Part of the  

Chip Design Magazine

  Network

About  |  Contact

Headlines

Headlines

High Level Synthesis (HLS) Splits EDA Market

Recent acquisitions and spin-offs by the major electronic design automation company’s reveals key differences in the design of complex chips.

Last week, Cadence Design Systems announced the acquisition of Forte Design. This announcement brought renewed interest to the high-level synthesis (HLS) of semiconductor chips. But the acquisition also raises questions about emerging changes in the electronic design automation (EDA) industry. Before looking at these wide-ranging changes, let’s see how this acquisition may affect the immediate HLS market.

At first glance, it seems that Cadence has acquired a redundant tool. Both Forte’s Cynthesizer and Cadence’s C-to-Silicon are SystemC-based applications that help chip designers create complex system-on-chips (SoCs) designs from higher levels of abstraction. “High-level synthesis (HLS) tools synthesize C/C++/SystemC code targeting hardware implementation, after the hardware-software trade-offs and partitioning activities have been performed upstream in the design flow,” explained Gary Dare, General Manager at Space Codesign,  a provider of front-end architectural EDA design tools.

Although both Cadence’s and Forte’s HLS tools are based on SystemC, they are not identical in function.

Forte’s strength lies in the optimization of data path design, i.e., the flow of data on a chip. This strength comes from Forte’s previous acquisition of the Arithmetic IP libraries, which focuses on mathematical expressions and related data types, e.g. floating-point calculations.

How do the data bits from arithmetic computations move through a SoC? That’s where the C-to-Silicon tool takes over. “Forte’s arithmetic and data focus will complement Cadence’s C-to-Silicon strength in control logic synthesis domain,” notes Craig Cochran, VP of Corporate Marketing at Cadence. The control plane serves to route and control the flow of information and arithmetic computations from the data plane world.

Aside from complementary data and control plane synthesis, the primary difference between the two tools is that C-to-Silicon was built on top of a register-transfer level (RTL) compiler, thus allowing chip designers to synthesize from high-level SystemC level down down to the hardware specific gate level.

The emphasis on the SystemC support for both tools is important. “Assuming that Cadence keeps the Forte Design team, it will be possible to enhance C-to-Silicon with better SystemC support based on Cynthesizer technology,” observed Nikolaos Kavvadias, CEO, Ajax Compilers. “However, for the following 2 or 3 years both tools will need to be offered.”

From a long-term perspective, Cadence’s acquisition of Forte’s tools should enhance their position in classic high-level synthesis (HLS). “Within 2013, Cadence acquired Tensilica’s and Evatronix’ IP businesses,” notes Kavvadias. “Both moves make sense if Cadence envisions selling the platform and the tools to specialize (e.g. add hardware accelerators), develop and test at a high level.”

These last two process areas – design and verification – are key strategies in Cadences recent push into the IP market. Several acquisitions beyond Tensilica and Evatronix over the last few years have strengthened the company’s portfolio of design and verification IP. Further, the acquisition of Forte’s HSL tool should give Cadence greater opportunities to drive the SystemC design and verification standards.

Enablement verses Realization

Does this acquisition of another HLS company support Cadence’s long-term EDA360 vision? When first introduced several years ago, the vision acknowledged the need for EDA tools to more than automate the chip development process. It shifted focus to development of a hardware and software system in which the hardware development was driven by the needs of the software application.

“Today, the company is looking beyond the classic definition of EDA – which emphasizes automation – to the enablement of the full system including hardware, software and IP on chips and boards to interconnections and verification of the complete system,” explains Cochran. “And this fits into that system (HLS) context.

The system design enablement approach was first introduced by Cadence during last month’s earning report. The company has not yet detailed how the “enablement” approach relates to its previous “realization” vision. But Cochran explains it this way: “Enablement goes beyond automation. Enablement includes our content contribution to our customer’s design in the form of licensable IP and software.” The software comes in many forms, from the drivers and applications that run on the Tensilica (IP) processors to other embedded software and codices.”

This change in semantics may reflect the change in the way EDA tool companies interface with the larger semiconductor supply chain. According to Cochran and others, design teams from larger chip companies are relying more on HLS tools for architectural development and verification of larger and larger chips.  In these ever growing SoC designs, RTL synthesis has become a bottleneck. This means that chip designers must synthesize much larger portions of their chips in a way that reduces human error and subsequent debug and verification activities. That’s the advantage offered by maturing high-level synthesis tools.

Cadence believes that SystemC is the right language for HLS development. But what is the alternative?

HLS Market Fragments

The other major high-level synthesis technology in the EDA market relies on ANSI-C and C++ implementation that involve proprietary libraries and data types, explained Cochran. “These proprietary libraries and data types are needed to define the synthesis approach in terms of mathematical functions, communication between IP blocks and to represent concurrency.” The ANSI-C approach appeals to designers writing software algorithm rather than designing chip hardware.

Kavvadias agrees, but adds this perspective. “Given the Synopsys’s recent acquisition of Target Compiler Technologies (TCT), it appears that the big three have different HLS market orientations: Cadence with a SystemC to ASIC/FPGA end-to-end flow, Snopsys moving on to application-specific instruction-set processor (ASIP) synthesis technology, while Mentor has offloaded its HLS business.”

“Further, Synopsys now has two totally distinct ASIP synthesis technologies, LISATek’s Processor Designer and TCT’s IP Designer,” noes Kavvadias. “They are based on different formalisms (LISA and nML) and have different code and model generation approaches. In order to appeal to ASIP synthesis tool users, Cadence will have to focus to the XPRES toolset. But I’m not sure this will happen.”

A few years ago, Mentor Graphics spun out it HLS technology to Calypto. But Mentor still owns a stake in the spin-off company. That’s why long-time EDA analyst Gary Smith believes that the Forte acquisition puts Cadence and Mentor-Calypto’s CatapultC way ahead of Synopsys’s Synfora Synphony C Compiler. “The Synopsys HLS tool pretty much only does algorthmic mapping to RTL, whereas Forte and Mentor-Calypto tools can do algorthmic mapping, control logic, data paths, registers, memory interfaces, etc. — a whole design.”

What does the Future hold?

Forte’s tool focus on data path synthesis and associated arithmetic IP should mean few integration issues with Cadence’s existing HLS tool C-to-Silicon. However, Kavvadias notes that the acquisition makes floating-point IP increasingly important. “It is relevant to algorithmists (e.g. using MATLAB or NumPy/SciPy) wishing to push-button algorithms to hardware.” The efficient implementation of floating-point functions is not a trivial task.

Kavvadias  predictions that, “if CDNS buys a matrix-processor oriented IP portfolio, then their next step is definitely a MATLAB- or Python-to-hardware HLS tool and maybe the MATLAB/Python platform beyond that.” Matrix processors are popular in digital signal processing (DSP) applications that require massive multiply-accumulate (MAC) data operation.

Today’s sensor and mobile designs require the selection of the most energy-efficient platforms available. In turn, this mandates the need for early, high-level power trade-off studies – perfect for High-Level Synthesis (HLS) tools.

Tags: , , , , , , , , , , ,

3 Responses to “High Level Synthesis (HLS) Splits EDA Market”

  1. jb Says:

    Great follow-on question from Phil Tomson via Twitter:

    @Dark_Faust “Python-to-hardware HLS tool” How likely is that? Would think that something like Julia would be a better language for that. @philtor

  2. jb Says:

    @Dark_Faust (jblyler) to @philtor
    Python-to-HLS might work if CDNS gets closer to MathLab esp. with matrix processing. Don’t know Julia. Is it better?

  3. jb Says:

    From Phil via Twitter:

    @Dark_Faust Julia is a new lang in the MatLab space. open source, parallelism and LLVM backend would probably make it easier to synthesize.

Leave a Reply