Breakpoint 2023: FPGA Working at 8M TPS

Summary

At Breakpoint 2023, Carve, a hardware architect from Jump Trading Group, delivered an engaging presentation on the capability of Field-Programmable Gate Arrays (FPGAs) to accelerate signature verification at an unprecedented rate. A seven-year-old FPGA was demonstrated to process 1 million transactions per second using just 50 watts of power, with minimal latency. Carve explained how utilizing a set of eight FPGAs could scale this up to 8 million transactions per second. The presentation covered the technical challenges overcome and innovations applied in this process, revealing insights on how this affects not only power consumption and efficiency but potentially the broader landscape of computing in finance and cryptology.

Key Points:

The Advancements in FPGA Acceleration

Carve introduced an impressive development in the use of FPGAs for accelerating signature verification – a critical process in financial transactions and cryptographic applications. He detailed how, after overcoming various technical challenges and design complexities, their team achieved a million signature verifications per second on a single FPGA. Notably, this feat was accomplished with 50 watts of power and a 200-microsecond latency without the need for transaction batching – a method commonly used to enhance throughput which also adds latency and complexity.

Significance of Heterogeneous Computing for Throughput

Diving into the mechanics behind this achievement, heterogeneous computing was highlighted as a key concept. This approach allows different regions of an FPGA chip to be programmed to perform distinct tasks. By using this methodology, Jump Trading Group optimized overall system performance, achieving 8 million transactions per second using eight FPGAs, with a total power consumption of only 400 watts. The ability to connect FPGAs directly to networks was cited as a great advantage for data network processing applications, distinguishing FPGAs from GPUs in the scope of parallelism and resource management.

The Role of FPGA in the Ecosystem of Computing

The discussion also expanded upon the nature of software and hardware paradigms. Carve demonstrated how hardware design allows for more parallel processing compared to the sequential nature of traditional software design. A detailed breakdown was provided on assembling a computational pipeline on an FPGA to optimize the throughput independent of latency. The importance of optimizing calculations for the available area on an FPGA chip reflects a significant mindset shift from software to hardware engineering, with parallelism being essential for scalable, ultra-high-performance systems.

Facts + Figures

• FPGAs can now process 1 million signature verifications per second.
• Power consumption is exceptionally low at 50 watts per FPGA.
• The latency figure is a remarkable 200 microseconds.
• By deploying eight FPGAs, throughput scales up to 8 million transactions per second with a peak power of just 400 watts.
• Reuse of functions and not reinventing processes were key to their optimization strategy.
• The FPGA used in the demonstration is already seven years old.
• Amazon's AWS cloud was chosen for the demo due to its accelerator style infrastructure.
• The SHA and SV0-SV3 parts of the signature verification algorithm were decomposed and allocated across the FPGA.
• A custom ECC CPU was developed on top of the FPGA for specific control flow heavy computation segments.
• The demo showed an actual operational process on an AWS machine running in real-time.

Top quotes

• "We can achieve 1 million signature verifications per second on a single FPGA, and that's using a peak power of 50 watts."
• "We are streaming transactions. We're not batching. Every single transaction is being processed individually."
• "To get this kind of throughput, you do have to batch transactions, and that causes issues of itself."
• "Nothing stops us from implementing our very own custom processor on top of our programmable fabric."
• "A properly pipelined design, the latency of the system has little to no impact on the throughput of the pipeline."
• "The FPGA does not have direct connectivity to the network."
• "Assigning area to accelerate one algorithm is area that can't be used to accelerate another."
• "Optimizing for area and trading latency for throughput are some of the key tricks and differences in mentality between hardware and software engineers."
• "By re-architecting low-level computational primitives... we can dramatically increase the performance of key algorithms while enhancing security."

Questions Answered

What is FPGA and why is it important?

FPGA stands for Field-Programmable Gate Array, which is a type of hardware that can be programmed to perform specific computational tasks after manufacturing. It is important because it offers flexibility and efficiency, allowing for custom hardware acceleration of complex processes. For instance, in the context of Jump Trading Group's demonstration, a single FPGA could perform a million signature verifications per second, making it a powerful tool for applications in cryptography and financial transactions.

How does FPGA acceleration compare to traditional software methods?

With FPGAs, computations can be carried out in parallel, as opposed to traditional software methods, which often operate sequentially. This allows for greater efficiency and performance, as evident in the significant increase to 8 million transactions per second showcased by Jump Trading Group without a significant power increase.

What makes the FPGA presented by Jump Trading Group notable?

What's notable is the ability of the seven-year-old FPGA used in the demonstration to process an exceptionally high volume of transactions quickly and efficiently. This highlights the robust potential of existing technology when optimized effectively for new demands, picking extra relevance in the context of growing transaction processing needs.

Could this breakthrough in FPGA technology affect everyday computing?

While the direct application is specialized towards signature verifications, the underlying principles and methods could influence the evolution of everyday computing. The focus on power efficiency, acceleration, and throughput optimization might inspire similar innovation in consumer electronics, data centers, and more.

Why did Jump Trading Group use a seven-year-old FPGA for this demonstration?

Despite its age, the FPGA demonstrated that with the right design and optimization, existing hardware could achieve remarkable efficiency and performance levels. By showcasing the viability of using older technology in new ways, Jump Trading Group emphasizes the potential longevity and adaptability of FPGAs. It implies that through ingenuity, the life and relevance of semiconductor components can be extended, having profound implications for sustainability and cost management in the tech industry.