Dojo appears good at first look, however its reported efficiency numbers and area of interest function place it exterior the ranks of true supercomputers.
Tesla’s AI Day occasion included the reveal of a number of new potential merchandise, and whereas bipedal robots are thrilling (if maybe a bit unrealistic), the actual information to observe is the reveal of Tesla’s new in-house designed supercomputer, referred to as Dojo.
To name Dojo a full-fledged supercomputer is a bit beneficiant, although: It hasn’t been totally assembled but, and its potential efficiency limits have but to be examined. What Tesla guarantees, although, is nothing in need of a supercomputing breakthrough.
Probably the most highly effective supercomputer on the earth, Fugaku, lives on the RIKEN Heart for Computational Science in Japan. At its examined restrict it’s able to 442,010 teraflops (TFLOP) per second, and theoretically it might carry out as much as 537,212 TFLOPs per second. Dojo, Tesla mentioned, might find yourself being able to breaking the exaflop barrier, one thing that no supercomputing firm, college or authorities has been able to doing.
SEE: Hiring Equipment: Video Recreation Programmer (TechRepublic Premium)
Placing that declare in perspective means understanding the size and capabilities of Dojo and different supercomputers.
First, Dojo is designed to do one explicit factor: Practice synthetic intelligence. Tesla is constructing Dojo to be used in-house, processing video information from the hundreds of thousands of Tesla automobiles on the street. Dojo is constructed on Tesla’s D1 chip, the second the corporate has designed. The chip is constructed utilizing seven-nanometer expertise and is independently able to 362 TFLOPs per second.
Dojo chips do not function individually, nonetheless, and the smallest unit Tesla has constructed is what it calls Dojo coaching tiles. These tiles are a connection of 500,000 nodes which are reportedly able to performing 9 petaflops per second (1 PFLOP = 1,000 TFLOPs). All of that unbelievable pace is obtainable in a tile lower than one cubic foot in dimension.
Tesla’s plans for Dojo tiles is to community them into bigger methods. Its first design objective is to construct a pc cupboard able to housing two trays, every containing six Dojo tiles. In that configuration, Tesla mentioned, it might have the ability to deal with 100+ PFLOPs per second. Past that, Tesla plans to construct what it calls an ExaPOD consisting of 10 Dojo cupboards that can have the ability to carry out 1.1 exaflops (1 EFLOP = 1,000 PFLOPs).
Fugaku, then again, takes up a complete room, and at its peak measured efficiency is able to 442 PFLOPs.
Again to that earlier perspective: Tesla believes that it is going to be able to doubling the efficiency of Fugaku with 422 fewer racks, and it is going to be ready to do this by subsequent yr regardless of solely having reached the milestone of constructing and testing a single tile.
The fact of Tesla’s Dojo claims
Dojo’s reported capabilities do not grant it true high-performance pc (HPC) standing, mentioned Gartner analysis vice chairman Chirag Dekate, largely as a result of it hasn’t been examined utilizing the identical requirements as Fugaku and different supercomputers.
“The Tesla Dojo is an AI-specific supercomputer designed to speed up machine studying and deep studying actions. Its decrease precision focus limits applicability to a broader HPC context,” Dekate mentioned.
The measurements supplied by Tesla point out that Dojo’s spectacular speeds had been measured utilizing three requirements: BF16, CFP8 and FP32, every of which point out the quantity of bits that an equation occupies within the pc’s reminiscence.
SEE: Digital transformation: A CXO’s information (free PDF) (TechRepublic)
“For probably the most half, HPC functions depend on higher-order precision (FP64) than those supported by Dojo, which is extra designed for extreme-scale deep studying and machine studying duties,” Dekate mentioned.
All of this is not to say that what Tesla has developed with Dojo is not spectacular: It might show to be an business chief in machine studying coaching. That mentioned, calling it a supercomputer and claiming it can break the exaflop barrier could also be a tougher promote when everybody else is score their methods utilizing requirements which are twice as sophisticated.