Fault injection

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
Posts: 902
Joined: Fri Jan 06, 2012 2:13 pm

Re: Fault injection

Post by mgates3 » Mon Aug 05, 2019 12:55 pm

That seems fine for testing.

MAGMA doesn't do anything to detect or prevent faults from bit flips. If the bit happens to be low-order (like 0.00000001), then the difference will be negligible, but if the bit happens to be high-order (like 1.) or an exponent bit, the error is likely to be significant.

There are fault detecting & tolerant linear algebra algorithms, we just haven't implemented them in MAGMA.

Note that ECC memory normally protects against single bit flips, so what you are doing would correspond to a double bit flip that the hardware was unable to correct, or using non-ECC memory (gaming GPUs generally use non-ECC memory).


Post Reply