a large compute plane, with extremely high
bandwidth, and low latencies. During the
event, Tesla revealed it adopted a top to
bottom approach to help scale performance,
with the smallest entity being the training
node to help address latency and bandwidth
issues. The high-performance training node is
capable of 1024 GFLOPS and 512 GB/s in each
cardinal direction – truly incredible igures.
A computer that the company showed of
ofered 354 training nodes capable of 362
TFLOPs, whilst the irm’s D1 Chip is a machine
learning machine capable that’s 362 TFLOPs
and is designed completely in-house. Tesla’s
unique integration process was designed to
preserve bandwidth, resulting in training tiles
with 9 PFLOP. Dojo will be the fastest AI training
computer in the world with four times the
performance, 1.3X better performance/W, and
5X smaller footprint compared to what exists,
and the company is recruiting “heavily” to
help speed up its artiicial intelligence journey
- with the talent to be deployed across the
entire leet. Tesla says that the latest iteration of
Dojo will be operational next year, but admits
that there are problems with scalability and
distribution which need to be ironed out irst.
The irm says that the diiculty is how to keep
the localities and that its primary ambition is
to train vast avoids of video, reducing training
time, and improving safety and performance,
but by the time the Tesla bot is introduced in
more than a decade’s time - perhaps in 2035,
or maybe even 2040, that technology will be
even further advanced and could unlock even
greater powers and abilities for the company.
In short, the Tesla Bot would serve as a