ResNet-50 on ImageNet now (allegedly) down to 224sec (3.7min) https://arxiv.org/abs/1811.05233 using 2176 V100s. Increasing batch size schedule, LARS, 5 epoch LR warmup, synch BN without mov avg. (mixed) fp16 training. "2D-Torus" all-reduce on NCCL2, with NVLink2 & 2 IB EDR interconnect
-
-
“God” is statistically very likely to be an
#AI from base reality trying to understand its surroundings by simulating an infinite#multiverse
.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
As always highly impressive. Can you give me some backup in my plan to use AI to turn Model 3 into kinda pacecar to regulate traffic? People getting interested over here and would be nice to hear this can be done.https://twitter.com/HansNoordsij/status/1062465481356050432 …
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
When you have a chance please dm me I would like to chat .
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Is this still Tesla and not Volta? Would be interesting to see what Volta can achieve with tensor cores.
-
Nvidia Tesla is not an architecture, but a product name. Tesla V100 is Volta-based and does have Tensor Cores (P is Pascal-based).
End of conversation
New conversation -
-
-
If the universe didn’t allow it we probably wouldn’t be here to appreciate it.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Cool. Is there a compute cost column for each row ?
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
-
-
My main takeaway is how little I know about neural nets.
Thanks. Twitter will use this to make your timeline better. UndoUndo
-
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.






