Cerebras' wafer-size AI chips play nice with PyTorch, TensorFlow

Good news for those who like their AI chips big: Cerebras Systems has expanded support for the popular open-source PyTorch and TensorFlow machine-learning frameworks on the Wafer-Scale Engine 2 processors that power its CS-2 system. The chip designer says the expanded support, announced today, is an important milestone because …

[Anonymous Coward]

They may be relatively new to the market, but I think this is a serious contender in their space, much like Teradata was at the peak of relational technology.

[katrinab]

What are production yields going to be like?

In normal chip manufacturing, how often do you get a wafer where every single chip on the wafer is functional?

[Anonymous Coward]

Never.

[Martin Gregorie]

Repairability?

Wafer repair is presumably not an option, unlike a conventional chassis full of replaceable system boards. As a result, if anything significant fails on a single wafer system, it seems very likely that you'd have to replace the entire wafer unless, of course, there are redundant subsystems designed into it. However, since I've seen nothing about redundancy being built into these systems-on-wafers, it seems likely that it doesn't exist. Otherwise it would be used as a selling point.

What did I miss?

[BOFH in Training]

Re: Repairability?

Redundancy is built in.

https://cerebras.net/blog/wafer-scale-processors-the-time-has-come/

According to this : https://techcrunch.com/2019/08/19/the-five-technical-challenges-cerebras-overcame-in-building-the-first-trillion-transistor-chip/?guccounter=1

they have a 1 to 1.5% redundancy in cores.

[Kevin McMurtrie]

400000 cores and Python

World's most compact space heater. I've seen people try to use 3 cores in Python and it's all spin locks and dead locks. Even the acceleration libraries are bad because efficient mathematics and efficient concurrency together are a lot to master.

[Tom 7]

Re: 400000 cores and Python

Not everyone gives up solving problems as easily as you seem to.