Pixel 6 AI explained | Lex Fridman

00:00:00.000 | Here's the new Pixel 6 Pro from Google.

00:00:03.280 | You're now seeing the result of it running a computationally intensive neural network

00:00:08.840 | in real time that I put on there for testing purposes.

00:00:12.440 | It's using TensorFlow Lite and the new Tensor chip that is optimized for AI.

00:00:19.040 | I am unboxing this AI because as a robotics and AI person, I'm interested in seeing how

00:00:26.320 | innovation in AI hardware and software is increasingly taking over the smartphone space.

00:00:33.600 | Also, unboxing AI makes me think of Pandora's box, the myth that serves as the metaphor

00:00:39.440 | for the mystery and the power of AI.

00:00:41.960 | I know this is just a phone, but let's pause for a second to think.

00:00:46.760 | This computer has over 200,000 times the processing power of the computer that first landed humans

00:00:52.460 | on the moon and over 2 million times the RAM.

00:00:56.180 | We are engineering our way to superhuman intelligence one phone at a time.

00:01:02.120 | One small step for phone and soon enough, one giant leap for a hybrid of machine and

00:01:12.160 | mankind.

00:01:14.360 | Let's talk about the specs.

00:01:16.000 | Here's the comparison of the Pixel 6 to the Pixel 6 Pro across various specs, highlighting

00:01:21.520 | what to me are the key differences in yellow and in green, what are the key similarities.

00:01:27.320 | Key differences, $300 in price.

00:01:30.400 | Also the Pixel 6 Pro has a slightly higher display, slightly higher resolution and a

00:01:34.800 | 120 hertz refresh rate versus the 90 hertz refresh rate.

00:01:39.080 | Though honestly, I've been using both phones for almost a week now and I don't feel any

00:01:43.800 | difference between them.

00:01:45.240 | The key similarity to me on the AI and the computational side is that they both have

00:01:49.800 | the same SoC system on a chip, the Google Tensor.

00:01:54.760 | To me, the RAM and storage are very important, but once you get to a certain threshold, it

00:01:58.720 | really doesn't matter.

00:01:59.720 | 12 gigabytes feels the same as 8 gigabytes and the same goes for the storage.

00:02:04.680 | Both phones have a 50 megapixel wide angle.

00:02:08.720 | Now that produces a 12 megapixel image because Google combines the 2x2 pixel groups into

00:02:14.720 | single effective pixel to cut noise and improve color and dynamic range.

00:02:19.120 | There's also the 12 megapixel ultra wide that's used to decrease noise for both the photos

00:02:24.280 | and the videos.

00:02:25.880 | And the Pro has a 48 megapixel telephoto lens, which results in one key difference, I think.

00:02:32.040 | It provides a 4x optical zoom plus a 20x digital zoom via machine learning via their super

00:02:39.200 | resolution algorithm.

00:02:41.160 | The rest to me is pretty much the same.

00:02:43.720 | Both phones feel amazing in my hand, but of course, me being who I am, I care mostly about

00:02:48.500 | what's on the inside, and that's the Google Tensor.

00:02:52.480 | Now let's talk about the Tensor chip.

00:02:54.600 | My main flagship phone for machine learning applications this year has been the Samsung

00:02:58.920 | Galaxy S21 Ultra 5G, pictured here with the Pixel 6 and the Pixel 6 Pro.

00:03:05.760 | The Galaxy Brain is powered by Snapdragon 888.

00:03:09.400 | The Pixel Brain is powered by the new Tensor chip.

00:03:12.860 | Both are truly amazing machines.

00:03:15.240 | I think AI innovation in both hardware and software will be what matters in flagship

00:03:19.260 | smartphones over the next decade.

00:03:21.480 | This is where the battle is.

00:03:24.800 | Let's now look at the details of the technical specs of the Tensor system on a chip and also

00:03:30.440 | the philosophical vision behind its architecture.

00:03:33.960 | The key components of the architecture of the Tensor system on a chip are the CPUs,

00:03:38.760 | the GPU, ISP, TPU, Context Hub, and the Titan M2.

00:03:43.920 | Depending on the application, various components of this chip can be used at the same time,

00:03:47.840 | leading to what Google is calling heterogeneous computing.

00:03:51.280 | For the CPU, there's two big CPU cores with the Cortex-X1, there's two medium CPU cores

00:03:58.800 | with the A76, and there's four small CPU cores with the A55.

00:04:05.600 | This is in contrast with the most common design for the flagships, which is one big Cortex-X1

00:04:10.960 | and four and three medium A78 cores.

00:04:13.960 | It's funny that Google says that having one big CPU core is good for benchmarks but not

00:04:18.960 | good for the experience.

00:04:20.680 | It's funny because, as you'll see in the benchmarks, the Pixel 6 actually performs really well

00:04:25.520 | on the single core Geekbench 5 test.

00:04:29.600 | Performance-wise, the benefit of having two Cortex-X1s is that you can distribute a thermal

00:04:34.760 | budget across them so there's less overheating on intensive tasks like 4K 60fps video.

00:04:42.480 | So in initial tests, there's a lot less overheating, so you'll be able to shoot video for much

00:04:46.360 | longer.

00:04:47.360 | Besides the CPUs, there's the GPU, there's an upgrade in that.

00:04:51.720 | There's the ISP, Image Signal Processor, that's optimized for image and video processing in

00:04:57.280 | terms of machine learning.

00:04:58.520 | And then there's the more general machine learning engine, that's the Tensor Processing

00:05:02.040 | Unit.

00:05:03.040 | The Geekbench 6 Hub does ultra-low power ambient computing, and the Titan M2 does hardware

00:05:07.920 | security.

00:05:08.920 | Like I said, I've been using the Galaxy S21 Ultra with the Snapdragon 888 for many months

00:05:14.880 | now, and so it's nice to take a look at some benchmarks for the CPU, GPU, and NPU for these

00:05:19.680 | two flagship chips.

00:05:22.000 | Now the big caveat here, as you probably know, is that benchmarks often don't reflect real-life

00:05:26.080 | performance so arguably they don't actually matter.

00:05:29.520 | But the main takeaway story here is that these are both amazing chips.

00:05:33.680 | In Geekbench 5 CPU benchmark, Pixel 6 outperforms the Galaxy S21 on single-core tests, and the

00:05:42.520 | Galaxy S21 outperforms Pixel 6 on the multi-core tests.

00:05:46.520 | For the Geekbench machine learning benchmark, the Snapdragon wins on the CPU and the GPU,

00:05:52.840 | and the Pixel 6 wins on the NPU.

00:05:57.880 | The Geekbench machine learning benchmark, by the way, uses TensorFlow Lite.

00:06:01.660 | And there's also the AI Benchmark 4 that's specialized for machine learning.

00:06:05.080 | It runs a huge number of different neural networks on the devices, and there, once again,

00:06:10.480 | Pixel 6 far outperforms Galaxy S21.

00:06:12.920 | In fact, it leads every other smartphone on the current AI Benchmark 4 leaderboard.

00:06:19.320 | Again, I think the takeaway here in terms of benchmarks is that Pixel 6 does well on

00:06:24.360 | machine learning tasks, and Snapdragon does well on CPU/GPU-centric tasks.

00:06:29.420 | But they're both, again, incredible machines.

00:06:32.740 | I think the important thing here is what does this heterogeneous computing enable in terms

00:06:36.940 | of software features?

00:06:38.800 | And the Pixel 6 provides a huge number of seamlessly integrated machine learning algorithms,

00:06:43.720 | increasing the vibrancy of the color with the HDR+ for the images and HDR-Net for the

00:06:48.040 | video, improving the accuracy and the efficiency of the face detection, again, both for images

00:06:53.120 | and video.

00:06:54.680 | And then there's just a huge number of cool features like face unblur motion mode that

00:07:00.200 | adds blur to moving objects.

00:07:02.200 | There's the magic eraser that's actually shown here on screen, where you can select certain

00:07:06.560 | parts of the object.

00:07:07.880 | They can be removed and then intelligently filled based on what the background is.

00:07:12.360 | Then for images, there's real tone that's looking at skin color, making sure this shows

00:07:15.760 | up looking great on photos.

00:07:18.160 | Honestly, the video is where the fun is.

00:07:20.280 | Like I said, HDR-Net, that's an incredible use of neural networks.

00:07:24.360 | I actually personally think super resolution algorithms are one of the coolest applications

00:07:28.040 | of computer vision in terms of its maybe simplicity and usefulness and impact.

00:07:34.440 | And there's a huge number of applications outside of visual domain, so speech, automatic

00:07:38.760 | speech recognition.

00:07:40.560 | You're talking about deployment of state-of-the-art ASR algorithms that pay attention to context,

00:07:47.160 | pauses, is able to do noise removal.

00:07:50.880 | And on the language side, there's neural machine translation.

00:07:54.800 | Obviously Google is taking natural language NLP really seriously in both the textual domain

00:07:59.860 | in speech, that's audio, and again, back to images and video.

00:08:06.100 | This is incredible leveraging to do heterogeneous computing on AI hardware to enable all kinds

00:08:12.140 | of cool computational photography features.

00:08:14.560 | Okay, let's look at some takeaways.

00:08:17.040 | My two favorite AI chips for Android now are the Google Tensor and the Snapdragon 888.

00:08:23.080 | Time will tell which wins for which applications, but for now, competition in this space is

00:08:28.200 | great for everyone.

00:08:30.200 | If you want me to talk about other AI systems or about running machine learning code on

00:08:34.640 | this and on other phones, let me know.

00:08:37.560 | I'll close with a quote from Eliezer Yudkowsky.

00:08:41.200 | "By far, the greatest danger of AI is that people conclude too early that they understand

00:08:47.880 | it."

00:08:48.880 | Thanks for watching, and hope to see you next time.

00:08:51.480 | Bye.

00:08:51.980 | [END]

00:08:54.980 | Eliezer Yudkowsky, Google Tensor, and Snapdragon 888

00:08:57.980 | Page 2 of 10

00:08:58.980 | Page 2 of 10

00:09:00.980 | Page 2 of 10

00:09:02.980 | Page 2 of 10

00:09:03.980 | you

00:09:06.040 | you

00:09:08.100 | [BLANK_AUDIO]

Pixel 6 AI explained | Lex Fridman

Chapters