Cerebras Systems is understood for its avant-garde chip designs. The chipmaker has as soon as once more managed to show heads with the announcement of Wafer Scale Engine 2 (WSE 2), the world’s largest chip, primarily based on 7nm course of node.
Measuring roughly round 46,225 mm sq. (50 instances the scale of the most important GPU) and a processor with nearly a million cores (850,000), WSE2 is a 123x improvement over the Ampere A100, NVIDIA’s largest GPU with 54 billion transistors and seven,433 cores.
When the semiconductor trade is striving to construct smaller parts, the introduction of the Wafer Scale Engine (WSE 1) caught everybody off-guard. Cerebras Systems has designed and manufactured the most important chip completely for optimised deep studying. Deep learning is without doubt one of the most computationally intensive workloads. Moreover, DL is kind of time-consuming because it makes use of a number of layer loops to extract higher-level options from the uncooked enter. The solely approach to scale back coaching time is to chop down the time taken for inputs to cross via a number of layer loops, which may be achieved by growing the variety of cores to extend calculation pace. The Wafer Scale Engine was constructed to deal with this want.
Two years in the past, Cerebras challenged Moore’s Law with the Cerebras Wafer Scale Engine (WSE). The earlier era of WSE chip with 1.2 trillion transistors trounced Moore’s regulation by an enormous margin. Moore’s Law states that the variety of transistors on a microchip doubles each two years, whereas the price of computer systems is halved.
So, what’s the catch?
According to Our World Data, until 2019, the following largest transistor rely for a microprocessor is AMD’s Epyc Rome processor with 39.54 billion transistors. WSE may as properly be the tipping level for the semiconductor trade bringing forth a brand new period of AI chips which quadruple the variety of transistors yearly. The newly developed AI chip WSE 2 with 2.6 trillion transistors is proof of that risk.
So, what was the aim of introducing an enormous chip? Was it simply to show Moore’s Law was incorrect? The reply is way more sensible. According to Andre Feldman, co-founder and CEO of Cerebras Systems, the logic behind the size of WSE is kind of easy. Usually, a considerable amount of information is required to speed up AI, and the processing pace is essential.
Size has at all times been a limitation in machine studying processors, however all the things modified when Cerebras launched its revolutionary design. By constructing a wafer-scale chip on a single piece of silicon, Cerebras prevented all of the efficiency points like gradual off-chip communication, low reminiscence bandwidth, distant reminiscence and so on. Since reminiscence is an important element, its placement impacts the processor drastically; the nearer it’s, the sooner the calculations and consumes much less energy whereas shifting information. High-performance deep studying requires every core to function at optimum, which requires shut collaboration and proximity between the core and reminiscence.
In different phrases, the WSE achieved cluster-scale efficiency with out the penalties of constructing massive clusters and the constraints of distributed coaching. The Cerebras WSE-2 presently powers the CS-2 system. The new processor has 2.6 trillion transistors with 850,000 AI cores; all packed onto a single silicon wafer to ship main AI compute density at unprecedented low latencies. It accelerated deep studying to compute and lowered the time it took to coach fashions, enabling the duty to be accomplished in a couple of minutes. Cerebras WSE-2 has pushed previous the boundaries in AI chip improvement, ushering in a change in AI chips design and efficiency.
Join Our Telegram Group. Be a part of a fascinating on-line group. Join Here.
Subscribe to our Newsletter
Get the most recent updates and related presents by sharing your e mail.