Ben Koska (@BenKoska) header
Ben Koska (@BenKoska) profile picture

Doing linear algebra & optimizing GPU kernels. Court jester @SFTensor

Tensor Cores
Joined September 2019
Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)Photo by Ben Koska (@BenKoska)
if you are selling B300s no commit at $5/hr, I will literally take up to 1000 nodes
last month i was begging gpu providers for b300s at $7.9/hr now they're begging me to take them at $4.6/hr, no commitment something big is happening soon. ai gave us all its dessert and now it's time for the veggies
25
11
5
527
66,585
Ben Koska retweeted
We ran a normal training script, with no checkpointing and no idea it was moving and ran it as-is on resumable compute. Despite being moved around constantly, loss converged the same as it did staying on a single cluster.
1
2
12
561
Today, there are 3 classes of compute: On-Demand is reliable but expensive, Reserved requires you to commit to your spend years out and Spot is cheap but can be reclaimed at a moments notice and anything running dies. What if we could just... continue a process on a new host? 🧵
6
8
3
113
10,107