Brainstorm AI London 2024: The Importance Of Ai Chip Development

Fortune

Presenter: Ami Badani, Chief Marketing Officer, Arm

Transcript

00:00 (upbeat music)

00:02 - Hi there.

00:05 I think we're all here because we believe

00:08 that AI will have a profound impact,

00:10 probably the most profound impact

00:12 that will be in our lifetimes.

00:14 It'll be bigger than the internet,

00:16 bigger than the smartphone revolution,

00:19 bigger than decades of innovations prior,

00:22 like the invention of the light bulb.

00:24 But AI is only on its tipping point.

00:28 In fact, you've seen use cases today,

00:30 like sports, financial services, entertainment,

00:34 and we're only getting started.

00:37 If you think about how AI will become mainstream

00:40 over the next decade, you think about ChatGBT.

00:44 It's at 200 million users in just over a year.

00:49 That's the fastest growing application

00:51 that we have ever seen.

00:53 Secondly, you think about self-driving cars.

00:56 One in 10 cars by 2030 will be self-driving cars,

01:01 and that will all be because of artificial intelligence.

01:05 Third, 75% of office applications,

01:08 office productivity applications,

01:11 will incorporate some form of generative AI

01:14 in some way, shape, or form.

01:16 So as you can see, AI will become mainstream.

01:20 But we're on a collision course with reality,

01:23 and it's not sustainable.

01:26 There are two things that we need

01:27 to really advance artificial intelligence,

01:30 and that's compute and power.

01:33 Let's talk about compute.

01:35 AI models are more hungry, more data hungry,

01:39 than they've ever been before.

01:42 They're increasing in complexity,

01:44 they're increasing in size,

01:46 and they're increasing in scope.

01:48 And what that drives is the need for more compute chips.

01:52 In fact, if you recall, OpenAI had their latest Sora model.

01:57 That's their latest text to video model.

02:01 It takes 100,000 AI chips,

02:04 working at full compute capacity and full power consumption,

02:09 in order to train Sora.

02:11 That's a huge amount of compute chips.

02:15 And that's just one large language model of many.

02:19 Then you think about smaller language models

02:22 that are gonna grow in scale,

02:25 and be quantized and applied to things like smartphones,

02:29 your thermostat in your home, and other home appliances,

02:33 your PC laptops, your inference in the cloud.

02:38 And then you think about large language models,

02:41 small language models, and that's really driving

02:44 the demand for compute chips.

02:47 But it's not about driving the demand for compute chips,

02:50 it's really about power efficient compute chips.

02:54 And that's really gonna help us carry

02:56 the advancements of artificial intelligence.

02:59 Power efficiency is in ARM's DNA.

03:02 In fact, 99% of smartphones

03:06 use an ARM power efficient chip.

03:09 And it's the same design philosophy

03:12 that we're applying to the advancements

03:14 of artificial intelligence.

03:17 In fact, if you think about AI, it comes with a cost.

03:22 And that cost is unfortunately power.

03:27 In fact, large language models

03:29 have an insatiable demand for power.

03:32 So without the need to address power,

03:35 we won't have advancements in artificial intelligence.

03:39 And if you think about the future

03:42 of artificial intelligence,

03:44 it's all gonna be about power efficient chip designs.

03:48 Now, if you look at this graph,

03:50 what it really shows is data center power consumption.

03:54 Most AI training and large language models

03:58 are all done in the data center today.

04:00 And that's 2% of the global electricity consumption,

04:05 which is an enormous amount.

04:07 But if I talk, you know, I spoke about earlier

04:09 that AI will become mainstream.

04:12 And when AI becomes mainstream,

04:15 the biggest tax that we'll see is on our power grid.

04:19 And we won't be able to continue the advancements of AI

04:23 without addressing power.

04:25 In fact, by 2030, artificial intelligence

04:29 will have triple the load than what it looks like today.

04:34 To put this into perspective,

04:35 25% of the US power grid

04:39 will be because of artificial intelligence.

04:42 So we really need to address this power problem head on.

04:46 And if you look at chatGBT,

04:48 it requires 15 times more energy

04:52 than a traditional web search.

04:55 So that's a huge amount of energy.

04:57 And that's what Arm is really focused on,

04:59 is driving power efficient compute chips.

05:02 In fact, if you look at customers like Google,

05:05 Microsoft, Amazon, Oracle,

05:08 they are all designing power efficient chips

05:11 because they know they have to solve the power problem first

05:16 in order to solve and drive

05:18 the advancements of artificial intelligence.

05:21 So in order for companies to have a competitive advantage,

05:24 they really need to think about power efficiency

05:27 in solving the power problem

05:29 in order to advance artificial intelligence.

05:31 [END]

Category

Transcript

Recommended