We are used to increasingly advanced processors with an amazing number of transistors, but we had never seen anything like the chip developed by a company called Cerebras .
Its “Wafer Scale Engine” or WSE of Cerebras is the exaggeration made chip: it has 1.2 billion transistors , 400,000 process cores and 18 GB of memory integrated in an absolutely huge SoC and that is larger than the Apple iPad, for example.
The announcement of this chip was made during the Hot Chips conference, one of the most relevant meetings of this segment worldwide. The company also did it with a specific approach to this development: that of artificial intelligence .
Sean Lie, co-founder of Cerebras with Andrew Feldman. In the image, its WSE Cerebras chip. It’s not a wafer, eye: that’s the chip.
That is the field in which, according to its creators, the WSE Brain can stand out from other competing products. The integration of those 18 gigabytes of memory multiplies by 3,000 what we had seen in some current high-end GPUs, and that allows according to the company that this great requirement of large applications of deep learning, for example, are benefited by this design .
This design is really impressive, but as they point out in TechCrunch it is not without problems. One of the challenges is of course manufacturing : on a chip of these dimensions the problems and potential defects multiply, although in Cerebras they say they have been solved thanks to their work with TSMC, the manufacturer of these process beasts.
In the coming months we can see if this is so: the manufacturer has already begun to send some samples to the first customers, but if they are successful we could be facing a unique trend in chip design. The data around the chip is as astronomical as its price is probably , but that data has not been specified, although it will surely be very, very high.
The comparisons are hateful
The chip is huge: it has an area of 46,225 mm 2 compared to 4,410.9 mm 2 of the AMD Epyc Rome , for example . This last chip, which is probably one of the densest in the market, has 32,000 million transistors on that surface, which makes clear the size of the WSE Cerebrae and its 1.2 billion transistors: that’s 40 times more than which has the AMD SoC, already gigantic in itself. To give you an idea, the mobile SoC is much lower in size, and a Snapdragon 855 round 112 square mm.
The AMD Ryzen 3000 with Zen 2 architecture have for example about 3.9 billion transistors. Intel has not offered the number of transistors of its processors for a long time, although estimates indicate that a Core i7-9700K should be around 3 billion transistors .
It is logical that they are below AMD in this area since Intel maintains the 14 nm lithographic process while AMD has just made the leap at 7 nm with those 3rd generation Ryzen. The mobile processors have long made that leap, and we have a good example of the achievements that this technology makes possible: the Apple A12 Bionic has 6.9 billion transistors, for example.
If we talk about graphics, the numbers grow: a GeForce RTX 2080 Ti has 18.6 billion transistors, while the recent Radeon RX 5700 / XT has 10.3 billion transistors (close to 10.8 billion of the RTX 2060/2070 with the that compete). They are equally remarkable figures, but all of them pale in the dimensions and characteristics of the Brain chip .