Nvidia CEO Jensen Huang announces new generation AI chips and software

At Nvidi­a’s devel­op­er con­fer­ence in San Jose on Mon­day, CEO Jensen Huang intro­duced the com­pa­ny’s new AI chips and soft­ware with remark­able improve­ment in per­for­mance and capability. 

The new gen­er­a­tion of AI graph­ics proces­sors is named Black­well, and the first chip in this line­up is des­ig­nat­ed as the ‘GB200’, which is sched­uled to be shipped lat­er this year. 

The Nvidia CEO announced that the GB200, with its two B200 graph­ics proces­sors and Arm-based cen­tral proces­sor, achieves 20 petaflops in AI per­for­mance, a huge improve­ment over the pre­vi­ous Hop­per H100, which man­aged only 4 petaflops. 

It’s a game-chang­er in AI com­put­ing if Huang is right. 

“Hop­per is fan­tas­tic, but we need big­ger GPUs,” CEO Huang said at the company’s devel­op­er con­fer­ence in California.

AI mod­els are get­ting big­ger and more intri­cate along with rapid advance­ments. There’s a high demand for robust hard­ware solu­tions due to the growth of cloud com­put­ing and big data ana­lyt­ics, and Nvidi­a’s Black­well chips aim to meet this demand effec­tive­ly, enabling AI com­pa­nies to train larg­er and more com­plex mod­els with ease.

The com­pa­ny has also intro­duced NIM (Nvidia Infer­ence Microser­vice), a rev­o­lu­tion­ary soft­ware solu­tion designed to sim­pli­fy the deploy­ment of AI mod­els. NIM sim­pli­fies the process of run­ning AI soft­ware on Nvidia GPUs, mak­ing it more acces­si­ble, even on old­er models. 

Nvidia exec­u­tives say that Nvidia is tran­si­tion­ing from sole­ly being a chip man­u­fac­tur­er to becom­ing a com­pre­hen­sive plat­form, sim­i­lar to Microsoft or Apple, where oth­er com­pa­nies can devel­op and deploy AI software.

“Blackwell’s not a chip, it’s the name of a plat­form,” Huang said.

“The sell­able com­mer­cial prod­uct was the GPU and the soft­ware was all to help peo­ple use the GPU in dif­fer­ent ways,” said Nvidia enter­prise VP Manu­vir Das in an inter­view. “Of course, we still do that. But what’s real­ly changed is, we real­ly have a com­mer­cial soft­ware busi­ness now.”

Das said Nvidia’s new soft­ware will make it eas­i­er to run pro­grams on any of Nvidia’s GPUs, even old­er ones that might be bet­ter suit­ed for deploy­ing but not build­ing AI.

“In my code, where I was call­ing into Ope­nAI, I will replace one line of code to point it to this NIM that I got from Nvidia instead,” Das added.

In the con­fer­ence, Nvidia also intro­duced a sys­tem capa­ble of han­dling a mas­sive 27-tril­lion-para­me­ter mod­el, sur­pass­ing even the largest exist­ing mod­els like GPT‑4, with 1.76 tril­lion para­me­ters. How­ev­er, the com­pa­ny did­n’t dis­close the cost, but ana­lysts esti­mate sim­i­lar sys­tems range from $25,000 to $200,000.

The goal is to encour­age cus­tomers to sub­scribe to Nvidia enter­prise at $4,500 per GPU per year for a license, accord­ing to Nvidia offi­cials. Nvidia col­lab­o­rates with AI com­pa­nies like Microsoft and Hug­ging Face to ensure their mod­els run smooth­ly on Nvidia chips. 

“It’s all made at home,” CEO Huang came up with a new way to intro­duce new sim­u­la­tions from the Omni­verse sys­tem, empha­siz­ing that the clips were gen­er­at­ed, not animated.

Accord­ing to Nvidia, devel­op­ers can deploy these mod­els on their servers or Nvidi­a’s cloud with­out much has­sle with NIM.

The icing on the cake, of course, was the demon­stra­tion of robot­ics and how Omni­verse can empow­er them. Along with the demon­stra­tion were two cute mini robots — one green and one orange — which did­n’t seem to be obey­ing the CEO very well.