The Future of Adaptable Computing in the Data Center with Xilinx

Hello everyone and thank you for joining us at this event. We'll be speaking to you today about the future of adaptive computing with Xilinx technology. I'll be joined by three of my Xilinx colleagues who will cover some of our new data center products and solutions and our vision for the composable data center of the future. We are looking forward to a lively exchange with you.

Well, here we are on another Zoom session. The pandemic has obviously been a catalyst for a huge surge in streaming content, and if you're a data center operator, you probably haven't been getting a lot of sleep over the past 12 months, worrying 24/7 about how to handle the unprecedented surge in traffic like this. But this past year has really just emphasized a challenge for data center operators that has been looming for years.

As you know, the uniprocessor performance of x86 processors hit a ceiling over a decade ago. Most of the performance gains since then have come from multi-core architectures, and now even that approach has hit scalability and power limits. As a result, the clear trend at data centers is the shift to heterogeneous compute architectures, which combine different types of domain-specific accelerators that are optimized for specific tasks.

As an example, Amazon's AWS data center now features 300 different types of compute instances for every conceivable type of compute or storage application. With the increasing diversity of applications and continuously evolving standards, it's very difficult for data centers like AWS to have just the right amount of each type of resource and avoid obsolescence due to changing requirements.

That's why adaptable devices such as FPGAs are gaining momentum as adaptive acceleration engines that can be customized to specific domains and configured on the fly. Today you'll be learning about how Xilinx is paving the way to a fully composable data center that can be instantly reconfigured to accelerate a diverse set of workloads. Xilinx is the FPGA market leader with over 50% market share.

For those of you who aren't familiar with FPGAs, these devices have massive parallelism and a customizable data path that can be optimized for different applications in just milliseconds by simply loading in a new bitstream. For example, as this illustration shows, the FPGA hardware structure can be reconfigured with software to be an encryption accelerator, a decryption accelerator, and a data analytics accelerator as the workload requires.

We are seeing FPGAs continue to expand their footprint in the data center, standing shoulder to shoulder in heterogeneous compute architectures with GPUs and CPUs. And that's because FPGAs are the clear choice for data centers that need deterministic, low-latency performance for demanding applications, while keeping the flexibility to adapt to changing requirements and new applications.

Another key advantage for FPGAs in the data center is the ability to bring adaptable computing closer to the data wherever it is at the moment, so we can also process the data when it's at rest on disk and in motion as it's flowing through the network.

Xilinx recently partnered with Samsung to deliver FPGA-equipped NVMe drives. These smart SSDs offload CPUs by bringing compute to where the data lives, with massive linear scalability, which makes them ideal for applications like search and storage. We also recently released the SN1000 series, our next generation of SmartNICs, to accelerate data in motion with processing that is optimized for specific application or resource needs. We'll deep dive into both of these products — the smart SSD as well as the SN1000 — during our time with you today.

FPGA-equipped devices like these have the ability to solve data center challenges that GPUs and CPUs can't, and we see FPGAs taking on a critical and indispensable role in data center heterogeneous compute architectures.

Now, the conventional FPGA narrative is that while they are powerful, they require arcane hardware development expertise to operationalize. That's proved a significant barrier for broader developers and ISV community, but Xilinx has demolished that barrier with off-the-shelf Alveo accelerators and a development environment that enables software developers to exploit FPGA performance.

The Vitis unified software platform, which we introduced in 2019, provides a comprehensive software stack with libraries, programming tools, runtime, and target platform. It is integrated with domain-specific application frameworks like TensorFlow and FFmpeg, so that application programmers can continue to work in their familiar environment and yet get the full benefit of FPGA acceleration.

With the introduction of the SN1000 series SmartNICs, we have extended Vitis to include application-level programming of smart packet processing using P4 programming language. As a result of this, Xilinx FPGAs and accelerator cards have become widely accepted acceleration platforms, and a large ecosystem of companies has formed to offer a variety of products and services for data centers.

For instance, Amazon F1 FPGA as a Service platform exclusively uses Xilinx FPGAs. And just last month, Azure announced general availability of their NP series of virtual machines built on Xilinx Alveo accelerator cards. The Azure environment enables developers to target both on-premise and cloud customers with a single codebase.

Additionally, a large number of independent software vendors now offer accelerated applications that run on our FPGAs, a number of which we'll detail for you today. Finally, some turnkey application platforms combine hardware and software to accelerate compute-intensive applications in areas like biotech, fintech, and data analytics.

Let's look at a couple of examples that illustrate the power of FPGA-based acceleration today. Illumina has been using our FPGA-based acceleration for a few years now. Their first application was to diagnose genetic diseases in newborns so treatments could be started immediately. The analysis process used to take about a day and a half, but with our acceleration, the entire human genome can be analyzed in less than 22 minutes.

In the past year, a new application has emerged which has placed even greater demands on performance and cost. That is to identify and track variants of the coronavirus and help trace clusters of infection. Fortunately, with the adaptability of their FPGA-based solution, Illumina was able to adapt their system for this task, and they have now sequenced and analyzed over 100,000 coronavirus samples to date.

Illumina's application was initially available only as an on-premise appliance, but with increasing demand, it has become a cloud-based service on AWS and now on Azure as well.

This second example is for accelerating compute-intensive functions in graph databases, which are a popular way to represent relationships among unstructured data. We have been working with a well-known graph database partner of ours, TigerGraph. This specific use case is for a leading healthcare provider, and it entails matching an incoming hospital patient with 10 million previous patients, identifying similarities based on 200 features using a cosine similarity function. The goal is to quickly identify a course of treatment based on treatments used for similar patients in the past.

Using five Xilinx Alveo accelerator cards, we can speed up this computation by nearly 400 times compared to two Xeon x86 servers and compute the result in under 100 milliseconds, as required by the customer.

Clearly, adaptable compute, network, and storage acceleration can have a major impact on performance and latency, and data centers are aggressively adding accelerators to the mix. However, like any assets, accelerators need to be used efficiently in order to maximize the data center investment.

A new architecture is emerging that promises to make a dramatic improvement in resource utilization. It's known as composable infrastructure. It entails decoupling and pooling of resources which are accessible from anywhere and configurable into virtual servers by software. This approach enables dynamic provisioning of workloads with just the right amount of resources, including processors, SSDs, and accelerators.