Comment by nickandbro
9 hours ago
I am curious what workloads Citadel Securities is running on these TPUs? Are you telling me they need the latest TPUs for market insights?
9 hours ago
I am curious what workloads Citadel Securities is running on these TPUs? Are you telling me they need the latest TPUs for market insights?
Training their own, closed, internal models on their own data sets? Probably a good way to squeeze out some market trading signals.
Reminds me of when hedge funds started laying increasingly shorter fiber-optic cable lines to achieve the lowest possible latency for high-frequency trading.
I thought these TPUs were primarily used for inference?
TPU8t is for training. But even still, once you’ve trained, you need to run the model too. And these kinds of models already have a huge latency hit so there’s not much hurting running it away from the trading switches.
As the article states, there's both training and inference dedicated chips.
Not Citadel, but Jane Street is training LLMs for trading:
https://www.janestreet.com/join-jane-street/machine-learning...
> We build on the latest papers in LLMs, computer vision, RL, training libraries, cuda kernels, or whatever else we need to train good models.
> We invent our own set of architectures and optimizations that work for trading.