AI report compares 5nm and earlier node wafer costs

September 22, 2020 //By Peter Clarke
AI report compares 5nm and earlier node wafer costs
A report on the costs of designing and making artificial intelligence (AI) processors reveals that foundry TSMC is likely to be charging of the order of $17,000 for each 5nm wafer.

This represents a near doubling of the price from the 7nm node. The report assesses that typically 71.4 chips would be contained on each 300mm-diameter wafer made using 5nm manufacturing process technology resulting in typical per chip price to fabless chip companies of $238.

This is based on a hypothetical 5nm GPU with similar specification to Nvidia's 16nm Tesla P100 in terms of die area – 610 square millimeters – but scaled up to 90.7 billion transistors – versus 15 billion.

The 72-page report is called AI chips: What they are and why the matter.  It was prepared by two authors from the Center for Security and Emerging Technology (CSET) at Georgetown University’s Walsh School of Foreign Service.

CSET is currently focusing on the effects of progress in artificial intelligence (AI) and advanced computing. The report asserts that the US government is considering how to control the spread of AI technology; and that because AI software, algorithms and datasets are not effective targets for control, hardware is the natural lever of control.

As part of their deliberations the authors point out that training a leading AI algorithm can require a month of computing time and cost $100 million.

The report details why leading-edge chips are more cost-effective than older generations of chips and why specialized AI processors are more cost-effective than general-purpose processors. To do this the report surveys the semiconductor industry and the AI chip design landscape.

As a result the authors not only come up with a chip manufacturing cost of $238 but also a design cost of $108 per chip and a packaging and test cost of $80 per chip. This yields a total chip cost to the fabless developer at the 5nm node of $426.

Table 9: Calculation of foundry sale price per chip in 2020 by node. Source: AI chips: What they are and why they matter.

The report's conclusion is that China’s lack of capability in EDA software, semiconductor manufacturing equipment and leading-edge fabs means the US and its allies maintain a competitive advantage in the production of leading-edge AI chips. It goes on to say that because AI chips have critical strategic value for advanced security it is vital to US and allied interests, and for global security, that this advantage is maintained.

Related links and articles:

cset.georgetown.edu

AI chips: What they are and why they matter

News articles:

Opinion: Nvidia's bad deal is not yet done

Universal processor startup goes for split tape-out

Graphcore launches 7nm AI processor

Groq gains funds to expand tensor AI applications

Share price soars as lossy AI 'unicorn' Cambricon goes public

'Starbleed' vulnerability revealed in Xilinx FPGAs


Vous êtes certain ?

Si vous désactivez les cookies, vous ne pouvez plus naviguer sur le site.

Vous allez être rediriger vers Google.