News TSMC mulls massive 1000W-class multi-chiplet processors with 40X the performance of standard models

[Admin]

TSMC is prepping a 9.5-reticle, 7,885 mm² multi-chiplet packages on 120×150 mm substrates with integrated power management for future AI and HPC processors.

TSMC mulls massive 1000W-class multi-chiplet processors with 40X the performance of standard models : Read more

 

[qxp]

Cerebras has been building wafer-scale computers for a while:

https://www.cerebras.ai/chip

 

[hotaru251]

gotta imagine the cost for that would be given how high the risk of defects are on such a large one

 

[ezst036]

We're gonna need a bigger nuclear power plant.

 

[qxp]

gotta imagine the cost for that would be given how high the risk of defects are on such a large one

The cost is not small, but that's because you are buying a whole wafer and it is a unique product. They work around defects by disabling some cores. The unique part is not just the more compact design, but also computational features, such as more bandwidth per operation. This can be a winning point if you compare it to GPUs, but, of course, the higher priced AI GPUs use HBM and negate some of the advantage.

Also, some defects are introduced when the wafer is cut, so using uncut wafer results in smaller defect rate.

 

[edzieba]

Cerebras has been building wafer-scale computers for a while:

https://www.cerebras.ai/chip

Tesla, too. Also fabbed by TSMC.
And as that article shows, TSMC announced SoW almost exactly a year ago (minus 1 day). With 2024 being a leap year, this could be an article entered into Toms' CMS with an off-by-one in the Year field.

 

[TJ Hooker]

I think Cerebras is a bit different, as their WSE is a monolithic piece of silicon fabbed from a single wafer. Tesla Dojo, and what is being discussed in this article, are bonding a bunch of separate chips onto a roughly wafer-sized package.

 

[fiyz]

Installing that chip is probably the only time I would ever consider wearing an esd strap.

 

[bit_user]

Tesla, too. Also fabbed by TSMC.

Tesla and Cerebras are different. Tesla stacks sub-reticle dies on a carrier, while Cerebras never cuts the dies out of the wafer, in the first place.

Like TJ said.