Home  Article

Aug 20 2019

What Is The Largest Semiconductor Chip In History?

Catalog

Article Core:Cerebras Systems has launched the historically largest semiconductor chip and what chanllenges they have met.
I
About 1.2 Trillion Transistor Chip
II
About Cerebras Systems
III
Challenges On WSE

About 1.2 Trillion Transistor Chip

According to ventruebeat, Cerebras Systems, an AI start-up, has launched the Wafer Scale Engine (WSE), the largest semiconductor chip in history. It is understood that the chip has 400000 nuclear (4 core processor 100000 times), the processing power to support a large amount of calculation, at the same time, it use the advanced packaging technology of silicon interconnection, achieving silicon level of high-speed communication and storage, which in theory that the entire chip can handle extremely complex AI tasks at the fastest possible speed.


Cerebras WSE 

At present, the key indicators revealed by the authorities are as follows:

Silicon wafer size 42,225 square millimeters (about 22cm on each side)

l1.2 trillion transistors

l400,000 AI cores

l18GB of on-chip memory

lMemory bandwidth 9PB/S

lThe bandwidth of architecture interconnection (Fabric type) is 100PB/S

lTSMC 16nm process was adopted

the key indicators 

Although this model has a huge area (up to 46,225 square millimeters), it still uses TSMC's 16-nanometer manufacturing process. It contains 400,000 cores, up to 18G of on-chip storage, 15,000 watts of power consumption (about equal to the power of 6 induction cookers), 9PB/ s of memory bandwidth, and 100PB/ s of Fabric bandwidth. WSE has 1.2 trillion transistors, the most recent Advanced Micro Devices processor has 32 billion transistors, and Advanced Micro Devices CEO Lisa Su announced that the second-generation Epyc processor code-named "Rome" is the most powerful x86 processor in the world. More cores, more local memory, and a high-bandwidth structure with low latency create the best architecture for accelerating AI work. The WSE is 56.7 times larger than the largest GPU, Nvidia's latest generation flagship GPU Titan V.

Cerebras WSE was designed for AI and contains a number of fundamental innovations that address decades-long technological challenges that have limited chip sizes, such as yield per wafer, power output, packaging, etc., "Cerebras WSE CEO Andrew Fieldman said in a statement.” All of these architectures are designed to optimize the performance of the AI's work. Overall, WSE provides hundreds or thousands of times the performance of existing solutions with minimal power consumption and space."

Chip size is important in AI because large chips can process information faster and produce answers in less time. Chip size is determined by chip designers and even the whole chip design team to communicate together, there are costs and benefits involved in many aspects of the problem. Generally speaking, under the conditions of the same architecture, the larger the chip design, the higher the specification, and the stronger the performance. However, due to various conditions, the chip cannot be expanded indefinitely. The size of the chip is closely related to the manufacturing process. If there is only one chip on a wafer, there is a 100% chance that it will have impurities, which will cause the chip to fail. But Cerebras' chips are designed to be redundant, so one impurity won't destroy the whole thing.

At the moment, the WSE's computing power is worth acknowledging, but the 16nm process and its ultra-high power make cooling a problem. One expert suggested that installing such chips in many data centres would be impractical. In the early 1980s, American firms took hundreds of millions of dollars in Trilogy funding to create their own superchips. However, the processor was too hot to test and not as powerful as initially thought. At the same time, the single WES price will not be low, at least far higher than the price of a dozen nvidia TitianV.

Dr. Ian Cutress, senior editor at the news website AnandTech, said advances in technology come at a cost. "One of the advantages of small computer chips is that they use less power and are easier to keep cool," he explains. "When you start dealing with larger chips like this, companies need specialized infrastructure to support them, which limits who can actually use it. "That's why it's good for AI development, because that's where the big money is going right now."

In response, Andrew said the chip would not be sold separately but would be packaged into a 'device' designed by Cerebras. The reason is that it requires a complex water-cooling system to dissipate heat. The water cooling system is an irrigation network that can offset the extreme heat generated by chips running at 15 kilowatts.

Cerebras WSE 

Because of cooling, the chip can't be plugged into any existing servers as a speed card. Under this design, it is expected to be very high power, but compared with the system cluster it is supposed to replace, its power consumption is very low.


About Cerebras Systems

The chip comes from a team led by Andrew Feldman, who founded SeaMicro, a micro-server company that was sold to AMD for $334 million.Feldman’s partner in crime, co-founder Gary Lauterbach, has been working on chips for 37 years and has 50 patents on the techniques and tricks of the art of design.  “It’s like a marriage,” says Feldman of the twelve years they’ve been collaborating.Feldman and Lauterbach have gotten just over $200 million from prominent venture capitalists because of a belief size matters in making A.I. move forward. Backers include Benchmark, which funded Twitter, Snap, and WeWork.


Challenges On WSE

First, a quick look at the chips that power phones and computers. Fabs such as TSMC take standard-sized silicon wafers and use light to etch transistors into chips, splitting them into individual chips. One of the challenges with this lithography process is that errors can spill over into the manufacturing process, requiring extensive testing to verify quality and forcing fabs to abandon poorly performing chips.

Cerebras has come up with the idea of etching a mass of individual chips onto a single wafer, instead of just using the wafer itself as a single giant chip. This allows all these individual cores to connect directly to each other, greatly accelerating the critical feedback loops used in deep learning algorithms - but at the expense of the huge manufacturing and design challenges of creating and managing these chips.

According to Feldman, the first challenge the team faced was dealing with "crossed" communication. While Cerebras' chips contain complete wafers, today's lithography devices still have to etch individual chips like silicon wafers. The second challenge is yields. Using a chip that covers the entire silicon wafer, etching a single defect in the chip may render the entire chip useless. For decades, this has been a barrier to wafer technology: because of the laws of physics, it is virtually impossible to repeatedly etch trillions of transistors with perfect precision. The third challenge Cerebras faces is dealing with thermal inflation. Chips get very hot as they work, but different materials expand at different rates. This means that the connectors that bind the chips to their motherboards need to expand at the same rate to avoid creating cracks between the two.

Once the chips are made, they need to be tested and packaged for shipment to original equipment manufacturers (OEMs), who add the chips to products used by end customers, whether in data centers or consumer laptops. But there is a challenge: there is absolutely nothing on the market designed to handle the entire wafer.

WSE is succeed because Cerebras has made great efforts and made great innovation and reform in all aspects, Feldman explains,"So we designed this whole manufacturing process because no one had done it." Cerebras' technology isn't just about the chips it sells - it's about all the machines needed to actually make and package them.

Hot Chips activity site introduction of WSE 

Finally, all the processing power in a chip requires huge amounts of power and cooling. Cerebras' chips run on 15 kilowatts of power - providing huge amounts of power to a single chip. Cerebras had to devise a new way to provide both functions for such a large chip. Cooling such a massive chip requires more than radiators and fans. Cerebras said it had installed a 'cold board' over the wafers and used vertical pipes to cool the chips directly. Because the chip is too big to fit into any conventional package, Cerebras designed its own chip, which combines PCB, wafer, custom connector to connect the two, and cold plate.


You May Also Like

Wireless Electricity?What is a Tesla Coil?

Integrated Circuit and Motherboard: Is the Motherboard an Integrated Circuit ?

What are Varistor and Metal-Oxide Varistor? Application and Selection of Varistor

0 comment

Leave a Reply

Your email address will not be published.

 
 
   
Rating: