IBM Research recently launched the AI chip NorthPole. The chip is said to be inspired by the operation of the human brain. Its inferred performance is said to surpass 4nm GPUs and is suitable for edge computing and other fields.
After inquiries that the NorthPole chip is the successor of the TrueNorth chip that IBM used to “simulate the operation of the human brain” in 2014. The chip development is also led by Dharmendra Modha, the head of the TrueNorth chip.
▲Image source IBM
It is reported that in the traditional semiconductor industry, chips mainly follow the same basic architecture, and the processing unit and stored information are separated from each other. Although this architecture simplifies the chip design model, it also causes “Feng” because the transmission speed cannot keep up with the processing speed. “von Neumann Bottleneck”, and Dharmendra Modha believes that the human brain is the most energy-efficient processor currently known, and therefore continues to look for ways to digitally replicate the human brain.
The biggest difference between the NorthPole chip currently launched by IBM and traditional chips is the “chip built-in memory”. Without the “Von Neumann bottleneck”, the AI inference capabilities of the NorthPole chip are better than those of competing products on the market.
Although NorthPole uses a 12nm process, 22 billion transistors are placed on 800 square millimeters, and it has 256 cores. At 8-bit precision, each core can perform 2048 operations per cycle. If it is at 4-bit or 2-bit precision, , the number of operations can be doubled.
▲ PCIe card equipped with NorthPole, picture source IBM
In terms of specific architecture, NorthPole claims to blur the boundaries between computing and storage, which makes it easy to integrate NorthPole into the system and significantly reduces the load of the equipment equipped with the chip.
IBM Research tested NorthPole on the ResNet-50 model. Compared with competing GPU products based on the 12nm process, NorthPole’s energy efficiency in identifying frames per second is 25 times that of competing products, and whether it is in terms of latency or computing space requirements, The performance is better than all mainstream architectures on the market, even better than GPUs based on the 4nm process.
However, NorthPole’s advantage is also its weakness. NorthPole can only easily read local data information integrated into the chip. When reading external data, it has no computing speed advantage.
Dharmendra Modha claimed that while NorthPole cannot be used to host GPT-4, it should meet the model inference requirements required by many enterprises.
Currently, IBM Research is still studying the applicable fields of NorthPole. Many edge computing applications that require real-time processing of large amounts of data may be very suitable for NorthPole, such as autonomous driving, remote sensing communications and other fields, where NorthPole can be used.
THIS IS A SPONSOR PROMOTION: >>>>>>>>>>>>>
Geekwills is an online shop that connects consumers with millions of products and brands around the world with the mission to empower them to live their best lives. Geekwills is committed to offering the most affordable quality products to enable consumers and sellers to fulfill their dreams in an inclusive environment.
