FPGA Acceleration: Exploration and Practice for Data Centers and Cloud Services

The speech was delivered by Zhang Heng, a Tencent Cloud FPGA expert, at the 2018 Trusted Cloud Conference jointly organized by the China Academy of Information and Communications Technology and the China Communications Standards Association.

He introduced the motivation for using FPGA acceleration in data centers and cloud services, emphasizing the rapid growth of data (≈30% yearly) and the increasing demand for high‑performance computing driven by AI. Traditional CPUs are hitting the limits of Moore’s law, prompting the search for higher‑performance chips such as GPUs, ASICs, and FPGAs.

Key characteristics of high‑performance chips were described: high throughput for massive data and low latency for real‑time interaction (e.g., 5G, IoT). Compared with CPUs (software‑programmable, general‑purpose) and ASICs (fixed hardware, high efficiency), FPGAs offer a unique combination of hardware‑level performance and programmability, allowing custom pipelines for each algorithm.

Industry examples were cited: Google’s TPU, Intel’s Nervana, domestic companies Cambricon and Horizon. In the FPGA space, Microsoft equips every server with an FPGA, while Chinese giants Baidu, Alibaba, and Tencent also deploy FPGA acceleration.

Advantages of FPGA acceleration were enumerated:

High performance and low latency through custom hardware architectures.

Flexibility and scalability – rich I/O and programmable logic enable acceleration beyond compute, including storage and networking.

Low power, cost‑effectiveness, and high reliability for data‑center deployment.

Hybrid CPU+FPGA solutions that allocate tasks to the most suitable processor.

Practical Tencent use cases were presented:

Image transcoding for QQ and WeChat: FPGA‑based codecs (JPEG, WebP, HEVC) achieve 3× lower latency and 6× higher throughput compared with CPU, enabling on‑the‑fly resizing and format conversion.

Content moderation in information‑security scenarios: Two‑stage AI models filter billions of daily user‑generated images, with FPGA accelerating both models to meet real‑time requirements.

AI inference acceleration: FPGA implements core operators (convolution, pooling, normalization, activation). Benchmarks on GoogLeNet show FPGA matching GPU throughput while delivering 10× lower latency and 50% lower total cost of ownership.

Gene‑sequencing acceleration: FPGA‑enhanced BWA/GATK pipelines reduce whole‑genome analysis from 30 h (CPU) to 2.8 h, a ten‑fold speed‑up.

The talk also covered the FPGA cloud service ecosystem: hardware vendors, IP providers, and solution integrators. By offering FPGA resources as a cloud service, Tencent lowers entry barriers, shortens development cycles, and provides a unified platform for developers and end‑users.

Challenges identified include fragmented cloud‑FPGA platforms lacking standards, high development difficulty due to low‑level hardware description languages, and an immature ecosystem with limited reusable solutions.

Future plans for Tencent FPGA cloud service were outlined: upgrading hardware platforms, expanding AI‑oriented IP markets, building an evaluation system to connect developers and users, and promoting edge‑cloud integration leveraging FPGA’s rich I/O and low‑latency characteristics.