As part of Processing-in-Memory (PIM), i.e. the outsourcing of computing tasks from a processor or computing accelerator to memory, Samsung has also experimented with computing-capable main memory. So-called Acceleration DIMMs or AXDIMMs emerged.
In contrast to the HBM-PIM, for example, the processing units are not located in the DRAM components, but in the buffer chip (Processing-in-DIMM, PID) that server RDIMMs traditionally use to improve signal quality and consequently to increase memory capacity. Upcoming AXDIMMs largely correspond to the DDR5 standard, so they fit into normal RAM slots instead of RDIMMs, but are significantly higher than normal modules.
More performance, less energy consumption
The idea: AXDIMMs can perform simple arithmetic tasks on servers themselves in order to reduce data movements between RAM and processor and thus save energy. In addition, memory accesses should be easier to parallelize according to the principle, because the computing units can access all memory banks at the same time.
For prototypes, Samsung relied on DDR4 components in dual-rank design and a programmable logic chip from Xilinx’s Zynq UltraScale + FPGA family. In conjunction with older Xeon CPUs from Intel’s Broadwell SP series, the system performance increased by a factor of 1.8 in initial, unspecified tests, while the system-wide power consumption fell by almost 43 percent.
Suitable application scenarios would be, for example, AI calculations directly on the RAM module, for example with FP16 cores like the HBM-PIM, or the outsourcing of database software. In an initial cooperation According to the announcement, Samsung worked with SAP.
Samsung did not comment on a date for the series production of AXDIMMs at the Hot-Chips Conference 33. This year, with Intel’s Sapphire Rapids, the first server processors will appear that can handle DDR5 RAM and are therefore suitable for the new type of memory bar.