Nvidia CEO Says Google Is the Company's Only Customer Building Its Own Silicon At Scale

An anonymous reader quotes a report from CNBC: Nvidia's CEO, Jensen Huang, has reason to be concerned about other chipmakers, like AMD. But he's not worried about Nvidia's own big customers turning into competitors. Amazon, Facebook, Google and Tesla are among the companies that buy Nvidia's graphics cards and have kicked off chip-development projects. "There's really one I know of that have silicon that's really in production," Huang told CNBC in an interview on Thursday. That company would be Google, he said. "But our conversation with large customers is intensifying," Huang said. "We're talking to more large customers."

Google first announced its entrance into the data center AI chip-making world in 2016. As it came up with new versions, the web company pointed to performance advantages over graphics cards that were available at the time. Google hasn't started selling data center chips for training AI models to other companies, though. (Google has started offering various products that use its Edge tensor processing unit chips, but those chips aren't as powerful as the TPU chips for training AI models in Google's cloud.)

We'll see...

[Spock9999]

We'll see. AMD's GPU cards for servers and virtualization have been really gaining ground on Nvidia. And there's also a big difference: Nvidia not only charges for the card but also requires annual per-user licenses. AMD only charges for the card and has no per-user licensing. That will be attractive to enterprise and big-business clients.

Re:

[AHuxley]

The big buyers buy up how many GPU products every set of months per generation as cards?
Why do that if you can "rent" a super computer from Google?
No power, no down time connecting cards, no heat, cooling.
Staff with skills network to the CPU?GPU they need.
What could not be like using a real GPU card in a real computer in front of the skilled staff?

A set of chips 1/2 around the USA on a fast network is the same "hardware" right?

Re:

[AHuxley]

The interesting part is when the cloud GPU asks what math is needed and has a product ready per seat for that.
Then asked again and again if project needs have changed?
The GPU and CPU becomes a question of math and money per project, not the set of last gen cards in place over months.
Who wants to buy a GPU for months/a year and be stuck with that last generation for next months of projects?
The competition has changeable CPU/GPU count in the cloud that is new and ready for bigger projects.
Scale and pow

Re:

[Rockoon]

It all depends on how much work you actually need computed.

Renting 365 days per year is a bad idea. Renting for 60 days per year on the other hand....

Re:

[AHuxley]

The market and very skilled people will work it out :)
GPU cards go the way of peripheral controllers :)

Re:Tesla is in production already

[Namarrgon]

Perhaps you missed where the headline specified only NVIDIA's customers. Tesla is not a customer of NVIDIA (anymore).

Re:

[drinkypoo]

Perhaps you missed where the headline specified only NVIDIA's customers. Tesla is not a customer of NVIDIA (anymore).

Right, they were right up until they started shipping their own silicon. So his argument looks good until you look at the facts. While his claim is technically true, it's also deliberately misleading.

Maybe you should read ALL the summary?

[SuperKendall]

Perhaps you missed where the headline specified only NVIDIA's customers.

From THE SUMMARY

Amazon, Facebook, Google and Tesla are among the companies that buy Nvidia's graphics cards and have kicked off chip-development projects.

Retard.

Re:

[Namarrgon]

From THE SUMMARY

Hasn't been true for many months. Huang clearly isn't counting Tesla as a customer, because they have been shipping all models with their own hardware since April [electrek.co].

Retard.

Heh.

Re:

[link-error]

    He also missed AWS which builds it's own chips... at scale.

https://www.theregister.co.uk/... [theregister.co.uk]

  They are also an NVidia customer.

he should be worried

[WindBourne]

For other car makers to keep up, they will have to do similar things. And Nvidia is not even close to what Tesla is doing.

Uh, Apple

[david.emery]

Builds and delivers A LOT of silicon, and will build even more if it moves the Mac from Intel. It's really hard to look at the iPhone and the iPad and not argue this is "silicon at scale" even if it's not servers.

Re: Uh, Apple

[Cmdln Daco]

Lots of companies "build" a lot of silicon. Some aren't even almost entirely fabless like Apple. But this article is about Nvidia silicon. Otherwise, if it's just chip-count, some company making embedded controllers for appliances is the "top" because they have incredible volume.

Re:

[Rockoon]

Perhaps you missed a key fact. Nvidia doesnt have any fabs either.

when a CEO says stuff like this, time to worry

[Locutus]

Tesla has already built its own hardware and it's way faster than what NVidia had. Google is doing TPUs and AMD is starting to get there too.

But here is NVidia CEO telling the world things are fine and its customers are not doing their own hardware at the expense of NVidia.

Reality is probably just the opposite from what the CEO stated. Short.

LoB

Re:

[Rockoon]

To be fair, some of NVidia's GPUs are outstanding.

With its last two GPU reference releases, AMD has been "all over the place" with two different GPU solutions for two different kids of workloads, but it isnt a star at either of them. If money is no object then you dont choose AMD for your vector processing.

Money, however, changes the equation. AMD isnt trying to "innovate into new markets" like NVidia. AMD's offerings at their price point has buyers that already exist, in markets that already exist.