This year, our focus on Baidu's work centers on its advancements in deep learning. However, the company also faces significant challenges with big data across a variety of applications that are not necessarily at the cutting edge. At this week’s Hot Chips conference, Baidu’s Ouyang Jian highlighted just how massive the data volume is—Baidu handles over 1 EB of data, processes around 100 PB daily, updates 10 billion pages each day, and generates more than 1 PB of log updates every 24 hours. These numbers are comparable to those of Google, and Baidu has adopted a similar approach to tackle large-scale bottlenecks.
As we discussed earlier, both Google and Baidu are exploring ways to go beyond Moore’s Law, with a strong emphasis on accelerating machine learning tasks. According to Ouyang, the gap between Baidu’s demand for high-end services based on its own data and the capabilities of traditional CPUs is expected to grow. This has led the company to look into alternative computing solutions, such as FPGAs, which offer flexibility and performance for specific tasks.
For Baidu’s vast data processing needs, a wide range of frameworks and platforms support data analysis, including knowledge graphs, multimedia tools, natural language processing systems, recommendation engines, and clickstream analysis. The core challenge of big data lies in matching these diverse applications with the overwhelming amount of data they must process.
When it comes to accelerating big data analysis, Baidu faces several hurdles. As Ouyang pointed out, the diversity of big data applications and the constantly evolving types of computation make it difficult to find a universal solution. Integrating everything into a distributed system is complex due to the changing platform and programming models (like MapReduce, Spark, streaming, or user-defined functions). Additionally, the number of data types and storage formats is expected to increase in the future.
Despite these challenges, Ouyang shared some common patterns his team discovered. He emphasized that SQL plays a central role in many of Baidu’s data-intensive tasks, with about 40% of their analysis done using SQL. Moreover, they can leverage existing SQL systems like Hive, Spark SQL, and Impala. The next step was to accelerate SQL queries, and Baidu found that FPGAs were the most effective hardware for this purpose.
The FPGA-based processing units (PEs) automatically handle key SQL functions during query execution. While the details are still somewhat vague, it’s clear that Baidu is pushing the boundaries of what’s possible with custom hardware. If the performance improvements seen in benchmarks are real, this could signal a major shift in how large-scale data is processed.
Ouyang mentioned that the current accelerator is limited by the FPGA’s memory bandwidth, but otherwise, performance could be even higher. In their tests, Baidu used two 12-core Intel E26230 CPUs running at 2.0 GHz with 128 GB of RAM. The SDA device features five processing units, each handling different SQL operations such as filtering, sorting, aggregation, joining, and grouping.
To speed up SQL queries, Baidu developed a special engine called the Processing Unit (PE) to optimize five key operations in the TPC-DS benchmark. These accelerators use an offload model, where the FPGA-based cards perform the heavy lifting. Data is sent in a columnar format to the accelerator, making queries significantly faster. Through a unified SDA API and driver, SQL operations are distributed to the appropriate processing unit for acceleration.
The SDA architecture follows a data flow model, where unsupported operations are returned to the database system for local execution. Although Baidu hasn’t fully disclosed how data and queries are distributed across the cluster, the results from their benchmarks have been impressive, especially in the Terasort test. We’ll continue to follow Baidu’s progress after the Hot Chips conference to see if more details emerge about the full implementation and how they plan to address the memory bandwidth bottleneck.
Piezoelectric Rings For Ultrasonic Machining
Ultrasonic processing is a special processing tool that uses ultrasonic frequency to vibrate at small amplitudes, and through the impact of the abrasive free of liquid between it and the workpiece on the processed surface, the surface of the workpiece material is gradually broken.
Yuhai company is engaged in produce high performance piezoelectric elements,
Custom Hifu Piezo Parts,Hifu Ultrasonic Focusing Part,Piezo Crystal For Hifu,Cheap Hifu Transducer
Zibo Yuhai Electronic Ceramic Co., Ltd. , https://www.yhpiezo.com