Over a trillion transistors, the’largest computer chip ever’

Cerebras Systems, a startup that makes chips dedicated to machine learning and deep learning, announced that it has developed the Wafer Scale Engine, the largest computer chip ever.

Wafer-scale engines are 20×20cm in size and are equipped with 1.2 trillion transistors. The company announced details about the wafer scale engine at Hot Chips, a conference on microprocessors held at Stanford University on August 19th. The area is 42,225mm2, and 400,000 SLA cores are connected with a bandwidth of 100Pb/sec. It is the largest chip for processing. The largest so far is the Tesla V100, NVIDIA’s data center GPU, but it is 56.7 times larger than that.

Itanium, a 64-bit microprocessor announced by Intel in 2002, has 221 million transistors. However, the wafer-scale engine contained 1.2 trillion units, which is 5,000 times that of this. Considering that IBM’s new Power 9 model has 8 billion transistors, it is a big part of the scale. The wafer scale engine also features 18GB SDRAM and supports 9Pb/sec memory bandwidth. It is said that the manufacturing process used TSMC’s 16nm manufacturing process.

However, a chip that is as large as this can contain defects in the cores on the wafer. Usually, bad chips with such defects are discarded, but the wafer-scale engine has more than 400,000 cores, and a redundant design is being made that assumes that some of the cores cannot be used due to defects.

Today is the perfect day to share our first tweet as we have announced the largest chip ever built to accelerate AI compute! Read all about the Cerebras Wafer Scale Engine (WSE) via @WIRED and @tsimonite https://t.co/ATaBRnxnGD pic.twitter.com/dX0xetyX6X

— Cerebras Systems (@CerebrasSystems) August 19, 2019

https://platform.twitter.com/widgets.js

In addition, giant chips have cooling issues as well as manufacturing. To cool a chip the size of a computer keyboard, it is difficult to solve it with a heat sink and fan, which are the existing cooling systems. It is necessary to install a water cooling system that directly cools the chip by directly installing a pipe through which the coolant passes. Because the chip is too large to fit into the existing package, the manufacturer said that it is preparing an optimized cooling system and connectors for wafer scale engines.

Of course, this product is not for personal PCs. It is assumed to be used in data centers that perform machine learning or deep learning experiments. The manufacturer expects the wafer-scale engine to be active in large-scale development projects such as AI and autonomous vehicles. Related information can be found here .