up to 100 times faster

As the technology industry seems to be reaching the limits of Moore’s Law, a daring Finnish company is proposing a solution that could well reignite the race for computing power by multiplying by 100 the performance of central processing units in data processing.

In recent years, artificial intelligence accelerator chips have captured the attention of technophiles, thanks in particular to the American company Nvidia to name just the most famous. On the contrary, central processing units (CPUs) have been gradually relegated to the background.

La Start-up Flow Computingwith its Parallel Processing Unit (PPU) technology, aims to change the game and radically transform the architecture of CPUs to put them back at the heart of computer systems. It tackles the challenge of optimizing processors from a completely new angle.

In short, the PPU works in tandem with the central processor and is able to handle many tasks simultaneously, providing advantages in various fields such as artificial intelligence, scientific computing, and virtual reality.

An Overview of the Parallel Processing Unit (PPU)

The Flow Computing approach is therefore based on the introduction of a parallel processing unit (PPU). Instead of packing 16 identical CPU cores into a laptop, manufacturers could opt for a configuration that combines 4 standard CPU cores and 64 PPU cores. This configuration could achieve up to 100 times the performance while maintaining the same hardware footprint.

The PPU thus offers significant acceleration for parallelizable tasks, bridging the gap between traditional CPUs and specialized GPUs. Professor Jörg Keller, a parallelism expert at theUniversity of Hagen in Germany, stressed the importance of this innovation: ” This is a game changer for smaller workloads, meaning parallelization can be applied to more places in the code. »

An architecture designed for versatility

Martti Forsell, CTO and co-founder of Flow Computing, said computing tasks typically fall into two categories: sequential and parallel. His company’s proposed architecture aims to Optimize the processing of both types of tasks within the same system. And to add: ” When a sequential workload is part of the code, the CPU part will execute it. And when it comes to parallel parts, the CPU will assign them to the PPU. This way we get the best of both worlds.. »

The Parallel Processing Unit (PPU) is an electronic component designed to work alongside a computer’s central processing unit (CPU). Its unique feature is its ability to handle multiple tasks simultaneously, whereas a traditional processor typically performs them one after the other.

The PPU stands out for its flexibility. As a Flow Computing spokesperson explains: “ Our parametric design allows for extensive customization, including the number of PPU cores, the variety and number of functional units, and the size of on-chip memory resources. »

This adaptability allows the PPU to adjust to different needs:

• The number of computing cores can vary from 4 to 256 or more
• The types of calculation units (arithmetic, floating point, etc.) are adaptable
• The built-in memory size is adjustable

Flow : SuperCPU

Advantages of PPU over conventional processors

The Flow Computing team identified four fundamental requirements for a computing architecture optimized for parallelism: tolerance to memory latency, sufficient bandwidth for communication between threads, efficient synchronization, and low-level parallelism.

To meet these requirements, the company has developed an entirely new architecturespecifically adapted to parallel computing. In addition, the PPU brings three major improvements over conventional processors:

1. Reduced waiting times

In a conventional processor, memory access can slow down the execution of tasks. The PPU solves this problem by executing other tasks during memory accesses, thus avoiding idle time.

2. Better synchronization

Current multi-core processors waste time synchronizing their different units. The PPU significantly reduces this synchronization time, allowing for smoother execution of parallel tasks.

3. Optimized execution of instructions

The PPU organizes its computational units in a chain, allowing each unit to directly use the results of the previous unit. This organization eliminates the slowdowns related to dependencies between instructions, common in conventional processors.

Encouraging results

Initial tests by Flow Computing have shown encouraging results. Martti Forsell added: “ Up to 100x improvement was achieved in our preliminary performance comparisons, assuming there would be a silicon implementation of a Flow PPU running at the same speed as one of the commercial processors being compared and using our microarchitecture. »

The team is currently developing a compiler for their PPU and is looking for partners in the field of CPU production. Their technology can be implemented with any instruction set architecture, thus offering potential for improvement for all existing CPU types.

Professor Keller concluded by emphasizing: “ The time has really come for this technology to come to market. We now have the need for energy-efficient computing in mobile devices, and at the same time, we need high computing performance..

Concrete applications for users

Integrating PPU into our everyday electronic devices could radically transform their capabilities.

Smartphones would become more powerful, allowing advanced AI applications to be run locally without relying on remote servers. In the field of virtual reality, headsets would offer more immersive experiences thanks to more realistic and responsive environments. Personal computers would see their performance increase tenfold, significantly accelerating complex tasks such as video editing or 3D modeling. Finally, the safety of autonomous cars would be enhanced, with ultra-fast processing of sensor data allowing for almost instantaneous reaction to potential dangers.

« Flow Computing develops an artificial intelligence tool to help application and software developers identify parallel parts of code and propose methods to optimize them for maximum performance. ” said a representative from Flow Computing.

An impact on different sectors

The PPU could transform many fields with its advanced computing capabilities. In artificial intelligence, faster computing could allow models to be trained more quickly and at lower cost, paving the way for more frequent innovations. In scientific computing, complex numerical simulations, essential in meteorology or particle physics, would benefit from significantly faster execution, potentially accelerating discoveries.

In the defense domain, ultra-fast data processing would provide a crucial strategic advantage for modern military systems, improving decision-making and operational responsiveness.

Google recently highlighted the importance of such advances when it announced its first Arm-based processor in April 2024: “ While our investments in compute accelerators have transformed our customers’ capabilities, general-purpose computing is and will remain a critical part of our customers’ workloads. Analytics, insights, ML training and serving all require massive amounts of compute power. Customers and users looking to maximize performance, reduce infrastructure costs, and meet sustainability goals have found that the rate of improvement in processors has slowed recently. Amdahl’s Law suggests that as accelerators improve, general-purpose computing will dominate the cost and limit the capacity of our infrastructure unless we make commensurate investments to keep pace.. »

If Flow Computing’s PPU technology delivers, it could usher in a new era in computing. The electronic devices we use every day could become dramatically more powerful, leading to new applications and innovations.

A word from Jörg Keller on PPU technology

Professor Jörg Keller introduces the concept of the Parallel Processing Unit (PPU) of Flow Computing and explains how it differs from traditional multi-core processors. Professor Keller highlights the limitations of current processors with independent cores and the overhead they incur when executing parallelizable code.

The PPU addresses these limitations by having a set of cores that work in unison, which significantly reduces overhead and enables efficient parallelization. Additionally, PPU cores are smaller and consume less power than traditional CPU cores, further enhancing their appeal.

Antti MÄKELÄ : Today we will delve into Flow Computing’s PPU technology. Jörg Keller, Professor at the Faculty of Mathematics and Computer Science at Fern University in Hagen, joins us. Professor Keller has carried out the technical verification of Flow’s PPU parallel processing unit.

Jörg KELLER : Today we have multi-core processors. These are processors with a larger number of cores in order to increase computing performance. For some time now, these computing cores have all been the same.

JK : Some time ago, manufacturers started introducing processors with different types of cores, like Arm’s “big little” technology. But all of these cores are independent of each other. Now, when we have code that can be parallelized, it would be nice to have a number of cores that are doing essentially the same thing and doing it without the overhead that we get when each core is rather independent of each other. This is where Flow technology and PPU cores come in.

JK : So, next to the standard CPU cores, we have a set of cores that are doing mostly the same thing at any given time, and can therefore speed up some code without each of them having to fetch its program for itself, because they are all doing the same thing. Also, they don’t need to maintain all the structures that normal CPU cores have to do to run a program. So we have a simple way to parallelize with less overhead.

JK : Additionally, PPU cores can be smaller than a typical CPU core, so their power consumption will also be lower.

Sources : Flow Computing

>> Discover the White Paper in PDF (Design Objectives, Advantages and Benefits of Flow Computing)