battle between AMD and NVIDIA

Graphics processing units (GPUs) have become an essential part of modern computing, evolving from devices used to render graphics to critical components for AI, data science, and high-performance computing. GPUs are composed of multiple compute units, mostly simple arithmetic logic units (ALUs), often supplemented by specialized cores such as tensor and RT cores. These specialized cores enable complex equations involving vectors, matrices, and tensors to be processed in parallel. In this article, we examine the competition between AMD and NVIDIA, looking at why AMD has consistently lagged behind NVIDIA in GPU acceleration and the AI ​​race, and examining the historical, technological, and strategic factors that have shaped this rivalry.

Understanding GPUs and their components

To understand what's going on between AMD and NVIDIA, it's important to understand what a GPU is and how they work. GPUs outperform CPUs at parallel processing tasks thanks to their large number of ALUs and specialized cores.

ALUs and specialized cores

ALUs form the core of a GPU’s computing power, performing the basic operations needed to render images and process data. Modern GPUs also include Tensor Cores, which are optimized for AI and deep learning, and RT Cores, which improve real-time ray tracing capabilities, bringing realistic lighting and shadows to graphics.

Discrete Mathematics and Linear Algebra on GPUs

The operations performed by GPUs are deeply rooted in discrete mathematics and linear algebra. Vectors and matrices are fundamental structures in these fields, and tensors extend these concepts into higher dimensions. Using these mathematical principles, GPUs can perform the complex calculations needed to render graphics, simulate physical systems, and train AI models.

NVIDIA's Strategic Investments

NVIDIA's dominance in the GPU market is the result of strategic investments in SDKs, developer support, and hardware innovations that span a wide range of applications beyond gaming.

SDK and developer support

NVIDIA recognized early on that robust SDKs and extensive developer support would be critical to maintaining its leadership. Its CUDA platform, introduced in 2006, greatly simplified GPU programming, allowing developers to leverage the parallel processing power of GPUs for general-purpose computing. This strategic move positioned NVIDIA as a leader not only in gaming, but also in scientific computing, AI, and data analytics.

Hardware innovations

NVIDIA has continually pushed the boundaries with hardware advancements. The introduction of Tensor Cores in their Volta architecture in 2017 was a significant step in improving AI processing capabilities. These cores are designed to speed up matrix multiplications, which are necessary for training neural networks. RT Cores, introduced in the Turing architecture in 2018, brought real-time ray tracing to consumer graphics cards, setting a new standard for the realism of visual effects in games.

AMD's Missed Opportunities

While NVIDIA was blazing a trail ahead, AMD was struggling to keep up. Several factors contributed to AMD's lag in the GPU race, including a lack of focus on AI and limited developer support.

Ray tracing and AI support delay

AMD didn’t introduce hardware support for ray tracing and AI until the release of the RDNA 2 architecture in 2020. By then, NVIDIA had already solidified its position with the RTX series, which featured both Tensor and RT cores. AMD’s late entry into these critical technologies left them playing catch-up in markets where NVIDIA had already set the standard.

Market focus and strategic mistakes

AMD’s strategic focus has historically been on consumers, particularly the gaming and console markets. This approach has secured partnerships with major console makers like Sony and Microsoft, but has left AMD less prepared to compete in the emerging AI and cloud computing markets. Moreover, AMD’s mobile GPU division was sold to Qualcomm, which integrated it into its Snapdragon systems-on-a-chip as an integrated graphics unit.

The move reduced AMD's presence in the mobile market and allowed Qualcomm to dominate the mobile GPU market, surpassing even Apple in market share and revenue. It also further widened AMD's gap in the graphics computing sector, as the company's mobile graphics team left the company along with the rights to the technology.
At the same time, NVIDIA, despite unsuccessful attempts to independently enter the mobile market as a manufacturer of gaming consoles and NVIDIA-Shield TV set-top boxes, has succeeded as a manufacturer and supplier of its mobile SoCs to large corporate clients. Thus, Nvidia Tegra is the basis of 141 million Nintendo Switch consoles alone and hundreds of thousands of Tesla electric cars, as well as many other electronics.

Although in the case of Tesla, from 2021 onwards, new models will also started equipped with CPU and GPU from AMD. Elon also plans to switch completely to his Dojo AI chips for both training and use in electric cars, but so far, judging by the information available in the industry, this is the case when Elon, promising to turn the market upside down, creates a solution that is several generations behind the analogues available on the market. The jammed Hyperloop sends greetings.

The announcement of the development of the AI ​​chip was made in 2019, although the development is presented as independent from Tesla, according to rumors, not only former top AMD engineers, such as Jim Keller, head of the autopilot development department at Tesla, but also AMD itself are involved in the development directly.

AI Race: NVIDIA Clearly Leads

NVIDIA’s aggressive approach to AI has put them well ahead of AMD. Their GPUs are ubiquitous in data centers, powering everything from AI research to cloud services.

Tensor Cores and AI Frameworks

The introduction of Tensor Cores gave NVIDIA a significant advantage in AI. These cores are optimized for deep learning tasks, allowing neural networks to be trained faster and more efficiently. NVIDIA also developed comprehensive AI frameworks such as TensorFlow and PyTorch, which are now industry standards. By building an ecosystem around its hardware, NVIDIA ensured that developers would prefer their GPUs for AI applications.

Data Center Dominance

NVIDIA GPUs have become the backbone of data centers around the world. Their performance, coupled with robust software support, makes them the preferred choice for AI and high-performance computing workloads. AMD, on the other hand, has struggled to gain a foothold in this lucrative market. Their focus on consumer GPUs has left them less prepared to compete in data centers dominated by NVIDIA’s purpose-built solutions.

Ray Tracing: A Deep Dive

Ray tracing, despite being a modern standard in computer graphics, has its roots in the work of Albrecht Dürer. In his treatise “Underweysung der Messung” (1525), he first described the technique that is now known as “ray tracing”. This technique allows for the creation of precise images of 3D objects on 2D surfaces by following the paths of light rays from the eye to the objects and back.

Ray tracing is based on solving a rendering equation that models how light travels through a scene. The algorithm works by tracing rays from the viewer's eye to light sources, determining how the light interacts with surfaces along those paths. This interaction involves calculating reflections, refractions, and shadows, which is necessary to achieve photorealistic images.

Path Tracing

Path tracing is an advanced form of ray tracing that traces hundreds or thousands of rays through each pixel, following their paths through multiple reflections and refractions until they reach a light source. This method allows for more accurate modeling of how light scatters in a scene, providing a high level of photorealism.

Developed by Jim Kajiya in 1986, path tracing applies statistical techniques to solve the rendering equation, using Monte Carlo methods to select a controlled number of paths to the light source.

RT cores

NVIDIA RT Cores are specifically designed to accelerate the most computationally demanding parts of ray tracing. These cores improve the intersection of bounding box volumes (BVHs), which are tree-like structures that organize the spatial relationships between objects in a scene. By accelerating BVH intersection, RT Cores dramatically reduce the time it takes to calculate ray-object intersections, enabling real-time ray tracing in games and applications.

The Future of GPUs: Can AMD Catch Up?

While AMD has made significant strides in recent years, the gap between them and NVIDIA remains significant. AMD’s RDNA 2 architecture has brought them closer to ray tracing and gaming performance, but NVIDIA’s dominance in AI and the data center poses a significant challenge.

AMD's Prospects

To close the gap, AMD will need to invest heavily in AI and cloud computing technologies. Their recent acquisition of Xilinx, a leader in adaptive computing, could provide the needed boost. Xilinx’s expertise in FPGAs and adaptive SoCs could help AMD develop more competitive AI solutions. However, overcoming NVIDIA’s leadership will require not only technological advances, but also a strategic shift toward supporting developers and building an ecosystem. This has not been seen so far, with most machine learning frameworks and libraries showing better performance on NVIDIA GPUs or only supporting CUDA.
The advantage of Open Source drivers for Linux is also no longer a trump card, as something incredible has happened and in recent years NVIDIA has made giant leaps towards opening up some of the sources of its drivers for Linux.

NVIDIA's development paths

NVIDIA shows no signs of slowing down. Their continued investments in AI, machine learning, and data center technologies will likely ensure that they maintain their current leadership. Upcoming architectures and innovations will likely further solidify their dominance. Additionally, NVIDIA’s entry into the ARM-based processor market Grace indicates a strategic expansion that can integrate with their GPU offerings.
And there is already a partnership with Mediatek looming on the horizon, and the Tegra line is doing well.

Conclusion

The battle between AMD and NVIDIA in the GPU market is a complex story of strategic decisions, technology innovation, and market dynamics. NVIDIA’s early investments in AI and developer ecosystems have given them a significant advantage, while AMD’s consumer focus and delays in supporting key technologies have left them behind. While AMD has made significant progress, catching up to NVIDIA will require a coordinated effort and significant investments in AI and data center technologies. As demand for GPU acceleration continues to grow across industries, the competition between these two giants will undoubtedly shape the future of computing.

Well, if you are interested in what is now, and not what will be in the future, we at ITGLOBAL.COM can offer our service AI Cloud, which will suit any of your needs in training and launching neural networks. By the way, in our Netherlands branch we had an extremely interesting addition to the ranks of green workers AI Cloud, straight from Uncle Huang. But more about that in another post.