Russell Baker on 4/18/12 wrote:
> 1. Nvidia will release a Tesla board based on GK104 that will not have ECC for both DRAM and on-chip SRAM.

Tesla sales are hundreds of millions of dollars per year. The cost of making a derivative of GK104 with ECC for both DRAM and on-chip SRAM is under ten million dollars, assuming ECC is the only addition. The management of Nvidia would have to be a bunch of idiots to sell a Tesla board without ECC for both DRAM and on-chip SRAM. I don't think the management of Nvidia is a bunch of idiots so I think this prediction is wrong.

> 2. Someone in the financial industry will use that Tesla board for Black-Scholes
> option pricing or some other calculation used for algorithmic trading.
>
> 3. A soft error will occur in the on-chip SRAM of the Tesla board which will cause
> a stock market "flash crash" just like when the Dow dropped 600 points in 5 minutes
> on May 6, 2010 due to algorithmic trading.

These predictions are plausible if prediction 1 is true but I think prediction 1 is wrong.

> 4. Nvidia will say "We're so sorry. We never imagined this could happen."

About 18 years ago, Intel had a bug in the floating point divide instruction of their Pentium microprocessor. Although Intel's bug was a hard error, the situation for the end-user was similar in that the chip occasionally produced the wrong answer. Intel was in denial for months. They claimed errors would be very rare. Most customers didn't want a computing chip that made any errors. Eventually, Intel took a $475M charge against earnings and admitted that they screwed up.