Read the full version
21.10.2023 13:09, Gennady Detinich
IBM Company reported, which has completed testing of a new prototype processor for artificial intelligence tasks. The new development, codenamed NorthPole, proved to be 4,000 times better than the company’s previous AI architecture, called TrueNorth, and “breathtakingly” outperformed all the most advanced central and graphic processors.
IBM NorthPole processor on PCIe cards. Image source: IBM
The NorthPole chip is manufactured using a 12nm process technology and contains 22 billion transistors over an area of 800 mm2. This is actually a “network on a chip” – this processor contains 256 cores with an extensive interface and built-in memory. It is thanks to the memory built into the chip that it was possible to achieve the industry’s best performance in terms of energy efficiency, reduced latency and effective area.
In one clock cycle, the NorthPole processor performs 2048 operations per core (with 8-bit precision). For 4- and 2-bit precision, the number of operations performed is doubled and quadrupled, respectively. This ability is aimed primarily at image processing. More precisely, for digital machine vision, and these are autopilots, autosurgeons, and so on.
The bottleneck of the von Neumann architecture was and remains the separation of memory and processor. IBM developers overcame this obstacle when they created a processor that stores all data within itself, without sending it to external storage devices.
Testing on the ResNet50 model, a 50-layer neural network for testing image recognition and classification solutions, showed that the energy efficiency of the NorthPole chip is 25 times higher than the energy efficiency of conventional 12nm GPUs and 14nm CPUs. Also, the latency indicators were 22 times better, which were lower for the IBM chip. The developers called it a “mind-blowing” result. Finally, in terms of chip area utilization (number of transistors), IBM’s architecture also outperformed all competitors, including even 4nm GPUs.
2023-10-21 10:09:00
#IBM #developed #chip #internal #data #storage #order #magnitude #faster #Nvidia #accelerators
