Intel, AMD, and Google Drop In on AI Acceleration Wave

January 03, 2024 by Aaron Carman

From commercial to cloud-based acceleration, the latest AI hardware helps designers build bigger, better AI models.

Many big tech companies are turning to dedicated AI acceleration to support the current and expected AI loads at both the data center and edge levels. While AI can certainly be deployed with traditional processors, dedicated hardware affords the scalability and performance necessary to develop more advanced AI models.


AI accelerators are available in many different architectures

AI accelerators are available in many different architectures and provide parallel processing to accelerate the training, testing, and deployment of complex models. Image used courtesy of Synopsys

Much like how application-specific circuits can accelerate compute-heavy tasks like Bitcoin mining, many companies, including AWS and Microsoft, have turned to custom silicon to support industrial-grade AI loads. In this article, we will take a closer look at three new AI accelerators from Intel, AMD, and Google to find out what makes each accelerator unique and how the broader trend toward AI-specific silicon could usher in a new age of computer intelligence.


Intel Drills Down on Core-Level AI Acceleration

First up, Intel recently launched its newest Xeon processors (code-named Emerald Rapids), which feature AI acceleration built into every core. As a result, Intel reports that the 5th-generation Xeon processors offer a 21% higher average performance and up to 42% higher performance on AI inference workloads than competitors. In addition, Intel claims the new generation cuts the total ownership costs by 77% following a five-year refresh cycle, making it a good candidate for system architects needing a boost in AI or HPCC performance for next-generation workloads.


5th-gen Intel Xeon processors

5th-gen Intel Xeon processors show improved general compute and AI-focused performance, giving new capabilities to edge and data center devices. Image used courtesy of Intel

The Xeon processor itself is not a standalone AI accelerator but rather is built to “address demanding end-to-end AI workloads before customers need to add discrete accelerators.” The improved AI performance afforded by core-level acceleration falls in line with Intel’s “AI Everywhere” launch since the processors can be used in both data center and edge devices. Designers can leverage the improved AI performance while maintaining a familiar ecosystem.

As a result, designers needing extra performance without the optimization of fully custom accelerators can make good use of Intel’s newest offering.


AMD Releases Discrete Accelerators

On the discrete side of AI hardware, AMD also recently announced the MI300 series of accelerators, bringing new improvements in memory bandwidth for data-heavy applications such as generative AI or large language models. The two new products, the MI300X and MI300A, each target unique applications that demand high levels of memory performance.

According to AMD, the MI300X accelerators offer 40% more compute units and up to 1.7x the memory bandwidth than previous generations, accelerating more complex and data-heavy applications. The MI300A accelerated processing unit (APU), on the other hand, delivers a combination of processing performance on AMD’s Zen 4 core architecture and HPC/AI performance.


AMD AI accelerators

AMD AI accelerators provide designers with the memory capacity and bandwidth required for complex, high-level AI models. Image used courtesy of AMD

Companies such as Microsoft, Dell, and HPE have already revealed AI-accelerated devices using the MI300 series, highlighting the growing need for dedicated AI accelerators.


Google Ups Cloud-Based AI Acceleration

For designers looking to leverage the benefits of AI acceleration without building a device from scratch, Google has recently shown off its latest AI model using custom Tensor Processing Units (TPUs). While Google TPUs are not commercially available, they address a key market for cloud-based AI solutions.


Google TPUs in a data center

Google TPUs in a data center provide software designers access to accelerated AI capabilities, enabling faster innovation and improved performance. Image used courtesy of Google

Although system architects certainly benefit from custom AI solutions, not all designers need commercial products to accelerate their work. As such, software-focused designers can simply leverage Google's acceleration to test, train, and deploy models virtually.

As a result, hardware and software designers alike can benefit from AI-focused innovation.


Setting New Limits

Despite the fact that each technology listed here targets a different market segment, each one illustrates the growing need for AI acceleration and AI-focused hardware. Especially as AI is considered a solution for more complex problems outside the engineering world, a shift toward dedicated acceleration hardware may be necessary to support the number of calculations required to train larger models.