When it comes to pushing the boundaries of AI training, Tesla isn’t messing around. Elon shed light on the Tesla’s AI ambitious plans for its Giga Texas.
Giga Texas training cluster will be a behemoth, packing a whopping 50k H100 GPUs from NVIDIA. Tesla is also throwing in 20k of its own HW 4.0 AI computers into the mix, creating a training powerhouse unlike anything we’ve seen before.
Now, here’s where things get really interesting. Elon confirmed that Tesla’s HW4 AI computers are part of a closed-loop training system, and they’re about to shake things up in a big way. Traditionally, Hardware-in-the-Loop (HIL) has been used for testing and verification, but Tesla is flipping the script and implementing HIL in the training phase, starting with the Giga Texas cluster.
This shift towards a training-focused HIL approach suggests that Tesla sees HIL as more than just a testing tool. Instead, it could be a way to optimize performance and accelerate feedback iterations during the training process itself. Elon vision involves pairing HW 4.0 with 5-megapixel cameras, end-to-end HIL training and testing, and consolidating everything into a single, massive cluster of 50k H100 GPUs.
Elon Musk expects model divergence – where HW 4.0 surpasses the capabilities of the previous HW 3.0 – to kick in by Q4 of this year. As time goes on, the gap between the two models is only going to widen, with no overlap on the horizon.
Details are still scarce on how exactly Tesla will leverage all this obscene AI muscle. Elon made one thing crystal clear: Traditional closed-door training just won’t cut it for his lofty full self-driving ambitions. For that, you need off-the-charts scale and unconventional thinking in spades.