IBM, a true tech dinosaur, unveils the NorthPole chip, fully optimized for AI and which offers frankly astonishing performance. A potential candidate to redefine the future of artificial intelligence?
In the field of AI, IBM stood out in May. His boss announced a suspension of hiring of several thousand positions in his company, precisely in relation to the explosion of this technology. Even if it is applied today on a large scale, there is no equipment specially developed for it yet.
In the IBM research laboratory located in Almaden (California), brains are boiling to make things happen. Their latest advances in chip architecture could well prove to be a very effective solution to overcoming current hardware limitations.
NorthPole: a new generation of chips
Since the invention of semiconductors, chips have remained structurally identical, which greatly affects the field of artificial intelligence. Known as ” von Neumann problem ” Or ” von Neumann bottleneck », the phenomenon limits the transfer of data between the processor and the memory of the systems. This is why IBM intends to tackle this structural problem by creating a neuromorphic chip imitating the functioning of the human brain.
Dharmendra Modha is the head of the NorthPole chip project, which is distinguished firstly by very low energy consumption. A considerable advantage when we know the energy impact of AI on a global scale. This chip is approximately 4,000 times faster than its predecessor, “TrueNorth”. The first test results were published in the famous journal Science and are very promising. When compared to GPUs and CPUs currently used for training AI models, NorthPole demonstrates 25 times greater energy capacity.
If NorthPole is so effective, it is because its structure is particular: the memory is integrated directly on the chip. It is this specific construction that eliminates the von Neumann problem. However, this design causes certain significant constraints. In fact, the chip only has access to onboard memory. If it is used to process huge amounts of data, this could pose a problem. IIBM believes it can circumvent this limitation with a technique called “scale-out”, which allows neural networks to be subdivided to make them more efficient.
Towards mass adoption in various fields
NorthPole is intended for a wide range of applications, in very different fields. It could be of valuable help in the sector of robotics or autonomous vehicles, for example. Digital assistants could also benefit, as could satellite-assisted agriculture. Its very compact design also eliminates the need for large cooling systems, making it easily adaptable to any type of environment.
NorthPole is only just getting started, but its results are already very encouraging. IBM is already considering the future, by further improving its engraving finesse to increase it from 3 to 2 nanometers. The computer industry owes a lot to this company, which has established itself for decades thanks to major innovations: the IBM PC, the System/360, the 8-inch floppy disk or the hard disk. Perhaps it will be the basis for the emergence of a new paradigm in the AI sector?
Source : IBM
0