Shulou(Shulou.com)11/24 Report--

Thanks to CTOnews.com netizens OC_Formula and Alejandro86 for the clue delivery! CTOnews.com, October 24 (Xinhua)-- IBM Research has launched the AI chip NorthPole, which is inspired by "the operation of the human brain" and deduces that its performance is said to surpass that of 4nm GPU and is suitable for areas such as edge computing.

After inquiry, CTOnews.com learned that the NorthPole chip is the successor to the TrueNorth chip that IBM "simulated the operation of the human brain" in 2014, and the chip development was also led by Dharmendra Modha, the head of the TrueNorth chip.

▲ image source IBM reported that in the traditional semiconductor industry, chips mainly follow the same basic architecture, and the processing unit and storage information are separated from each other. Although this architecture simplifies the chip design pattern, there is a "von Neumann bottleneck (von Neumann Bottleneck)" because the transmission speed can not catch up with the processing speed, while Dharmendra Modha believes that the human brain is the most energy-efficient processor known so far. So keep looking for ways to replicate the human brain digitally.

The biggest difference between the NorthPole chip currently launched by IBM and the traditional chip lies in the "chip built-in memory". In the absence of "von Neumann bottleneck", the AI inference ability of NorthPole chip is better than that of competitive products on the market.

Using 12nm process, NorthPole has 22 billion transistors on 800mm and 256cores. Each core can perform 2048 operations per cycle with 8-bit precision, and the number of operations can be doubled with 4-bit or 2-bit precision.

▲ is equipped with NorthPole's PCIe card. In terms of specific architecture, NorthPole claims to blur the line between computing and memory, which makes it easy for NorthPole to be integrated into the system and significantly reduces the load of devices carrying chips.

IBM Research tests NorthPole on the ResNet-50 model. Compared with the GPU competitor, which is also based on the 12nm process, NorthPole is 25 times more efficient in identifying frames per second than the competitor, and the performance is better than all the mainstream architectures in the market, whether in terms of latency or computing space requirements, even higher than the GPU based on the 4nm process.

However, the advantage of NorthPole is also its weakness. NorthPole can only easily read the local data information integrated in the chip, but there is no computing speed advantage when reading external data.

Dharmendra Modha claims that although NorthPole cannot be used to host GPT-4, it should meet the model inference requirements required by many enterprises.

At present, IBM Research is still studying the applicable fields of NorthPole. Many edge computing that researchers need to process a large amount of data in real time may be very suitable for NorthPole, such as autopilot, remote sensing communication and other fields. NorthPole has the opportunity to show its ability.

Welcome to subscribe "Shulou Technology Information " to get latest news, interesting things and hot topics in the IT industry, and controls the hottest and latest Internet news, technology news and IT industry trends.