In today's computing landscape, CPUs and FPGAs have been collaborating for quite some time. As heterogeneous computing platforms composed of coprocessors like GPUs, FPGAs, and other intelligent engines become more common, they are increasingly being used to enhance computing performance. This has become a major focus in both academic research and industrial applications. Let’s explore how the combination of CPU and FPGA is applied in image data storage systems.
With the rapid growth of the mobile internet, images have become a central part of daily information exchange. In many ways, we’ve entered the "image reading era." According to research, global device-generated internet data was expected to reach 24.3 EB per month in 2019, with 60% of that traffic coming from images. In this "big image era," where are all these high-quality images stored? How can we address the rising costs of storage?
One effective solution is the WebP image lossy compression scheme accelerated by FPGA. This approach enables fast conversion between JPEG and WebP formats. Compared to traditional methods, it achieves up to 14 times higher efficiency and supports real-time image retrieval and transmission with much higher concurrency.
**Using computing power to “replace†storage space**
Currently, common image formats on the internet include JPEG and GIF. However, their limited compression ratios lead to excessive use of server storage resources. To tackle this issue, Google introduced the WebP format, which significantly reduces file sizes while maintaining image quality. WebP files are 39.8% smaller than JPEG, 26% smaller than PNG, and 64% smaller than GIF. Using WebP can reduce page transfer time by 33% and page load time by 10%.
In China, platforms like Tencent News and QQ Space have started adopting WebP. This transition has led to a reduction of 9GB in peak traffic bandwidth and a decrease of 100ms in picture and data download delays. However, WebP still requires significant computational power to process, effectively using compute resources to replace storage space.
Because of its more complex compression algorithm, WebP encoding and decoding demand more processing power than JPEG. This results in a 10x or greater drop in processing efficiency. Since CPUs aren't optimized for high parallel tasks, developing efficient hardware acceleration for image encoding and decoding is key to the widespread adoption of WebP.
**FPGA-accelerated compression improves overall performance by 14 times**
While a single image codec task may seem small, when dealing with large volumes of images concurrently, the workload becomes massive. A serial CPU-based system is limited by clock speed, making it inefficient for such tasks.
Inspur addresses this by leveraging the parallel computing power of FPGAs. By hardware-accelerating the parallel parts of high-performance algorithms, Inspur’s FPGA platform can execute multiple instructions simultaneously, greatly improving throughput. This makes the system function like several traditional CPUs working in unison.
Inspur’s WebP image compression solution uses the industry’s highest-density FPGA card, the F10A. It embeds an optimized WebP codec algorithm tailored for FPGA environments, utilizing hardware pipelines and task-level parallelism to dramatically boost performance.
According to Inspur’s measurements, the FPGA solution is 14 times more efficient than traditional CPU-based solutions. In one test, converting 1200 JPEG images (2048 x 1536 resolution) took 33.4 seconds on a dual Xeon E5-2690v3 server, handling only 35 images per second. With the FPGA acceleration, the same task completed in just 2.39 seconds, achieving 502 images per second — a 14.37x speedup.
In today’s world, images are essential across e-commerce, instant messaging, image search, social platforms, and more. The increasing reliance on visual content puts heavy pressure on data center storage and distribution. Inspur’s WebP image compression solution not only enhances transcoding efficiency but also speeds up image delivery and reduces backend storage requirements.
Currently, Inspur offers FPGA-based solutions for WebP image compression, Gzip data compression, and ResNet models. These solutions provide a significant performance-to-power ratio advantage over traditional implementations. Looking ahead, Inspur plans to expand FPGA-based general-purpose systems into areas like deep learning, network acceleration, and storage optimization, bringing these benefits to more customers and application fields.
In the future, the CPU+FPGA combination could emerge as a new form of heterogeneous computing, gaining wider adoption in data centers and AI-driven applications.
Nvidia Graphics Card
Boluo Xurong Electronics Co., Ltd. , https://www.greenleaf-pc.com