IBM will fight for AI with its 64 Core chip and analog memory

IBM has developed a chip with analog memory for AI which could be the alternative to the high-end GPUs that are being used to run most of the algorithms in this discipline that are so booming lately. It is a processor inside your HERMES project. What is unique about this invention compared to other processors oriented towards artificial intelligence?

IBM has updated its HERMES project, which consisted of a chip for artificial intelligence made up of 34 tiles connected to each other in a matrix that they presented back in 2021. In the new version of 2023 they have increased the number of units to 64 and have implemented the use of analog memory to store the data inside the chip used for the corresponding calculations. Therefore, its greatest particularity is that it does not work with digital or binary values ​​to carry out its calculations.

IBM introduces an AI chip with integrated analog memory

Today everyone who is dedicated to chip desEsports Extrashas their own processor for artificial intelligence, Although all of them follow a series of common rules, in reality it is that each one of them makes the difference in one way or another from the competition. However, in the case at hand, IBM has opted for what is called analog in-memory computing (AIMC). Each 1 of the 64 cores on the chip has a PCM (Phase Change Memory) infrastructure able to store a matrix of 256 by 256 pesoson which they can perform matrix multiplications.

All this means that the memory cells they do not store data in binary format, but analogfor this it is necessary to use digital-to-analog converters (DACs)which generate information taking into account the amplitude and time of each wave. It is not the first chip to use alternative memory to store data internally to be used for AI calculations, as solutions using resistive memory have already been seen, but this is the first to use non-digital memory.

At the moment it is an experimental chip, of which we do not know its size, but we sEsports Extrasthis due to the fact that it has been built under a node of 14 nanometers, in terms of its power per square millimeter, this is 0.4 TOPS. In any case, it must be clarified that IBM has been working with analog memory for AI for several years, so it is not the first system developed under this concept.

The goal is to replace GPUs for AI

One of the biggest problems facing many AI-related supercomputer builders is the need to have a large number of graphics cardswhich consume much more than a CPU, including server ones. If we start from the assumption that when setting up one of these systems, the total energy consumption is a specification, then we will see that it is a problem to be solved in the coming years.

The idea of ​​IBM, with its chip for artificial intelligence with analog memory, is that it be an alternative to using GPUs. However, we must start from the fact that for the moment it is still an investigation and it is not a final product that they or a third party can benefit from. We’ll see where all this ends with the passage of time.