CUDA was a “guess” that prompted Nvidia’s inventory market worth to plummet by a superb 80 p.c. CEO Jensen Huang explains the way it took place – and why he caught to the technique anyway.
Nvidia CEO Jensen Huang not too long ago spoke at size in regards to the historical past of his firm in a podcast. Specifically, the CUDA platform launched in 2006 was mentioned, which in accordance with Huang nearly ruined Nvidia:
That [CUDA] was the primary strategic choice that got here closest to an existential menace.
Step one over 20 years in the past
The technical basis for CUDA was laid again in 2003, when Nvidia constructed IEEE-compatible 32-bit floating level calculations – also called FP32 – into its shader items.
This meant that scientific code that was really designed for CPUs might, in precept, additionally run on an Nvidia GPU.Researchers already knew the right way to make the most of this on the time, so CUDA (“Compute Unified Area Structure”) adopted as a logical step within the context of a fully-fledged structure.
A choice that halved Nvidia’s inventory market worth
Nevertheless, the concept of this know-how was not the actual drama, however the industrial implementation.
Huang made the momentous choice to deliver CUDA not solely to costly workstation GPUs, however to each single Geforce card. Even the most affordable gaming GPUs from Nvidia help CUDA.Pure PC players didn’t know what to do with it, so it was laborious to persuade them to pay a premium for an structure that was quite unimportant for his or her favourite pastime. Nevertheless, the manufacturing prices of a GPU elevated for the implementation of CUDA.
CUDA elevated our prices by round 50 p.c, and on the time we have been an organization with a gross margin of round 35 p.c. Our margin dropped by about one and a half {dollars} per chip.
The consequence on the inventory market was brutal: Nvidia’s market capitalization fell from round eight billion to only underneath 1.5 billion US {dollars} after the CUDA launch – downright ridiculous figures in comparison with right this moment’s inventory market worth.
The “CUDA guess” paid off
Nvidia’s CEO nonetheless felt it was important to deliver CUDA to prospects through Geforce graphics playing cards: If CUDA was to have an opportunity as a brand new computing structure, it needed to find yourself within the arms of as many individuals as doable.
In Huang’s eyes, the precept behind this technique is rapidly defined: “The set up base defines an structure. […] Every little thing else is secondary”.
It took a number of years for the tide to show. In 2012, the neural community “AlexNet” beat all opponents by greater than ten proportion factors within the ImageNet competitors.The underlying {hardware}? Nvidia graphics playing cards with CUDA – and all of the sudden everybody was speaking in regards to the structure.
Background:1 / 4 of a century in the past, a scholar related 32 Geforce graphics playing cards to play Quake 3. That is how CUDA was born
Trying again, Huang believes that Nvidia’s success is predicated on Geforce; in any case, it was these graphics playing cards that introduced CUDA “to everybody”.
No less than the present market scenario proves him proper, as CUDA is by far the dominant platform for AI coaching and inference – and is not the face of a self-destructive Nvidia guess, however one of the crucial precious tech firms of our time.
