인텔의 새로운 AI 칩: 모델 트레이닝의 판도를 바꾸는 칩

### 인텔의 새로운 제온 프로세서: AI 개발을 위한 게임 체인저
* 🚀 인텔은 제온 프로세서와 AI 가속기를 출시하여 **라마 2 벤치마크에서 50% 향상된 속도와 비용 효율성**을 보여주는 인상적인 성능으로 엔비디아에 대한 우려를 불러일으켰습니다.
* 💡 벤치마크 결과 **새로운 인텔 칩이 엔비디아의 제품을 능가하는 것**으로 나타나, 엔비디다의 CUDA 독점 체제를 넘어서는 보다 다양한 에코시스템의 필요성이 강조되고 있습니다.
* 🖥 이 새로운 아키텍처를 통해 개발자는 **기존 GPU 코드베이스를 인텔 플랫폼으로 쉽게 마이그레이션**할 수 있으며, **CUDA 라이선스 비용에 대한 부담 없이 PyTorch 및 TensorFlow와 같은 인기 라이브러리**를 지원할 수 있습니다.
* 📊 인텔 프로세서는 뛰어난 부동 소수점 연산 속도를 자랑하며 **부동 소수점 8 모델의 경우 2배, 부동 소수점 16 모델의 경우 4배 더 빠른 성능**을 달성하여 AI 학습 기능을 향상시킵니다.
* 🌐 새로운 AI 가속기는 **텍스트에서 동영상 생성에 이르기까지 다양한 데이터 양식**을 처리하도록 설계되어 개발자에게 다용도 옵션이 될 수 있습니다.
* 🛠 이 동영상에서는 비용 효율적이고 효율적인 학습을 위해 새로운 인텔 칩을 활용하여 **개인 YouTube 콘텐츠에 맞게 미세 조정된 Meta의 Llama 3 모델**을 사용하여 개인 AI 비서를 배포하는 데 초점을 맞추고 있습니다.
* ☁ 이 프로세스에는 **인텔 프로세서 전용 AWS 인스턴스**를 설정하고 최적의 성능을 위해 적합한 **AMI와 지역**을 선택하는 작업이 포함됩니다.
* 💻 최종 구현은 미세 조정 프로세스를 자동화하는 **단일 Python 스크립트**를 통해 간소화되므로 사용자는 최소한의 비용으로 로컬에서 AI 모델을 실행할 수 있습니다.

인텔의 새로운 AI 칩은 모델 학습에 있어 50% 더 빠르고 저렴하다는 점에서 엔비디아의 제품보다 성능이 뛰어납니다.

hello world. It's SJ, and Nvidia is terrified of this new AI chip that Intel just released. It's called the Xeon Processor and the associated AI accelerator. And the reason that Nvidia is so afraid is because when tested out on Llama 2 benchmarks, it performed 50% faster and 50% cheaper when it came to training these models. And the source of that information is St il. A. We can see those benchmarks right here, where they tested it out on a bunch of different multimodal diffusion transformers, and they found that it was 50% faster and 50% cheaper. Nvidia is terrified of that because the whole ecosystem is locked in with Cuda, and there needs to be less of a monopoly. There needs to be more options in the environment. There needs to be more of an availability for developers to build on different chips. Intel has just released that chip, and in this video, we're going to talk about that chip. I'm going to explain a little bit about how it works, and then we're going to deploy our own AI model to our

own infrastructure. Our own AWS infrastructure that allows us to create our own personal assistant. And what we're going to use for that AI is Meta's Llama 3 model, which they recently released a few months ago, and it is still a state-of-the-art in open-source foundation models. And we're going to fine-tune this Llama 3 model on my YouTube videos to make it into a personal assistant that I can use. And we're going to deploy it to my AWS account. And we're going to use all my YouTube videos as training data for that. And all this is going to be trained using the new Intel chip, the new GTI processor. And it's going to be cheaper, and it's going to be faster than NVIDIA. So, let's talk a little bit about this chip that Intel just released. So, H100s are in very, very high demand right now. Everybody wants them. Zuckerberg's, you know, building a million of them. Everybody's building them, and they're in short supply. So, this is a

good alternative to an H100. We can use an AI accelerator as a GPU alternative, and they work with the CPU, which is called Xeon. It's the new generation of Xeon that works together, so no CUDA, no CUDA licensing costs. If you have an existing GPU codebase, you can migrate it in a few lines of code. Py supports it, TensorFlow. We'll talk a little about the libraries here, but essentially, it's very easy to switch between this Intel GAI accelerator and your existing GPU workflow. And it works with one node or thousands of nodes, so it scales, and it's price performant with the H100. It's actually better in terms of cheaper when it comes to training Llama 2 models and Llama 3, which we're going to do in this video. Intel really has a long history of—they literally invented Moore's law, right? Which is the law that computing power doubles every year. And in this case, you can see that floating-point operations, which are used in AI matrix operations.

are two times faster for floating point 8 models and then four times faster for floating Point 16 models. What does that mean? That means different ranges of compute density. How much information are we packing into this stream of neural network processing chipset that we're using? And the most important part here is that it's freedom to scale without vendor lock-in. And Nvidia really has monopolized this market, and we really need to break up the monopoly and allow for a democratized AI ecosystem. And whatever type of application that we're building with this, it's going to work with this system, whether it's 3D Generation, video generation, text, image, audio, any data modality is going to work on the Intel chipset just like it does on the Nvidia chipset. And it's going to be faster, and it's going to be cheaper. This architecture of this new accelerator, it has many more tensor cores, and tensor cores are specifically designed to process the type of computing operation that neural networks use, and that is in a little wrap input times weight at a bias activate.

very simple how it works. This chip is designed for that. So, that's a little bit about the chip. It runs any type of AI model. It runs PyTorch. Let's get right into building something with it because I like doing actionable things with information. So, we know that Intel just released God3. It's a great GPU alternative accelerator. We don't have to use CUDA. It's cheaper. Great. Okay. What do we do with this? The next step is for us to train our model on the cloud. So, let's find a suitable cloud provider, and I'm going to set a timer here to do this in less than 5 minutes. And the first one that you might want to look at is Intel's own developer cloud. And what I've done is I've requested an instance, and it's pending review because I just did it today. But they're pretty fast with acceptance, so you should submit here to try on Body 3. But for now, since we don't have access to the latest preview version on the developer Cloud, we're going to use AWS directly. So, we're

going to run this AWS instance. This is still a way for us to try out the model on Intel hardware and avoid Nvidia for now to see what it's like. So let's get right into how we're going to do this. Okay, so the first thing we're going to do is we're going to launch our new instance, and that instance is going to be dedicated to this Intel processor and to fine-tuning llama 3 on my YouTube videos. So we're going to call it llama 3 finetune intel, and I'm going to now find that in my AWS dashboard under EC2. What we are creating here is a new computing instance, and a lot of cloud providers are trying to make this easier for you, like Vercel and others, but really, we all want to use AWS. We should all be using AWS because it's cheaper, and it's really the source of all these other startups. Heroku comes to mind, and I know it's easier for beginners to start off using other services, but ultimately, we should be using AWS directly. So once we create

a new instance. It's going to ask us what the machine image is, and what I'm going to do is I'm going to try to find that instance type, and I need to be in the right region. I need to be in Us East one to do this properly, so I'm going to find that instance, and I've found the Ami image that I need in the right region. You have to always remember that the region is very important to AWS. If you don't have the right region, then you won't get the right Ami, and I found a bunch of great Amis here. I'm going to use this one, which uses the god, and I'm going to subscribe now to this instance, and that is going to create a new version of that. It's connected to it. I'm going to launch the instance. It's going to create a new key pair, and that is going to authenticate me to this instance, and I'm going to actually proceed without a key pair, and that is now creating my Intel instance, and now

once we have that instance, we can run our code to finetune llama 3 on my YouTube videos. And how do we do that? I asked Chad gb01 preview, of course, to do all that in a single class file of Python. It was able to do that in this file right here. So, it's using the Hugging Face library to fine-tune llama 3 on my YouTube videos by just feeding it a single Google Drive link. And then, it's going to fine-tune on all that data by transcribing it to txt and then using that as a fine-tuning trainer object in the Hugging Face library to create this final version. And all of this is done in a single Python file, and we can run this inside the browser without any worries. Let me just show you exactly how to do that. So, if we go to [Music] instances, if we go to connect in our instance state, we can then connect to that instance in the browser. And once we've connected inside of the browser, then we can just paste in our Python

file and run it. So, we're going to type in Nano, and then we're going to call this script finetuner dopy. And we will copy the code for that into the new file. And now we can run it, and that's now running on our new instance. And it's going to install all the necessary packages that are necessary, and it'll run that script all in one go, thanks to Chad gb01 previews amazing ability to create working python code in one shot. And that lets us train our own AI model, and we can run this locally. We host it on our own instance, and this is very cheap, actually, right? A few bucks a month, and we have our own AI fine-tuned on our own data. Very useful stuff. Thank you for watching this video. Thank you, Intel, for sponsoring this video. I'm proud to be an Intel Ambassador. Check out helpful links in the video description. If you like the video, smash the like button. It really helps promote the video. Hit the Subscribe button, and for now, I got to

keep training my AI model, so thanks for watching. [Music]

Why NVIDIA is Terrified of This New AI Chip

Hello World, it's Siraj!  I'm a technologist on a mission to spread data literacy. Artificial Intelligence, Mathematics, Science, Technology, I simplify these topics to help you understand how they work. Using this knowledge you can build wealth and live a happier, more meaningful life. I live to serve this community. We are the fastest growing AI community in the world! 

Join my AI Sports Betting bot community: https://www.wagergpt.co 

Twitter: http://www.twitter.com/sirajraval
Instagram: https://instagram.com/sirajraval/
Facebook: https://www.facebook.com/sirajology/

If you found my videos useful, join my community: https://www.chatgptschool.io

Some Research -https://drive.google.com/file/d/0BwUv84lNDk72Q1gzaXgwR2U3U2NWVlZSOFk4amZIRmV1QXI0/view?usp=sharing&resourcekey=0-WgUkgIPZd-qCklnTUElNlg 

In the event of my demise, you must finish what I've started here.

Siraj Raval

인텔이 새로 출시한 제온 프로세서와 AI 가속기는 라마 2 벤치마크에서 50% 더 빠르고 저렴한 모델 트레이닝을 제공하면서 엔비디아의 우려를 불러일으키고 있습니다.

이 칩의 성능은 엔비디아의 독주에 도전하며 개발자에게 더 많은 옵션을 제공하는 것을 목표로 합니다.

이 비디오에서는 이 칩의 기능을 살펴보고 AWS 인프라에서 새로운 인텔 칩을 사용하여 Meta의 Llama 3 모델을 개인 비서로 배포하는 방법을 시연합니다.

차세대 제온 CPU와 함께 작동하는 AI 가속기는 기존 코드를 쉽게 마이그레이션할 수 있는 H100 GPU에 대한 비용 효율적인 대안을 제공합니다.

한 노드에서 수천 노드까지 확장 가능한 솔루션을 제공하며, 라마 2 및 라마 3 모델을 훈련하는 데 더 저렴합니다.

더 빠른 부동소수점 연산과 더 많은 텐서 코어를 갖춘 이 가속기는 AI 생태계의 민주화와 Nvidia의 시장 독점을 깨는 것을 목표로 합니다.

인텔의 새로운 칩으로 AI 모델 구축하기

인텔의 새로운 칩은 파이토치를 포함한 모든 AI 모델을 실행하도록 설계되었으며, CUDA를 사용하지 않고도 GPU를 대체할 수 있는 저렴한 대안을 제공합니다.

첫 번째 단계는 클라우드에서 모델을 훈련하는 것으로, 인텔의 개발자 클라우드와 AWS가 잠재적인 제공업체로 사용됩니다.

이 과정에는 모델 미세 조정을 위해 인텔 프로세서 전용의 새로운 AWS 인스턴스를 시작하는 것이 포함됩니다.

Llama 3 미세 조정을 위한 AWS 인스턴스 설정하기

이 프로세스는 올바른 인스턴스 유형과 지역을 선택하는 것으로 시작되며, 올바른 AMI 이미지를 위해 미국 동부 인스턴스에 중점을 둡니다.

인스턴스를 구독하고 실행한 후에는 인증을 위한 키 쌍이 생성되지만, 키 쌍 없이 진행할 수도 있습니다.

마지막 설정은 Hugging Face 라이브러리를 사용하여 Python 스크립트를 실행하여 YouTube 동영상에서 라마 3를 미세 조정하는 것으로, 전체 프로세스는 브라우저에서 관리할 수 있습니다.

Nano를 사용하여 스크립트를 생성하여 필요한 패키지를 설치하고 AI 모델 학습을 위한 Python 코드를 실행합니다.

이 과정은 한 달에 몇 달러에 불과한 개인 인스턴스에서 호스팅할 수 있어 비용 효율적입니다.

이 동영상은 인텔의 후원을 받았으며, 발표자는 시청자에게 좋아요를 누르고 더 많은 콘텐츠를 구독하도록 권장합니다.