Efficient General-Purpose-Computing in the World of GPUs and AI Accelerators

The need for computing continues to grow exponentially across applications ranging from edge IoT to data centers. This growth has resulted in a growing pressure to get maximum performance for every joule of energy and square millimeter of silicon area. Each year, the semiconductor market sets new benchmarks by moving to smaller semiconductor nodes, squeezing in more transistors per mm². With this increase in the number of transistors, the performance increase has been radical, but the energy efficiency improvements have largely just been a dividend of this semiconductor node reduction. As a result, our computing capabilities today are limited by the joules of energy available to power the processors.

Graphics processing units (GPUs) and artificial intelligence (AI) accelerators are typically heralded as a solution to this energy challenge because of their high computational capabilities. The reality is that fully-fledged computational applications are more complex and only partially addressed by the recent wave of AI-focused chips. We have major unaddressed gaps in the processor market.

The Promise – and Limitations – of GPUs and AI Accelerators

General-purpose processors – i.e., CPUs – have been used for decades as a versatile option capable of handling a wide variety of computing workloads. Over time, they have become faster, but continue to be very energy inefficient, performing tasks sequentially and wasting most of the energy fetching instructions from the memory and moving data around. To overcome this architectural challenge, in the last 10 years, the industry has widely adopted GPUs and AI accelerators as specialized alternatives. These processors are designed to perform massive parallel computations per watt of energy and are heavily used in large AI data centers and high-performance computing (HPC). 

Compared to traditional general-purpose processors, GPUs and accelerators are much more energy efficient for AI use cases, but they come with a major limitation. Their strength applies only to a subset of computing tasks:

• Dependent on structured parallelism: GPUs by design excel in processing large amounts of data that are highly structured and explicitly parallelized. They can churn through large amounts of parallelizable data at a small energy cost. Unfortunately, a significant portion of computing tasks in the real world are difficult or impossible to parallelize. Unlike AI data centers, most of the functions in real-time, physical environments at the edge (e.g., industrial automation & factories, infrastructure IoT, smart sensors, and autonomous vehicles) require general-purpose computing. On average 40 to 50% of the energy used by a processor is spent on irregular, unstructured tasks that are difficult to parallelize, making them a poor fit for GPUs and accelerators. If 50% of the computing tasks cannot be parallelized and made to fit the GPU, then even with infinite efficiency the maximum energy improvement a device can achieve is 2x; far from sufficient to make radical improvements to solve computing’s energy problem! 

• Sequential Task Support: Most edge devices operating in real-time conditions necessitate sequential processing of the data. For instance, sensors monitoring pressure in oil & gas pipelines or detecting machine vibrations in a factory must react to unpredictable environmental inputs and trigger actions in a specific order. These processes are inherently sequential, and GPUs and accelerators do not benefit them.
• Lack of Programmability: Modern computing and edge devices require versatility to support a variety of functions, from sensors to cameras and monitors. However, AI accelerators and application-specific integrated circuits (ASICs) are designed and trained to support only specific tasks, and usually singular functions. This rigidity limits their applicability and future adaptability, often necessitating the deployment of entirely new devices to accommodate evolving capabilities.

In summary, what the computing world needs are high-efficiency processors that can support a variety of workloads and the ability to be re-programmed. So far, the industry has over-corrected and adopted AI accelerators and GPUs as the main workhorse to solve the efficiency problem, but, in the process, it has limited itself in capabilities to specific tasks. This is where Efficient Computer addresses the gap in the market - addressing efficiency and programmability together.

Re-defining the New Possible of High-Efficiency and General-Purpose Computing

Efficient Computer has developed a new computer architecture and introduced the world’s most energy efficient processor, offering more than 166x higher performance per watt while supporting general-purpose computing. This far exceeds anything that exists in the market and unlocks new capabilities unimaginable today.

Unlike accelerators or GPUs that are restricted to specialized or parallel tasks, Efficient can support a variety of computing workloads while offering greater than two orders of magnitude higher efficiency.  Additionally, with a co-designed software compiler, Efficient offers an exceptional development experience using any high-level language or AI/ML frameworks that no other processor provides.

Efficient’s E1© chip is specifically targeted towards high-performance edge devices and effcc compiler and will unlock new capabilities not possible before with processors, GPUs, and accelerators today. Contact sales@efficient.computer or https://www.efficient.computer/contact for more information.

‍