Is Intel Making A GPU?

January 29, 2023

Is Intel Making A GPU?

Intel Enters the GPU Arena

In Q2 2022, Intel released its first series of discrete desktop GPUs named Arc. For the past two decades, Nvidia and AMD have been the discrete GPU market’s heavyweights, but Nvidia has always maintained at least 60% of the market share in the past eight years and even reached a peak of 86% in Q3 2022. On the other hand, AMD’s market share has shrunk from as high as 37.9% in Q2 2014 to only 10% in Q3 2022.

While this does not necessarily mean AMD has been doing poorly in the past eight years, since AMD’s revenue has grown by about four times since then, it means that Nvidia has expanded its reach to not only gamers, content creators, and crypto miners, but also machine learning researchers who require Nvidia cards’ unique Tensor Cores.

What are GPUs?

GPU stands for graphics processing unit, and GPUs are electronic circuits that can render graphics for display on devices by manipulating memory to accelerate digital image creation in a framebuffer, which is a portion of RAM containing a bitmap that drives a video display. GPUs were initially introduced in 1999 to the consumer market and are more popularly known for video creation and games.

Discrete vs. Integrated Graphics

CPUs, or central processing units, are responsible for executing processes necessary for the computer to function. Integrated graphics refers to the ability of a chip to perform the functions of both a CPU and a GPU. Many AMD CPUs do not contain integrated graphics, and only a select few of their models have them (AMD G-series CPUs). On the other hand, many Intel CPUs contain integrated graphics, and each desktop Intel CPU has a product line suffix: “K,” “F,” “KF,” or nothing.

K: CPU is unlocked for overclock and has integrated graphics
F: CPU is locked for overclock and has no integrated graphics
KF: CPU is unlocked for overclock and has no integrated graphics
Nothing: CPU is locked for overclock and has integrated graphics

Thus, if the CPU does not have integrated graphics, a discrete GPU is needed for a display on the electronic device with said CPU.

Discrete graphics is a GPU that is separate from the CPU. Discrete graphics cards have memory that is not shared with the CPU, known as VRAM. While discrete graphics cards consume significantly more power and generate more heat than their integrated counterparts, they can provide much higher performance because of their memory and power sources.

Nvidia’s Breakthrough in AI Hardware

Before the release of Nvidia’s Volta GPU microarchitecture in December 2017, Nvidia cards since 2007 contained only CUDA cores. CUDA is an abbreviation for Compute Unified Device Architecture, and CUDA is a software layer that gives the user direct access to a GPU’s architecture and parallel computation elements. Furthermore, CUDA is accessible because it works with programming languages such as C, C++, and Python. CUDA cores can execute calculations and accelerate computation by running processes in parallel.

Every Nvidia card released since the Volta series contains CUDA and Tensor cores. Tensor cores help run computations with tensors more quickly. Tensors are algebraic objects that describe physical properties and generalize scalars and vectors; a scalar is a zero-rank tensor, and a vector is a first-rank tensor.

While nearly all CPUs today can handle simple tensor math, more complicated matrix multiplications are computationally intensive and require additional computational power. Thus, Tensor Cores serve this purpose. Tensor math is critical in various subjects, including fluid mechanics and astrophysics. Most notable, however, is machine learning, specifically deep learning. TensorFlow, a deep learning framework developed by Google, was streamlined for Nvidia cards’ CUDA and Tensor Cores.

Although the uses mentioned above of Tensor Cores do not apply to the average graphics card buyer, Tensor Cores still play an essential role. For example, the Turing series of Nvidia GeForce cards have Tensor Cores for a feature called DLSS, or Deep Learning Super Sampling. DLSS uses the card’s Tensor Cores to render a frame at a lower resolution and increase the resolution at the end of the process to match the native screen resolution of a monitor.

AMD’s AI Plan

During AMD’s Financial Analyst Day 2022, AMD CEO Lisa Su laid out the company’s ambitions clearly to become a stronger competitor in the chip industry by focusing on making better and faster chips, building AMD’s AI footprint, and making more custom silicon. Regarding AI, AMD announced a plan to release a unified software interface for programming AI applications on various chips. In early 2022, AMD acquired Xilinx, an FPGA designer.

According to Xilinx, FPGAs are semiconductor devices that can be reprogrammed to desired application or functionality after manufacturing. AMD intends to use Xilinx’s AI engine for future Ryzen processors and Epyc CPUs, which are AMD’s CPU line for servers. AMD claims that users can run massive models on the AI engines embedded in the Ryzen CPUs that would have required a GPU with less power required. As a result, laptop users gain access to simple tools such as background noise suppression and background blur reduction with less power. At CES 2023 earlier this month, Su announced that all AMD’s devices, including GPUs, CPUs, and Adaptive Accelerators, will have AI capabilities.

Intel Gets Serious

President Barack Obama waves to workers during a tour of the microprocessor manufacturing facility at Intel Corporation in Hillsboro, Ore., Feb. 18, 2011. (Official White House Photo by Pete Souza)

In 2017, Intel recruited chip architect Raja Koduri from AMD’s graphics division at Radeon. Signaling Intel’s ambition to compete with Nvidia and AMD in high-end discrete graphics. After scrapping previous attempts to develop discrete GPUs, such as the Larrabee Project.

On June 14, 2022, Intel released its first discrete desktop GPU under the codename Alchemist. Intel’s first line of discrete GPUs, Arc, uses Intel’s new computer architecture, Intel Xe. The Xe GPU line’s target audience range from integrated to high-performance, gaming, and data center/computing.

Regarding deep learning, Xe GPUs become optimized for ML with the OpenVINO inference engine. With better cost-efficiency than other AI computing platforms. Furthermore, Intel added a technology called Deep Link to these GPUs, which allows intelligent power-sharing between the CPU and GPU to boost performance for ML tasks.

While Intel’s announcement was exciting for many, it came with a few issues and a massive loss for the company.

According to Jon Peddie Research, a leading graphics market analysis firm; Intel became estimated to have lost about $3.5 billion in discrete GPU development since its establishment in Q1 2021. Thus, speculation arose over whether Intel should axe its R&D department in discrete GPUs, AXG. As its GPUs were released late in a post-pandemic economy when the crash of cryptocurrencies heavily impacted GPU prices and sales.

However, in August 2022, Intel’s Tom Petersen stated:

“I’d just like to be clear: we’re not going anywhere!” As a response to hearsay on whether AXG will become shut down.

Official Release and Software Issues

In early 2022, Intel’s Arc Graphics cards had one big issue: the software. Intel tried to use software drivers from its integrated graphics products as the foundation for the drivers of the new discrete cards. The integrated graphics drivers bottlenecked the performance of the new cards. As a result the launches of Arc 750 and Arc 770 became delayed by several months.

Finally, in October 2022, the Arc 750 and Arc 770 officially became released. And received generally positive reviews because critics found that the Arc 770 outperformed Nvidia’s RTX 3060 on average during gaming. However, another issue arose: erratic performance. Users found that the cards’ performance in old games takes a huge hit, and according to Intel, the reason is that these games run on older versions of DirectX. Thus, Intel released a driver that increased average frame rates for games like League of Legends by up to 80%.

Besides the delay and software issues of Intel’s new cards, they generally received positive reviews and more anticipation for Intel’s next generation of discrete GPUs codenamed Battlemage, which uses Intel Xe 2. Furthermore, is due to become released in 2024. Furthermore, in Q3 2022, Intel grabbed some of AMD’s market share in discrete GPUs alongside Nvidia. Intel discrete GPUs hold 4% of the market share, a vast improvement from 0.4% the previous quarter.

Intel’s Poor Performance

It is no surprise that Intel has missed its earnings estimates recently after the semiconductor giant’s poor performance in the past few years. In Q4 2022, Intel reported a loss impacted by a slump in PC purchases and increased competition from industry rivals. Furthermore, Intel issued a gloomy outlook for Q1 2023.

According to research firm Gartner, PC shipments fell by 28.5% in Q4 2022. Due to consumer spending becoming slammed by rising interest rates. According to Intel CEO Pat Gelsinger; “Despite the economic and market headwinds, we continued to make good progress on our strategic transformation in Q4; including advancing our product roadmap and improving our operational structure and processes to drive efficiencies while delivering at the low-end of our guided range.” Thus, are the next couple of quarters’ earnings transitory? Or is Intel only continuing downhill? In other words, do we “trust the process?”