Learn Your AI ABCs …… Analog Compute in Memory …. Mark Stansberry ….. The StatisticsMatrix
Contact: mark@statisticsmatrix.com
Analog Compute in Memory, also known as ACIM, differs from Digital Compute in Memory (DCIM) in that it does not use a digital two state memory element (like a SRAM or DRAM cell, latch or flop-flop). Instead, ACIM uses a multistate, continuous voltage memory, based on non-volatile memristor technology like Phase Change Memory (PCM). The primary difference between the two memory cells is that a PCM cell can store much more than the one-bit of data that a SRAM or flip-flop can, some say 256 bits (256 voltage or conductance level).ACIM can be designed to work with a digital or analog processor AI system.
Many are reporting progress on ACIM mixed systems which are systems that use ACIM tiles and digital processors, like the ARM microprocessor. For example, Systems Laboratory (ESL), IBM Research Europe, and 3HEIG-VD in their paper “ALPINE: Analog In-Memory Acceleration with Tight Processor Integration for Deep Learning” report significant power / performance gains with their ACIM based mixed AI system over Single Instruction Multiple Data (SIMD) all digital AI implementations. The paper indicated that performance power ratios were in the order of 20 times better for mixed signal ACIM based systems.
Although the paper does discuss accuracy and precision, the paper didn’t provide a definitive accuracy and precision comparison between ACIM and DCIM based system. The paper does hint that high accuracy and precision ACIM systems are obtainable that could could replace power-hungry DCIM systems. Their AMIC tile, according to the paper, contains a digital-to-analog and analog-to-digital converter, (DAC and ADC) dedicated registers, a localized controller, and a PCM memory crossbar array.
What appears to be going on inside the AMIC tile is that multiple digital bytes are multiplied together in the analog domain, (DAC multiplier accumulator technology) and then stored as analog voltages in the PCM memory cell. This eliminates the need to perform the power hungry step of a DCIM digital multiplication and accumulation. The multiplied accumulated product is available as a digital word once converted with an onboard ADC. Because there are multiple tiles matrix vector multiplication can be done quickly.
About the AI Top 100 Study
100 companies have been included in my study. This includes analog, mixed signal and digital AI solutions and platforms. For further information about the study or to be included in the study, contact me directly via my email.
#ai #flash #MRAM #memristor #semiconductor #CIM #IBM #analog #mixed #marketresearch #memory #GPU #digital #PCM #tile #ACIM #DAC #ADC #IPCores #SRAM #DRAM #technology #deeplearning #power
