Data is now defining the future of computing technology as massive data generation combines with a competitive imperative to increase analysis and produce actionable insights.
Most, if not all, areas of infrastructure that handle and process data are experiencing significant market transformation. Demands for increased agility and flexibility are increasing, along with greater diversity in data types and new methods and algorithms for deriving value from data. New ways to handle data and demands for new data services are cropping up with increasing frequency, compounding the complexity for decision makers, solution providers and application developers.
For example, practical applications of artificial intelligence (Al) continue to evolve and it is still unclear exactly how Al can and will be deployed in many areas. This uncertainty means that flexibility will be needed to handle these evolving Al workloads. It is also unclear what level of precision may be needed for many Al applications. To handle different levels of precision, a flexible approach is required.
Edge and embedded
With a wide range of businesses and industries tied to the Internet of Things (loT) and the installed base of connected devices expected to reach nearly 31 billion by 2020, more data is being created, processed and transmitted at the edge. This brings an expectation that embedded devices at the network edge will be acting on structured and unstructured data and generating insight in near-real-time.
Processing the raw, unfiltered data, structuring it and conducting deep learning inference at the edge demands new levels of performance and capacity.
As data is ingested and moves from the edge through the network core, ever-increasing amounts of traffic must be processed. When network capacity relies on centralised hardware, value-add features can only be deployed where the specific hardware is in place to support them. But today, more intelligence and flexibility are necessary at several key areas at the network level to both transport and process the massive data loads more quickly and to decentralise functions that were implemented at the equipment level so they can be deployed where needed.
As a result, NFV (network function virtualisation) has emerged as a means of providing the necessary resource flexibility to ensure high utilisation and high levels of programmability to cover diverse networking workloads. NFV can be built on commercial off-the-shelf (COTS) infrastructure to provide cost-effective deployment.
But virtualised network functions can have limitations in maintaining quality of service (QoS) due to the increasing amount of data and the rise of low-latency, high-bandwidth applications. FPGA acceleration optimises the utilisation and cost-effective scaling of Intel architecture-based NFV and extends the capabilities of SmartNICs to accelerate networking as users, data and applications increase over time.
Also, 5G places new demands on networks – they must be highly available and scale efficiently to accommodate a massive number of connected devices with the exponential rise of data traffic and different quality-of-service levels generated by different use cases. Current solutions that use specialised appliances will not be sufficient to support future 5G needs. Proprietary hardware may be prohibitively expensive for broad deployments and may not be flexible enough to respond to evolving customer needs, as well as leverage an available developer ecosystem. Here, too, flexibility is needed to accommodate a changing customer landscape and a variety of implementation and deployment options.
Many market sectors now conduct critical analysis based on data centre processing utilising data sets of ever-increasing size, with use cases ranging from financial analytics, database acceleration and genomics to video transcoding, network and storage acceleration and security.
Moreover, the kinds and types of data processing happening within the data centre are in constant flux, in part because data centre customers are increasing demands to perform different types of data analysis in real-time. In addition, data centre operators may not know what kinds of data services their customers will want in advance. If data centre operators rely on a solution that only accelerates a specific operation or algorithm, they won’t have the flexibility to respond to future acceleration demands. Advanced planning is required to handle processing tasks that have not yet been defined – another reason that flexibility at the hardware level (the type of flexibility FPGAs provide) is essential.
Utilising data centre processing resources most efficiently is crucial and acceleration can be instrumental in supporting data-centric computing. Because Intel Xeon Scalable processors power many of the world’s data centres, there is a need to accelerate functions that Intel Xeon Scalable processors perform with as low delay and as little latency as possible to deliver peak performance for the most valuable operations, in the most resource efficient way possible.
The need: Flexibility combined with market-specific requirements
As end customers expand their service offerings and solutions, there is a need for highly customisable processing of data wherever it is generated, processed, transported, or stored. Processing at this level requires specific functionality to achieve the optimal power and performance. Each application class across the various markets has specific, unique requirements for data handling. These can include functions that are power- and cost-efficient for small form factors or complex designs where power, space and cost are at a premium, such as edge and embedded devices or vehicles; handling the highest data traffic and Ethernet speeds in the network; or providing high-bandwidth, low-latency computing acceleration in data centres.
To meet their specific needs, product developers have the option to use standard off-the-shelf products, or to design and deploy custom ASICs. However, standard products may not precisely fit specific industry requirements or allow sufficient market differentiation. Custom ASICs can perform these specific functions, but are typically time- and cost-intensive to develop and have high ROI (return-on-investment) hurdles to be economically viable.
The ideal solution combines the best of two worlds: FPGA flexibility and ASIC-level performance and power efficiency for specific functions.
Architecting the ideal solution
A new architectural approach is needed to address these challenges. With Intel Agilex FPGAs, Intel is leveraging a new and disruptive approach to FPGA architecture that creates tailored FPGA products designed to address the unique challenges in each application class.
The ideal solution combines flexibility with maximum power and performance efficiency. To make this possible and build the next generation of programmable logic, Intel’s transformative approach to FPGA architecture enables the integration of a wide range of semiconductor elements in a single system-in-package (SiP).
This approach combines a high-performance FPGA core die built on the Intel 10 nm SuperFin manufacturing process with function-specific chiplets, all integrated heterogeneously into a single product with advanced 3D packaging. This enables Intel to address a broad array of acceleration and other applications with tailored, yet flexible, solutions.
The chiplets provide functionality such as PCle Gen5, 116 G transceivers and cache-coherent interfaces (Compute Express Link (CXL)) to Intel Xeon Scalable processors. Other chiplets are also possible, including other transceiver types, custom I/O and custom computing functions. Leveraging Embedded Multi-Die Interconnect Bridge (EMIB) and other leading-edge proprietary integration and packaging technologies, the new architectural approach allows the combination of traditional FPGA dice with purpose-built semiconductor dice to create devices that are uniquely optimised for target applications.
The Intel Agilex FPGAs and SoCs enable next-generation, high-performance applications via higher fabric performance, lower power, gains in digital signal processing (DSP) functionality and higher designer productivity compared to previous-generation FPGAs. The Intel Agilex FPGA meets the myriad challenges of data-centric computing while opening up new possibilities for business and industry. It brings together a general-purpose fabric for flexibility with highly efficient processing at the silicon level for the specific, customised functions demanded by each market.
Custom functions can also be rapidly integrated into Intel Agilex FPGAs through the proprietary and unique Intel eASIC device technology. With Intel eASIC technology, customer FPGA designs can be converted into function-specific die that provide ASIC-level performance and power efficiency, integrated into a single component package along with other functions for advanced customisation.
Flexibility combined with agility to meet target application requirements
FPGAs have been highly valued for their flexibility to meet evolving market requirements. However, there are some functions that have stabilised and no longer require as much flexibility. It is desirable to ‘harden’ these functions as much as possible to get the most power efficiency and performance. Hardening these functions produces these benefits because added flexibility always comes with a necessary trade-off in lower power efficiency and higher power.
Traditionally, FPGAs were designed with a single monolithic die or multiple instances of the same monolithic die type. Now, new advanced packaging technologies from Intel are enabling multiple, disparate silicon dice within a single package. By integrating dice from different process types and functions, Intel offers unprecedented flexibility and customisation. Examples of purpose-built dice include:
• Interfaces for low-latency, cache-coherent processor acceleration.
• Advanced analog functions like 116 G transceivers (XCVRs and data converters).
• Custom computing engines for application-specific functions.
• Memories of different types and configurations that can be closely coupled to the logic fabric.
Intel Agilex FPGA elements
Advanced 10 nm FPGA fabric
The FPGA fabric die at the heart of every Intel Agilex FPGA is built on Intel’s 10 nm SuperFin chip manufacturing process technology, the world’s most advanced FinFET process. The fabric die leverages the second generation of Intel Hyperflex FPGA architecture, which uses registers, called hyper-registers, throughout the FPGA, optimised for leading performance on 10 nm. The second generation of Intel Hyperflex FPGA architecture, combined with Intel Quartus Prime software, delivers the optimised performance and productivity required for next-generation systems.
The FPGA fabric also features architecture optimisations for accelerating Al functions and DSP operations through dedicated structures for half-precision floating-point (FP16) and BFLOAT16, as well as increased DSP density compared to prior generation FPGAs.
Intel Agilex FPGAs can implement fixed-point and floating-point DSP operations with high efficiency. The DSP blocks provide double the number of 9x9 multipliers compared to the prior generation. This also doubles the amount of INT8 operations that Intel Agilex FPGAs can deliver per DSP block. The addition of new modes for FP8 and FP16 supports highly efficient implementations for specific Al workloads, such as convolutional neural networks (CNNs) for image and object detection, with a lower device utilisation and lower power compared to implementation with FP32.
Second generation Intel Hyperflex FPGA architecture
The innovative second generation of Intel Hyperflex FPGA architecture supports levels of performance not possible with conventional architectures. Like the first generation, the second generation Intel Hyperflex FPGA architecture employs additional registers, called hyper-registers, everywhere throughout the core fabric.
These registers are available across the routing structures and at the inputs of all functional blocks. The hyper-registers provide a fine-grained solution to the problem of how to increase bandwidth and improve area and power efficiency. When hyper-registers are used to implement these techniques, all other FPGA logic resources are available for logic functions instead of being sacrificed as feed-through cells to reach conventional LUT registers.
In the second generation Intel Hyperflex FPGA architecture, several advances have been made to improve overall fabric performance while minimising power consumption. One of the most significant improvements is the addition of a high-speed bypass to the hyper-registers, as shown in Figure 2.
On the left of Figure 2 is a representation of an Intel Stratix 10 FPGA hyper-register. You can see that there is a signal path that goes through the register and another signal path that bypasses it. Both signal paths go through a mux (multiplexer), which is controlled by configuration RAM. One of the ways Intel has improved Intel Hyperflex FPGA architecture in the second generation is by accelerating the speed of the hyper-register bypass path. This improvement increases performance for both Intel Hyperflex FPGA architecture-optimised designs and designs that are not optimised for Intel Hyperflex FPGA architecture.
In Figure 3, we see two design examples. The one on top is optimised for Intel Hyperflex FPGA architecture; the one on the bottom is not. The adaptive logic modules (ALMs) in the Intel Agilex FPGA are shown in the large, light-blue boxes and the hyper-registers are shown in two colours: orange for the unused hyper-registers and blue with a grey outline for the used ones.
As you can see, the top design has used hyper-registers, indicating that it has been optimised for the Intel Hyperflex FPGA architecture, whereas the bottom one has not used hyper-registers, indicating that it is not optimised for the Intel Hyperflex FPGA architecture.
In Figure 4, we see the signal delays from register to register in each design, which determine the critical path and ultimately, the fmax of each design.
You can see in Figure 5 that adding the hyper-register high-speed bypasses and ALM performance improvements accelerates any potential critical path in either of the two design types. The signal travels faster through each of the unused hyper-registers (the orange boxes) and also faster through ALMs where there are no synchronous clears and clock enables. In this way, the second generation Intel Hyperflex FPGA architecture delivers higher performance for all designs, whether they are optimised for the architecture or not; although those that are optimised will receive the greatest performance benefit.
Overall, the improvements to the second generation of Intel Hyperflex FPGA architecture contribute significantly to the performance benefits delivered by Intel Agilex FPGAs.
DSP acceleration functions
To meet demands for higher-precision signal processing, Intel developed the industry’s first variable-precision DSP block. In prior generations of Intel FPGAs, variable-precision DSP blocks were enhanced to include single-precision floating point (FP32) support. In Intel Agilex FPGAs, variable-precision DSP blocks have been enhanced to include support for half-precision floating-point (FP16) and BFLOAT16 (BF16). With thousands of floating-point operators available via these hardened DSP blocks, Intel Agilex FPGAs provide up to 40 TFLOPs FP16 or BF16, or up to 20 TFLOPs FP32 DSP performance.
Development advantages of chiplet-based architecture
Historically, FPGAs have integrated functions monolithically, adding individual functions like XCVRs, memory and I/O into the same die as the programmable logic. However, this approach has three distinct limitations. First, it requires all the individual functions on the FPGA to be developed on the same process. Second, it requires a new tape-out whenever any of the individual functions are updated or improved. And third, it requires a new tape-out to get different resource mixes and ratios of individual functions (i.e., more or fewer transceivers with the same amount of logic). Tape-outs are time consuming and expensive, so they need to be minimised to accelerate the pace of product innovation.
A better, more versatile approach can be achieved with Intel chiplets. A chiplet, or tile, is a physical intellectual property (IP) block designed to integrate with other chiplets through package-level integration and standardised interfaces. Like Lego blocks, chiplets enable mix-and-match infrastructure.Chiplets allow new products to be created with greater functionality, improved agility and shorter time to market, as shown by Intel FPGA early support for 116 G XCVR, CXL and PCle Gen5 technology. This advantage is shown in Figure 6.
The chiplets are built on the optimal semiconductor process for each function. The chiplets talk to the Intel Agilex FPGA fabric and can be mixed and matched to produce the exact combinations needed by a specific application class.
The manufacturing process utilises advanced packaging technology that acts as a substrate, allowing the multiple chiplets to be assembled into a single package. For example, multiple cores, I/O, FPGAs, analog, network and communication and memory accelerators can be integrated via chiplets. Third-party ecosystem technologies could be combined with Intel silicon.
Transceivers with high-speed I/O allow data to get on and off the chip very quickly. The high-speed connectivity allows processing of massive amounts of data. With the Intel Agilex FPGA generation, Intel has increased speed from 58 gigabits/sec (58 G) to 116 G with a single transceiver channel. With chiplets, the number of transceivers is not limited by the number of channels that can be floorplanned into the FPGA fabric die. The number of transceiver channels can be increased or decreased simply by using the desired amount of transceiver chiplets, rather than re-floorplanning the die to integrate a different number of channels.
There is a general industry perception that smaller is better – hence the race to 14 nm, 10 nm and so on. When compared to the competition’s
7 nm FPGA portfolio, the Intel Agilex FPGA delivers roughly two times better fabric performance per Watt.
Sometimes, there are reasons not to use the latest semiconductor process for some functions. For example, some analog and memory functions do not scale as well as general-purpose logic with process geometry. Also, some older silicon IP functions may have been developed on older process nodes, where the benefit to move to the new node does not outweigh the cost and time involved. With Intel’s chiplet methodology, each function can be built with the specific semiconductor process ideal for its implementation. Chiplets can then be mixed and matched to achieve the right solution for a specific market need.
Advanced 3D packaging
To put everything together in a single package, you need technology that can cost-effectively integrate different dice from different functions. Traditionally, one silicon die is used in a package. Now with advanced Intel technology, it is possible to put multiple dice together from different process nodes into the same package. For example, one of the advanced 3D packaging techniques used in Intel Agilex FPGAs is Embedded Multi-Die Interconnect Bridge (EMIB), which provides a high-performance, cost-effective means of integrating chiplets together with the FPGA die in the same package, as shown in Figure 7.
These latest innovations in Intel packaging technology enable the integration of heterogenous elements together into a single package. Purpose-built silicon addresses specific market needs and a flexible core fabric joins the dice together – resulting in a massively integrated system.
With this kind of integration, Intel is coupling the FPGA fabric along with purpose-built silicon to build targeted products for specific application spaces. The resulting architecture meets the needs of many different use cases and market sectors and can work for specific functions, purpose-built for a particular space.
High-performance processor interfaces
With additional cores per processor and the much higher memory and I/O bandwidth in Intel Xeon Scalable processors, the increased demands on the intra-chip interconnect can become a performance limiter.
To facilitate high-performance in-line and offload acceleration of processor functions, Intel Agilex FPGAs support the latest generation of high-performance processor interfaces, including PCle Gen5 and CXL.
Cache- and memory-coherent interface to Intel Xeon scalable processors via CXL
Intel Agilex FPGAs support CXL, which is a high-performance, low-latency cache- and memory-coherent interface between CPUs and workload accelerators. CXL technology maintains memory coherence between the CPU memory space and memory on attached devices, which allows resource sharing for higher performance, reduced software stack complexity and lower overall system cost. This permits users to simply focus on target workloads as opposed to the redundant memory management hardware in their accelerators. More information about CXL is available at www.computeexpresslink.org.
PCle Gen5 interface
Intel Agilex FPGAs support PCI Express interfaces up to Gen5 x16. PCle Gen5 brings a leap in transfer speeds, enabling greater performance capabilities to developers of PC interconnect, graphics adaptors and chip-level communications, among other applications using this ubiquitous technology.
Advanced memory hierarchy
Memory capacity and bandwidth can be critical bottlenecks for next-generation platforms. Computing elements, such as FPGAs or CPUs, operate with data types of varying sizes and access requirements. This means that various memory types and sizes are required to facilitate the most efficient and high-performance processing.
Intel Agilex FPGAs support a broad hierarchy of memory resources, including embedded memory resources, in-package memory and off-chip memory via dedicated interfaces. The hierarchy is shown in Figure 8. The hierarchy begins with embedded, on-chip memory, including MLABs, block RAM and eSRAM, each of which offers different capacities to accommodate different processing requirements.
High Bandwidth Memory (HBM) is the next generation of high-speed memory built into Intel Agilex FPGAs using SiP technology. HBM enables the highest memory bandwidth, not feasible with other solutions. Multiple DRAM layers are connected to a base I/O layer to form 3D, high-speed memory, connected to and controlled directly by, hard memory controllers built into the Intel Agilex FPGA.
Integrating the HBM die directly into the FPGA package reduces board size and cost and simplifies and reduces power requirements.
Intel Agilex FPGAs integrate HBM2E memory next to the core fabric. The interconnect between the core fabric and memory is significantly shorter, compared to discrete memories like DDR, which reduces the amount of power traditionally spent driving long printed circuit board (PCB) traces. The traces are unterminated and there is reduced capacitive loading, which results in lower I/O current consumption. The net result is lower system power and optimum performance per Watt.
Finally, Intel Agilex FPGAs include interfaces to memory components external to the device, including DDR4/5, QDR and RLDRAM, as well as Intel Optane persistent memory, which provides a new, breakthrough option for FPGA users.
Putting it all together
Intel FPGA and SoC innovations combine to enable unprecedented levels of customisation and flexibility in Intel Agilex FPGAs. The Intel Agilex FPGA fabric and innovative chiplet architecture – enabled by Intel’s EMIB packaging technology for integration of a variety of chiplets within a single system in package – delivers an extensible FPGA platform that scales across a wide range of device densities and brings these key features and benefits:
• The latest generation of high-bandwidth processor interface interconnect including PCle Gen5 x16 support with data rates of 32 GTps per lane for next-generation data centres.
• High-performance processor interface speeds for acceleration of cache-coherent operations via dedicated support for CXL.
• Architecture optimisations for accelerating Al functions and DSP operations through dedicated structures for half-precision FP16, BFLOAT16, as well as increased DSP block density, delivering up to 40 TFLOPs FP16/BFLOAT16 or up to 20 TFLOPs FP32 DSP performance.
• 116 Gbps serial transceiver links to support the most demanding bandwidth requirements in next-generation data centre, enterprise and networking environments.
• Rapid integration of customer IP with Intel eASIC device options for better cost and power efficiency.
Accelerating application development for Intel Agilex FPGAs
Design development support for Intel Agilex FPGAs and SoCs is provided via multiple options. The first is with Intel Quartus Prime software tools. The Intel Quartus Prime software includes everything you need to design for Intel FPGAs and SoCs, from design entry and synthesis to optimisation, verification and simulation. Dramatically increased capabilities on devices with multi-million logic elements are providing designers with the ideal platform to meet next-generation design opportunities.
In addition, Intel is opening up FPGA development access to software developers starting with the Intel Distribution of OpenVINO toolkit and continuing with oneAPI toolkit. Intel’s oneAPI development environment simplifies the programming of diverse computing engines across CPU, graphics processing unit (GPU), FPGA, Al and other accelerators. The oneAPI platform includes a comprehensive, unified portfolio of developer tools for mapping software to the hardware that can best accelerate the optimised applications, middleware and frameworks.
The Intel FPGA add-on for oneAPI base toolkit will allow more designers to take advantage of Intel Agilex FPGAs. With oneAPI toolkit, developers can explore different implementations of a design across diverse hardware choices, so they can understand the particular power and performance capabilities of each implementation and ultimately choose the best fit for their specific needs.
Sample use cases
Intel Agilex FPGAs target the acceleration and computing needs for a wide range of applications from the edge to the core to the cloud – customised for specific market segments and applications – wherever data is generated, transported, stored, or processed.
|Tel:||+27 11 923 9600|
|Fax:||+27 11 923 9884|
|Articles:||More information and articles about Altron Arrow|
© Technews Publishing (Pty) Ltd | All Rights Reserved