BDTI's DSP Insider Archives
|
||
BDTI's DSP Insider Archives |
|
|
| HOME << |
|
|
|
||
This month:
DSP Software Library Supports x86, XScaleIn April Intel announced version 2.0 of its Integrated Performance Primitives (IPP) library of DSP functions. This library is unusual in that it supports a diverse set of processor architectures: Itanium (IA-64), Pentium 4 and Xeon (IA-32), XScale, and StrongARM. Intel claims IPP 2.0 is highly optimized for each architecture, allowing application software developers to gain the benefits of hand-optimized code without delving into architecture-specific features. IPP 2.0 uses the same function call syntax on all processors, easing the migration of IPP-based applications among these diverse processors. This capability is particularly relevant for applications like audio and video codecs, which often start out on the desktop and later migrate to embedded platforms. According to Intel, IPP 2.0 includes over 3,000 functions for signal and image processing. (Intel says this number represents numerous variants of “a few hundred” distinct operations.) These functions include basic vector and matrix operations, primitives for common DSP algorithms like signal generation and filtering, and functions for speech and image recognition. IPP 2.0 also includes key building blocks for speech compression algorithms like GSM-AMR and audio and video codecs like MP3 and MPEG-4. Not all library components are supported on all targets; for example, the image recognition primitives are available only for the Itanium and the Pentium 4/Xeon. For many embedded DSP applications, blazing processor performance is less important than the quality and availability of DSP-oriented development tools and software components; established DSPs are typically backed by sophisticated DSP development tools and vast libraries of DSP software. It will be interesting to see how Intel evolves products like IPP 2.0 to match these offerings. (In related news, Intel says future versions of its compilers and its VTune Performance Analyzer will support XScale and StrongARM as well as IA-32 and IA-64.)
IPP 2.0 is available now for the Itanium and the Pentium 4/Xeon with
prices starting at $199. IPP 2.0 for XScale and StrongARM is expected
to be available this quarter and will be free of charge. IPP 2.0
supports most Microsoft Windows operating systems, including 98/ME,
NT/2000, XP, and CE .NET, as well as Microsoft PocketPC 2002 and
Linux. For more information, visit
http://www.intel.com/software/products/ipp/.
TigerSHARC Swims in Crowded WatersIn February Analog Devices announced that the 250 MHz TigerSHARC had at long last reached full production. Analog Devices says this processor is particularly suitable for 3G cellular base stations; TigerSHARC also targets high-performance fixed- and floating-point defense, medical, and video applications. At its announcement in October 1998, TigerSHARC’s projected performance was head and shoulders above that of the existing competition. For example, a 250 MHz TigerSHARC can perform up to 2.0 billion 16-bit fixed-point or 500 million 32-bit floating-point multiply-accumulate (MAC) operations per second. The fastest DSPs available at that time, Texas Instruments’ 200 MHz ’C6201 and 167 MHz ’C6701, could muster only 400 million fixed-point and 334 million floating-point MACs per second, respectively. Today, however, TigerSHARC faces a much more formidable pool of competitors, particularly considering its announced price of $196. In comparison, Texas Instruments’ 600 MHz ’C6414 can perform 2.4 billion fixed-point MACs per second at the far lower price of $111. The TigerSHARC’s price is in the same league as that of a 933 MHz Motorola MPC7445; this PowerPC processor can perform 1.87 billion floating-point MACs per second and is priced at $205. (All prices are for orders of 10,000 units.)
Although production delays have eliminated its performance lead,
TigerSHARC still has some unusual advantages. Primary among these
advantages is TigerSHARC’s support for a broad range of data types,
including 8-, 16-, and 32-bit fixed-point data and 32-bit
floating-point data; similar data-type agility is also found in
AltiVec-enhanced PowerPCs, but is rare among DSPs and embedded
general-purpose processors. TigerSHARC also has a remarkably high
internal memory bandwidth of 4 billion 16-bit words per second. The
challenge for Analog Devices is to convince the market that these and
other advantages are sufficient to differentiate TigerSHARC from the
growing ranks of high-performance competitors.
BDTI Case Study
This Month: Application-Level BenchmarkingBenchmarks can be one of the most important processor selection tools. Ideally, benchmarks cut through marketing hype to reveal the hard truths about processor performance. However, benchmarks cannot meet this lofty goal unless they accurately reflect the characteristics of the target application. Simple benchmarks like Dhrystone are too arbitrary to meaningfully represent an application, and hence are poor predictors of processor performance. In contrast, kernel benchmarks measure performance on functions designed to represent key portions of a target application. Because they model key application components, kernel benchmarks are often good predictors of processor performance in the target application. However, kernel benchmarks often reflect best-case performance that ignores the impact of resource sharing. Kernel benchmarks typically run in isolation with unimpeded access to processor resources like caches, whereas real-world application tasks must share these resources with other application tasks, an operating system, and so on. Application-level benchmarking addresses these problems by running multiple application tasks (such as audio and video decompression) simultaneously along with an operating system. By more closely modeling the behavior of an application, application benchmarks can provide excellent prediction of real-world processor performance. Although a benchmark’s predictive capabilities can improve as it becomes more elaborate, increasing the benchmark’s complexity also increases the difficulty of ensuring fair and consistent results. For example, obtaining good results from an application-level benchmark requires careful specification of the type, number, mix, and operating conditions (including operating system and test data) of the application tasks.
BDTI has extensive experience with kernel-level benchmarking of
processors for DSP applications and is now bringing the same rigorous
approach to application-level benchmarking. BDTI is initially
focusing on processors and applications intended for next-generation
portable information appliances like multimedia cell phones and PDAs.
To learn more about BDTI’s application benchmarking plans, contact
Jeremy Giddings (giddings@BDTI.com).
Impulse Response, by Jeff Bier
1600 lbs. of GorillaA battle of the titans is shaping up for supremacy in embedded application processors—processors intended for next-generation portable information appliances like multimedia cell phones and PDAs. The stakes in this battle are enormous: according to Micrologic Research, worldwide cell phone shipments alone are expected to exceed 450 million units this year. And the contestants are appropriately gargantuan: Texas Instruments, the long-time king of digital signal processors (DSPs), is pitted against Intel, the 800-pound gorilla of general-purpose processors (GPPs). The processors engaged in this struggle reveal much about each company’s background: TI’s entrant, OMAP, combines a DSP and a GPP on a single chip, while Intel’s challenger, the PXA2xx, is a single-core GPP—albeit one with significant DSP enhancements. TI contends that the increasing signal-processing demands of handheld applications warrant a separate DSP, while Intel argues that data, not signals, are the real driving force behind these same applications. Of course both companies have taken self-serving stances, but both architectural approaches have merit. This contest is particularly interesting given the intertwined histories of these two companies. For example, Intel’s 2920, introduced way back in 1979, was one of the first commercially available DSPs. TI didn’t introduce its first DSP, the TMS32010, until 1982. Of course, Intel can take little pride in its early lead: the 2920 was a flop, while the TMS32010 became the progenitor of today’s most popular DSPs. Ironically, one reason the TMS32010 succeeded is that TI was the first company to provide PC-based development tools for a DSP. Hence, TI’s success in the DSP marketplace was closely tied to Intel’s success in the PC marketplace.
So who will win this contest of the giants? The answer may depend on
how quickly users take up sophisticated next-generation devices and
which of these devices’ features emerge as “killer applications.” And
as last year’s cell phone glut demonstrated, predicting consumer
demand and preferences is a tricky business indeed. In any event, the
increasingly similar capabilities of their products assure that Intel
and TI will square off in more Godzilla-sized battles in the coming
years.
“Inside” the ARM7, ARM9, and ARM9EOne of BDTI’s most popular reports, “Inside the ARM ARM7, ARM9, and ARM9E,” provides benchmarking data and analysis of the DSP capabilities of the most widely used family of processor cores. This report provides insight into how the widely-used ARM7, the newer ARM9, and the ARM9’s DSP-enhanced sibling, the ARM9E, stack up against both each other and competitors. For excerpts from the analysis and a sample benchmark, go to http://www.BDTI.com/products/reports_arm.html.
For a sample of a BDTI “Inside” report, look on BDTI’s Web site at
http://www.BDTI.com/articles/info_articles.htm#summary_reports.
First BDTI “Focus” Report: “FPGAs for DSP”In the first volume of its newly announced series of technical reports, BDTI examines FPGAs as a solution for digital signal processing applications. “FPGAs for DSP” evaluates DSP enhancements of FPGAs and explains why FPGAs are a practical solution for some DSP applications. Includes:
Originally announced for April publication, this report will now be released in early summer 2002 in order to allow development of the first DSP benchmarks for FPGAs.
For more information on BDTI “Focus” reports, go to
http://www.BDTI.com/products/reports_focus.html.
BDTI Compiler EvaluationBaffled by competing vendor claims of compiler efficiency for DSP application development? BDTI has developed a methodology to quantifiably and qualitatively assess the relative merits of C compilers for DSP applications. Developed over several years, BDTI’s C compiler evaluation methodology is ready for commercial roll-out. BDTI welcomes a small number of early participants in a multi-client study.
Interested parties, processor vendors, and compiler vendors should
contact Jeremy Giddings at BDTI (giddings@BDTI.com) for further
information.
About BDTIBDTI is an independent source for DSP technology analysis and optimized DSP software. From rigorous technical analyses of processors for DSP, such as the “Inside” series of processor analyses, to highly regarded technology training classes, BDTI is the trusted independent source for reliable information on DSP technology.
For more information, visit our Web site at http://www.BDTI.com.
The next issue of BDTI’s “DSP Insider” is coming in June. Previous issues of BDTI’s “DSP Insider” are archived on BDTI’s Web site. Follow the link from http://www.BDTI.com/dspinsider/dspinsider.html. If you have comments, suggestions, or other feedback about the “DSP Insider,” please send email to dspinsider@BDTI.com. BDTI’s DSP Insider is a free monthly electronic newsletter published by Berkeley Design Technology, Inc. If our newsletter was forwarded to you and you would like to receive it regularly, please register at http://www.BDTI.com/dspinsider.htm.
If you no longer wish to receive the DSP Insider, send an email
message to dspinsider@BDTI.com with the words “Remove me” in the
subject line.
BDTI's DSP Insider © 2002 Berkeley Design Technology, Inc. |