Tesla's Dojo is impressive, but it won't transform supercomputing

3 weeks ago 8
PR Distribution

Dojo looks bully astatine archetypal glance, but its reported show numbers and niche intent spot it extracurricular the ranks of existent supercomputers.


Image: Tesla

Tesla's AI Day event included the uncover of respective caller imaginable products, and portion bipedal robots are breathtaking (if possibly a spot unrealistic), the existent quality to travel is the uncover of Tesla's caller in-house designed supercomputer, called Dojo.

To telephone Dojo a full-fledged supercomputer is simply a spot generous, though: It hasn't been afloat assembled yet, and its imaginable show limits person yet to beryllium tested. What Tesla promises, though, is thing abbreviated of a supercomputing breakthrough. 

The most almighty supercomputer successful the world, Fugaku, lives astatine the RIKEN Center for Computational Science successful Japan. At its tested bounds it is susceptible of 442,010 teraflops (TFLOP) per second, and theoretically it could execute up to 537,212 TFLOPs per second. Dojo, Tesla said, could extremity up being susceptible of breaking the exaflop barrier, thing that nary supercomputing company, assemblage oregon authorities has been susceptible of doing. 

SEE: Hiring Kit: Video Game Programmer (TechRepublic Premium)

Putting that assertion successful position means knowing the standard and capabilities of Dojo and different supercomputers.

First, Dojo is designed to bash 1 peculiar thing: bid artificial intelligence. Tesla is gathering Dojo for usage in-house, processing video information from the millions of Tesla vehicles connected the road. Dojo is built connected Tesla's D1 chip, the 2nd the institution has designed. The spot is built utilizing seven-nanometer exertion and is independently susceptible of 362 TFLOPs per second. 

Dojo chips don't run individually, however, and the smallest portion Tesla has built is what it calls Dojo grooming tiles. These tiles are a transportation of 500,000 nodes that are reportedly susceptible of performing 9 petaflops per 2nd (1 PFLOP = 1,000 TFLOPs). All of that unthinkable velocity is disposable successful a tile little than 1 cubic ft successful size. 

Tesla's plans for Dojo tiles is to web them into larger systems. Its archetypal plan extremity is to physique a machine furniture susceptible of lodging 2 trays, each containing six Dojo tiles. In that configuration, Tesla said, it would beryllium capable to grip 100+ PFLOPs per second. Beyond that, Tesla plans to physique what it calls an ExaPOD consisting of 10 Dojo cabinets that volition beryllium capable to execute 1.1 exaflops (1 EFLOP = 1,000 PFLOPs). 

Fugaku, connected the different hand, takes up an full room, and astatine its highest measured show is susceptible of 442 PFLOPs. 

Back to that earlier perspective: Tesla believes that it volition beryllium susceptible of doubling the show of Fugaku with 422 less racks, and it volition beryllium capable to bash that by adjacent twelvemonth contempt lone having reached the milestone of gathering and investigating a azygous tile. 

The world of Tesla's Dojo claims

Dojo's reported capabilities don't assistance it existent high-performance machine (HPC) status, said Gartner probe vice president Chirag Dekate, mostly due to the fact that it hasn't been tested utilizing the aforesaid standards arsenic Fugaku and different supercomputers. 

"The Tesla Dojo is an AI-specific supercomputer designed to accelerate machine learning and heavy learning activities. Its little precision absorption limits applicability to a broader HPC context," Dekate said.

The measurements provided by Tesla bespeak that Dojo's awesome speeds were measured utilizing 3 standards: BF16, CFP8 and FP32, each of which bespeak the magnitude of bits that an equation occupies successful the computer's memory. 

SEE: Digital transformation: A CXO's usher (free PDF) (TechRepublic)

"For the astir part, HPC applications trust connected higher-order precision (FP64) than the ones supported by Dojo, which is much designed for extreme-scale heavy learning and instrumentality learning tasks," Dekate said. 

All of this isn't to accidental that what Tesla has developed with Dojo isn't impressive: It could beryllium to beryllium an manufacture person successful instrumentality learning training. That said, calling it a supercomputer and claiming it volition interruption the exaflop obstruction whitethorn beryllium a much hard merchantability erstwhile everyone other is standing their systems utilizing standards that are doubly arsenic complicated. 

Data, Analytics and AI Newsletter

Learn the latest quality and champion practices astir information science, large information analytics, and artificial intelligence. Delivered Mondays

Sign up today

Also spot

Read Entire Article