NVIDIA의 GPU 및 AI 가속: 종합적인 개요

강력한 성능과 텐서 코어와 같은 특수 AI 가속 하드웨어의 통합으로 인해 NVIDIA의 GPU는 TPU와의 경쟁에도 불구하고 계속해서 우위를 점하고 있습니다.

### AI 가속화를 위한 NVIDIA의 전략: 텐서 코어 및 트랜스포머 엔진을 통한 GPU의 진화
* 🔮 엔비디아는 파스칼 아키텍처부터 AI 연산에 특화된 텐서 코어를 통합하여 GPU 성능을 향상시켜 왔으며, 현재 **행렬 연산을 위한 주의 메커니즘을 포함하는 트랜스포머 엔진**을 확장하고 있습니다.
* 💪 16비트, 8비트, 4비트 연산의 빠른 처리를 지원하도록 하드웨어를 설계함으로써 **NVIDIA는 다른 MPU의 등장에도 불구하고 성능 우위**를 유지할 수 있습니다.
* 🎮 NVIDIA의 AI 기반 업스케일링 기술인 **DLSS는 사실적인 그래픽과 더 높은 해상도를 구현**하여 게임용 GPU 채택의 핵심 동력이 되어 왔습니다.
* 🔄 소비자 그래픽 카드에 **레이 트레이싱과 DLSS를 통합**한 것은 더 많은 사용자에게 고급 AI 기능을 제공하려는 NVIDIA의 전략을 반영한 것입니다.
* 🔀 엔비디아의 미래 전략은 **트랜스포머 엔진을 통해 GPU와 MPU의 하이브리드 구조**를 만들고, **NV링크 및 멜라녹스와 함께 네트워킹 기능을 활용**하여 **완벽한 AI 패키지**를 제공하는 것입니다.
* 🚀 AI 시장은 아직 초기 단계에 있기 때문에 **NVIDIA는 지속적인 혁신과 AI 가속 기능 확장**을 통해 리더십을 유지하는 것을 목표로 하고 있습니다.

Basically, there are operations that perform 32-bit, 64-bit, or even 16-bit operations. At the same time, there are things that correspond to operations, such as this tensor core. This tensor is specialized in operations, specifically matrix operations or vector operations, so that when you throw a command at once, hardware supports processing it immediately. Isn't it strange that NVIDIA's GPU continues to be used, even though Google has designed TPUs up to the 6th generation now? No matter how efficiently you design something quickly, NVIDIA GPUs are still being used. Tesla is also bringing in non-GP GPUs for some reason. Hello, I'm Andae Engineering. NVIDIA has released a new next-generation server-type GPU by extending the platform of the Black. This is a GPU, but ultimately, it is used for AI in AI data centers.

Think of it as a processor, but now, if you look at newspaper articles, even in our country recently, there are talks about mergers like Samsung's acquisition of Arm, or other startups specializing in designing MPUs, such as Purley. When you look at those, you often hear discussions about how much better they are, several times or even tens of times better. So, the most common question I've received from people around me is whether the demand for MPUs like Nvidia's GPUs is decreasing because so many MPUs are being produced. Nvidia is still showing tremendous performance, though. Therefore, before diving deep into the technical aspects related to this, let's briefly grasp the concept first. Then, we will delve into the details later. I would like to start from here. Nvidia RTX.

The 20 series, so before the era of GPT, it's been several years already. Various discussions arose, including mining, and even the situation where gamers who wanted to play games couldn't buy NVIDIA graphics cards. At that time, NVIDIA generated a significant amount of revenue. So, in the case of graphics you are seeing now, such as the RTX 20 series, this is a graphics card called the 2060 Super. It can be seen as a graphics card that sold a lot for gaming purposes and mining purposes. But why did it sell well? It is necessary to think quietly about what has changed from before and why it sold so much. Of course, graphics cards, GPUs themselves, were used a lot for mining compared to CPUs, and there was innovation for gaming purposes. I have created a lot of content through this engineering channel.

This is a content related to DLSS and FSR. For those of you who are hearing about this for the first time, let me explain. Currently, NVIDIA's secret weapon, which is opening an eye-catching engineering channel, is the third or fourth video I have created. If the graphics card is well-made, it can enhance the graphic quality when we play games by combining AI. Therefore, DLSS stands for Deep Learning Super Sampling. As for AMD FSR, it stands for FidelityFX Super Resolution. These technologies play a role in increasing the resolution. You can find more detailed information by checking it out. Why did DLSS become famous? It is a concept of super sampling. For example, you can upscale from Full HD to 4K. When you see an example, it's like this. When producing graphics for gaming, you can do things like this in real-time.

We need to process this, but by using a certain technology called DLSS, it expresses much more realistically than the previous versions. It applies features such as ray tracing that shoots light, upscaling, increasing the frame rate, and even AI assistance. DLSS 3.5 has been continuously upgraded, and even with an initial DLSS, using the RTX 2060 series in the direction of increasing frame rates, even at the same 4K resolution, allows for more details to be preserved, doesn't it? By extensively learning AI game graphics, when such things are sufficiently supported in hardware, it means that by the inference of a specific AI model, it can be rendered at a much higher resolution.

DLSS technology was able to accurately handle pixels. So, what is needed to support this DLSS technology? What you are currently seeing is the G100 die shot, which contains the GPU corresponding to the Ampere A100. When you look at it combined with the Grasshopper, you can see parts labeled as SM. Streaming Multiprocessors, commonly known as CUDA cores, are units where GPUs can process things in parallel. If you look at it as processor units gathered by unit, you can consider it as GPU cores. If you zoom in a bit more, you will see a streaming multiprocessor with caches and schedulers inside. Internally, it handles 32-bit integer processing, 32-bit floating-point processing, 64-bit processing, and specially, there are tensor cores here.

It includes doing basic things, basically 32-bit, 64-bit, or 16-bit, such operations that correspond to operations, and at the same time, there are things like this tensor core. This tensor is specialized in operations, in other words, matrix operations or vector operations, so that when you throw a command at once, it can be processed immediately by the hardware to support fast processing. Isn't it a bit strange that NVIDIA's GPU accelerates AI, and in the case of GPUs like this, they handle various tasks such as gaming, AI processing, Bitcoin mining, and legal use, so they are GPUs for this purpose. I understood that they are making MPUs specialized for AI only. This hardware that handles matrix operations separately is already installed, so when you actually look at these operations, NVIDIA has been

TensorFlow uses specialized operators for matrix operations that are optimized for AGO, allowing you to maintain 16 bits in the case of pre-season and expand the range much wider. By adopting such capabilities, the focus is on designing something much more efficiently. When talking about the recent Blackwell GPU, it has been mentioned that the computing speed has significantly increased as it has risen based on the floating point 16, then up again from floating 8, and once more from floating 4. Although some argue that this calibration is not accurate, they are still designing hardware in advance to support these capabilities that can be run quickly at a very low level of pre-season.

What does this mean? In the case of MPUs, many applications are designed to be specialized for specific floating points for fast processing. However, NVIDIA has already implemented hardware for things like Port 16, 8, 4, and made it possible to operate without giving MPU a chance to enter. So, if you look closely at the tuning architecture that went into the RTX 20 series, you will see that the Turing architecture includes Turing SMI. Here, we have Tensor Cores. What are Tensor Cores? When playing ray tracing games, they trace back the reflected light to calculate how the light is reflected, showing a realistic mirror image, implementing shadows, and reflecting something in the water.

We have included a special module dedicated to ray tracing that even expresses the shadow effect. In addition to that, we have integrated AI inference such as DLSS to accelerate it even further. By incorporating Tensor Cores, we have designed an architecture that allows Tensor Cores to handle all the elements that used to be processed in floating point 16 8 4. So, we prepared the hardware in advance to enable Tensor Cores to handle everything. When you mention engineering rendering on Google, or just rendering, realistic 3D images are created. In the process of creating lifelike 3D images in games, there is an explanation about the rendering pipeline, how 3D scenes are created in games. Ultimately, it's about shading in 3D to create the image. Although we included ray tracing for that purpose, we also separately designed it with Tensor Cores.

In the Pascal architecture, this is the Pascal p00 era, and after v1, a 100 came out. So, during the time when shaders existed, separate cores for tensor core and ray tracing core were designed separately as hardware. This is why NVIDIA has been gradually expanding these aspects, not just increasing the number of GPU cores. As they progressed, in NVIDIA's case, by processing matrix operations back and forth in the backend, when you put it in the 00, the tensor floating 32-bit gradually widens the slope. So, in a way, not only did they continue to bring out the strengths of the GPU, but they also expanded a separate MPU area to enhance performance. Therefore, NVIDIA GPUs themselves, starting from the 00, have been speeding up by basically shading, then doing ray tracing, and even further accelerating through tensor cores.

You can consider it as an increase, so what you keep seeing are the calculations by ray tracing, used for gaming graphics. By embedding this ray tracing technique into consumer graphics cards, it supports processing these ray tracing effects quickly. In fact, there were conflicts among users when DLSS and AMD FSR were introduced in the past. Because AMD simply presented signal processing algorithms for upscaling, while NVIDIA integrated hardware to significantly increase prices before offering it to users, some expressed opinions like, What kind of technology is this? But at that time, if you only looked at the graphics cards,

When you consider Jensen Huang's thoughts on drawing the big picture in the AI world, it makes sense that AI inference is necessary in the server environment right now. If you think about it this way, protecting not only GPU but also MPU capacity while adopting this directly shows how NVIDIA has designed its core. So, you can understand how much NVIDIA has been drawing the big picture. Therefore, with the integration of Tensor Cores starting from Pascal, it gradually speeds up, and as it goes up to INT8, the bit number decreases, making it much faster. Going down to INT4 makes it even faster, so you can see the structure as getting faster. When you go down to INT8 and INT4 dedicatedly, the performance does not increase significantly when going to 32-bit floating point. By coming down specifically to INT8 and INT4, there is room to utilize Tensor Cores, so Microsoft inevitably has to buy into NVIDIA's overwhelming performance.

You have to buy Meta, and this is the result that Google has designed TPUs up to the 6th generation now, but still have to keep bringing NVIDIA GPUs, and this time, Tesla is also bringing in MB GPUs. The reason for doing this is to design sensor cores and TP designs on their own. Even if there are talks about this, it seems that they cannot give up NVIDIA because they show very powerful performance with these things. Moreover, they added a Transformer engine here. Not only Tensor Cores, but this Transformer engine includes attention, which is a very famous diagram now. This model architecture, the Transformer, is a multiple, so it's about matrix operations. But this part is specialized in Transformers. NVIDIA is already doing this, so in the past, for training purposes, they used to...

However, it is worth trying something more with inference, but Nvidia is defending well. So, if we actually compare A00 to H00, the die size of the chip is around 850mm squared, precisely one. With the current equipment, there is one large reticle size due to the one die. When it comes to black, beyond the reticle, it is structured to combine two chips. Furthermore, going beyond the architecture, in Pascal, there was only a shading architecture, but with Volta, they started incorporating tensor cores. As we move from Ampere to Hopper, we see an expansion to transformer engines. So, you can see that not only did the CUDA cores increase on their own, but the GPU cores have also increased. Additionally, the tensor cores, which are dedicated to AI operations, have also significantly expanded.

This is possible because NVIDIA continues to create such solutions. This means that in the future, NVIDIA's Transformer will have a very powerful performance. Of course, in addition to Intel's Xe-HPG, AMD's performance, like the MI300, was actually tested by Microsoft, and it turned out that the MI300's performance was the best. But the GPU performance alone is not the only important factor. NVIDIA is not only focusing on GPU performance but also on enhancing the Transformer engine and Tensor Core. By ultimately integrating a hybrid structure of GPU and MPU, NVIDIA is able to respond to the MPU camp while also independently possessing Mellanox, a subsidiary company. In addition, by having NVLink and Infinity Band, NVIDIA is comprehensively responding to networking, allowing them to create a complete AI package, which sets them apart from others.

Even if you try to collaborate, for example, even if Intel itself thinks that the GP from this side is good, if you want to hurt something through networking, you just need to buy MBD and connect everything, install MV link switches, configure everything in racks like that, and build a huge computer in the structure of MV 70, which can be seen as very fast. So, even Tesla quickly brings this and competes with Intel by quickly bringing it and competing with competitiveness. Since the AI market is still in its early stages, it can be understood that you need to take the lead now. Therefore, NVIDIA is approaching all these configurations in a comprehensive manner, and in fact, now, to call this a GPU is not correct because it is not for graphic processing but for AI acceleration, so you can interpret it as having everything in terms of AI acceleration.

In the future, if you are more interested in this, we will cover more in-depth content related to this. So far, we have discussed the basic aspects.

NVIDIA GPU는 수 년 전부터 NPU까지 고려하여 설계되었다...!  빅테크 자체칩 설계에도 NVIDIA가 여전히 강력할 수 밖에 없었던 이유

공학박사 '에러'가 전하는
최신 공학/테크/IT/신기술

문의 : Unrealtech2021@gmail.com

안될공학 - IT 테크 신기술

기술 산업에서 NVIDIA의 GPU 및 MPU가 미치는 영향

여러 세대의 MPU가 개발되었음에도 불구하고 NVIDIA의 GPU는 여전히 널리 사용되고 있으며, 이에 따라 NVIDIA 제품에 대한 수요에 대한 의문이 제기되고 있습니다.

2060 슈퍼를 포함한 엔비디아의 RTX 20 시리즈는 게임과 채굴에 사용되며 게임용 혁신으로 인해 상당한 판매량을 기록했습니다.

NVIDIA의 비밀 무기인 DLSS와 FSR은 AI 기반 슈퍼 샘플링을 통해 그래픽 품질을 향상시킴으로써 회사의 성공에 기여했습니다.

DLSS 기술의 진화와 NVIDIA의 GPU 기능

DLSS 기술은 그래픽 렌더링, 업스케일링 및 프레임 속도를 개선하기 위해 지속적으로 업그레이드되어 왔으며, 특히 RTX 2060 시리즈에서 더욱 향상되었습니다.

Ampere A100과 같은 NVIDIA GPU에는 스트리밍 멀티프로세서(SM)와 텐서 코어가 장착되어 있어 AI, 매트릭스 및 벡터 연산을 위한 병렬 처리 및 특수 연산을 처리합니다.

이 GPU는 효율성과 고속 처리에 중점을 두고 게임, AI 처리 및 특수 작업을 지원하도록 설계되었습니다.

최근 블랙웰 GPU는 정확성에 대한 일부 우려에도 불구하고 부동 소수점 16, 8, 4 연산을 중심으로 컴퓨팅 속도를 크게 향상시켰습니다.

NVIDIA GPU 아키텍처에서 텐서 코어의 역할

NVIDIA의 GPU 아키텍처, 특히 Turing SMI에는 게임에서 사실적인 3D 이미지 렌더링을 위한 레이 트레이싱을 처리하는 전용 텐서 코어가 포함되어 있습니다.

이러한 텐서 코어는 점차적으로 확장되어 GPU에 통합되어 성능을 향상시킴으로써 레이 트레이싱 효과를 빠르게 처리할 수 있게 되었습니다.

DLSS용 하드웨어를 통합함으로써 NVIDIA는 가격이 크게 상승했다는 비판에 직면했지만, 궁극적으로 소비자 그래픽 카드에서 빠른 레이 트레이싱 처리를 지원하는 기술을 제공하게 되었습니다.

서버 환경에서의 AI 추론에 초점을 맞춘 NVIDIA의 설계는 파스칼에서 INT4로 점차 속도를 높이며 압도적인 성능을 제공하여 Microsoft의 채택을 필연적으로 이끌어냈습니다.

파스칼의 강력한 성능과 트랜스포머 엔진은 GPU 코어와 AI 전용 작업을 지속적으로 확장하는 등 AI 기술을 위한 중요한 선택으로 주목받고 있습니다.

파스칼에서 호퍼로의 진화는 텐서 코어의 대폭적인 확장과 트랜스포머 엔진의 통합을 보여주며 AI 기술에서 NVIDIA의 입지를 공고히 하고 있습니다.

AI 가속화를 위한 NVIDIA의 포괄적인 접근 방식

엔비디아는 GPU 성능뿐 아니라 트랜스포머 엔진과 텐서 코어에 중점을 두고 강력한 솔루션을 개발하는 데 주력하고 있습니다.

GPU와 MPU의 하이브리드 구조를 통합하고, 멜라녹스, NV링크, 인피니티 밴드를 보유함으로써 완벽한 AI 패키지를 제공한다는 점에서 차별화됩니다.

NVIDIA의 접근 방식은 단순한 GPU를 넘어 포괄적인 AI 가속화를 목표로 하며, 이를 통해 초기 단계의 AI 시장을 선도할 수 있습니다.