in the world of this kind.
The most trivial way to explain what it is to those who do not have a very good relationship with technology is the following: take a processing unit like the ones mentioned above and add it to the RAM memory in the same chip, so as to carry out multiple operations with the available data. This is HBM-PIM.
But let’s get to the more technical aspects, so as to analyze in detail the importance of this product. The AI system that comes with these memory banks has been specially designed to lighten the load of moving analyzed data between the processing units and the memory, thus ensuring a processing power of 1.2 TFLOPS per chip and reducing overall system energy consumption by 71%, precisely as there is no need to move data from HBM-PIM to other drives. Additionally, using PIM for machine learning increases system performance by double.
Although this chip remains, as said by the vice president of the Memory Product Planning division of Samsung Electronics, Kwangil Park, “The industry’s first programmable PIM solution tailored for different AI-based workloads “, there are also cons as well as pros: in particular, HBM-PIM occupies the space normally taken up by memory banks, thus reducing the total capacity. Furthermore, it will take time before we see this product arrive on the market although the time-to-market remains particularly low, also because Samsung intends to focus more on the data center and HPC systems sector.
The first Exynos SoC with AMD GPU and the first range of smartphones with a rollable or expandable screen appear among the other great products expected to be launched on the market.