Northpole, the first AI-dedicated microprocessor built by IBM Research, supposedly boasts 22 times quicker performance than comparable industry solutions. The information comes from a research published in the journal Science on IBM’s forthcoming AI accelerator dubbed NorthPole. With the influx of the “AI frenzy” in the industry, many chip manufacturers are shifting toward developing their own solutions in order to outperform the computing performance of the industry leaders and meet the increasing demand for AI.
IBM Research states that the “NorthPole” AI processor is poised to create new industry records
Dharmendra Modha, the project’s leader, is optimistic about the chip design. Now, IBM Research’s implementation is that the company integrates neural inference frameworks into chip processing, which Modha classifies as a “human brain.” NorthPole‘s high performance is due to its efficient CPU interconnection and all-digital architecture, which allows for significantly faster inter-communication.
In terms of specs, the NorthPole AI chip uses a 12nm node processing technology, which is somewhat ancient in the industry; nonetheless, IBM Research thinks that the chip outperforms recent 4nm AI GPUs as well, thanks to the use of the ResNet-50 neural network model. This achievement contradicts Moore’s Law as well as, to some extent, Huang’s Law, which focuses on individual chip stacks rather than process downsizing.
If we consider NorthPole’s potential impact on the AI industry as a whole, it is almost probably confined to “model inferencing,” as the processor lacks capability for large-scale neural networks such as GPT-4. However, the chip’s purpose is not to target mainstream AI markets, but rather ones focused solely on inference, which explains why its influence is limited.
Also Read:
