How in the world is this possible?! (By write in your GPU code, I assume you mean the write was GPU-initiated?)
-
-
Replying to @cr1901
Yes. But in times of CUDA uniform memory a GPU write is not limited to GPU memory.
1 reply 0 retweets 4 likes -
I've managed to simply crash my machine with a buggy GPU kernels many times. Sometimes it just kills my desktop but can still login via ssh.
3 replies 0 retweets 2 likes -
Theoretically the memory protection scheme used by the OS is replicated on the GPU. But in practice we don't seem to be there yet..
2 replies 0 retweets 1 like -
That should be handled by the IOMMU unit, not the GPU though. BTW does your CPU have one? (DMAR in dmesg)?
1 reply 0 retweets 2 likes -
Replying to @astarasikov @cr1901
Hmm.. Should have (AMD FX, aka AMD Bulldozer), but no "DMAR" in dmesg. Maybe there's a BIOS switch that I have to enable? Rebooting..
2 replies 0 retweets 0 likes -
Sorry I don't know about AMD, need to look it up, I think DMAR might be Intel's name
1 reply 0 retweets 1 like -
Replying to @astarasikov @cr1901
Yep. Was disabled in BIOS. What a reasonable default! Grrr. Now that its enabled this should drastically improve my GPU hacking productivitypic.twitter.com/QdWWln9qHk
2 replies 0 retweets 9 likes -
And afaict from a quick test the performance of my GPU kernels is unchanged. Thank you so much, this was very helpful!
1 reply 0 retweets 0 likes -
Replying to @oe1cxw
Did enabling IOMMU prevent system crashes from the same GPU kernels that caused crashes before? (So far you only tweeted that it "should".)
1 reply 0 retweets 1 like
Hard to tell. Even before the change, in most cases an illegal memory access would just terminate the kernel w/o noticeable side effects.
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.