Will Infineon Break Intel's Slump?

Intel's purchase of Infineon's wireless business adds considerable support to the company's attempt to break into the smartphone market with its Atom processor--but only if Intel can manage this acquisition better than it has in previous situations. For Microprocessor Report, I did a study of Intel's 15 largest acquisitions (excluding Wind River, which is too recent to evaluate). In every one, Intel later shut down or sold off the acquired products (in most cases for much less than the original purchase price). In only one or two cases does Intel continue to use the acquired technology.

Six of these acquisitions were, like Infineon's wireless business, public companies (or units thereof) employing hundreds of people. In 1998, Intel acquired Chips & Technologies, ostensibly for its PC-graphics chips, but Intel already had its own graphics technology in house, and the impact of the C&T technology on Intel's graphics business remains unclear. Within the next few years, Intel also took on Digital Semiconductor, Dialogic, Level One, and DSP Communications. All of these business withered and were eventually sold off in a 2006 purge.

Xircom is perhaps the most successful of Intel's major acquisitions. In 2001, Intel paid $750 million for the maker of PC networking cards. Before the acquisition, the 1,900-employee company reported annual sales of nearly $500 million with a net income of 5%. Less than two years later, Intel shut down the business after revenue plummeted. But Xircom's wireless technology bolstered Intel's development of Wi-Fi chips, a market in which it remains a leader today.

Infineon ranked fourth in cellular-baseband shipments last year with 10.7% unit share, according to a recent report from The Linley Group. Infineon is a major supplier to Apple, Nokia, Samsung, and others. But Intel's focus on smartphones, and its .000 batting average with previous acquisitions, leaves the future of Infineon's non-smartphone products in doubt.

The key to this deal is not the products but retaining the Infineon engineering team. Although both vendors have a detail-oriented, big-company culture, Intel's aggressiveness may clash with German stoicism. Assuming Intel can make this combination work, it will greatly enhance the company's smartphone story. Intel will be able to develop products that combine the application CPU and cellular baseband on a single chip, offer leading-edge 3G and 4G cellular technology, and add power-management and RF components to its portfolio. The connectivity side remains a bit sketchy, but Intel now has solid FM and GPS technology and the capability to develop Wi-Fi and Bluetooth. The company needs this set of technologies to offer complete smartphone solutions that compete against Qualcomm's and ST-Ericsson's. Given the high stakes involved, Intel must find a way to make this deal succeed. --Linley

Complete coverage of this acquisition appears in a Microprocessor Report article titled "Intel Shakes Up Cellular Market."

Linley Gwennap, principal analyst

Broadcom Wins iPhone GPS

Teardowns of the iPhone 4 revealed the usual suppliers: Infineon for the cellular baseband and RF, Samsung for the application CPU (designated Apple A4), and Broadcom for the Bluetooth/Wi-Fi combo chip. One big change, however, is that the new iPhone substitutes Broadcom's BCM4750 GPS chip for Infineon's PMB2525 (Hammerhead II). This change, foreshadowed by Apple's use of the BCM4750 in the iPad released earlier this year, will upend the GPS-chip vendor rankings.

We forecast that Apple will build about 35 million iPhone 4s this year (including some that will ship in 2011) and 5 million cellular iPads (the non-cellular version does not have GPS). Thus, we estimate Broadcom's share of the GPS chip market will jump from just 6% in 2009 to 24% this year, putting in a dead heat with CSR/Sirf as the leading vendor of standalone GPS chips. Infineon, in contrast, will fall from 21% share in 2009 to an estimated 9% this year. Broadcom could gain further share if Apple adds GPS to the new iPod Touch in September.

Both the Broadcom and Infineon chips are based on a GPS design by Global Locate, which Broadcom acquired in 2007. Because of this acquisition, Infineon has not been able to update its 130nm Hammerhead design; it has since developed a new GPS product called Xposys. Instead of adopting Xposys, however, Apple chose to remain with the proven Global Locate technology. The BCM4750 uses 90nm technology to reduce power, and Broadcom has already announced a 65nm version, the BCM4751, that is likely to be used in next year's iPhone. Without its lead customer, Infineon will struggle to establish its new Xposys technology. --Linley

Complete GPS market share appears in our new report "Mobile and Wireless Semiconductor Market Share 2009."

Linley Gwennap, principal analyst

For analysis and news on semiconductors for mobile and wireless, subscribe to our free newsletter, Linley on Mobile

ARM Targets Xeon

At our recent Linley Tech conference, Marvell's Viren Shah disclosed plans to take on Xeon in the server market. The company is developing new processors based on its 2GHz Sheeva CPU, an ARM-compatible design from Marvell's own CPU team. These processors will start with 4 CPUs, but the company plans to offer 8-CPU and 16-CPU versions as well. First samples are due later this year, with the first systems likely to appear in 2H11.

Another company working on an ARM-based server chip is Smoothstone. This Austin-based startup is still in stealth mode, but it appears the company is designing its own ARM-compatible processor for servers. Smoothstone CEO Barry Evans was formerly director of marketing for Intel's (later Marvell's) PXA processors, ARM-compatible chips designed for smartphones and other mobile devices.

A stock smartphone processor lacks the memory and I/O capacity required for server products. An experienced engineering team, however, could easily design a chip combining multiple ARM-compatible CPUs with a large L2 cache, a high-bandwidth DRAM interface, and server I/O such as PCI Express, Ethernet, and SATA controllers. Such a processor could provide respectable server performance.

Most servers today use Intel's Xeon processors. The recently introduced Gulftown processor, for example, offers 6 Nehalem CPUs at up to 2.267GHz in a 60W configuration or up to 3.333GHz in the 130W version. We estimate that Nehalem delivers 2-3x the per-MHz performance of an ARM Cortex-A9 CPU on server workloads. Thus, a processor with 16 Cortex-A9 CPUs at 2.0GHz could have performance similar to that of the 2.267GHz Gulftown. Marvell's Sheeva CPU is comparable to Cortex-A9.

ARM is known for low power, but that is because it has focused on small CPUs at relatively low clock rates. To achieve 2.0GHz, ARM pushed the power of its A9 core to 1.9W. Including a large on-chip cache as well as high-speed memory and I/O interfaces, a 16-CPU ARM processor could end up at 50W, not much lower than Gulftown.

The bigger problem for Intel is its pricing model. Having vanquished all competing server processors, it takes premium margins on its Xeon products. The aforementioned 2.267GHz Gulftown model has a list price of $996; slower four-CPU models cost $440. If Marvell or another vendor can deliver competitive server processors, even at the lower end of the Xeon range, Intel could be forced to cut prices to defend its market share, a move that could grind a few gears in the profit machine. --Linley

Linley Gwennap, principal analyst

The Linley Group Acquires Microprocessor Report

Expands Services and Coverage with Addition of Award-Winning Publication and Respected Processor Analysts

MOUNTAIN VIEW, Calif.—May 7, 2010—The Linley Group, the industry's leading source for independent technology analysis of semiconductors for networking, communications, mobile, and wireless, announced today it has signed an agreement with In-Stat to acquire the Microprocessor Report, the leading newsletter for processor technology. Effective May 6, the agreement expands the services offered by The Linley Group and adds two respected processor analysts, Jim Turley and Tom Halfhill, to its research team. Financial terms of the agreement were not disclosed.

Microprocessor Report and The Linley Group both have long histories of providing high-quality, technology-focused analysis to their customers. The Linley Group will assume immediate responsibility for the Microprocessor Report’s weekly online articles and monthly subscriber newsletter, which it intends to continue without interruption.

Linley Gwennap, president and principal analyst of The Linley Group, will oversee the Microprocessor Report as publisher and editorial director, reprising roles he played when the publication won four Computer Press Awards for Best Industry Newsletter. “By bringing the publication into The Linley Group, we can use our deep domain expertise in the embedded, networking, consumer, and mobile segments to enhance and sharpen the coverage,” said Gwennap. “As processors integrate more systems functions and become more application-specific, looking beyond the CPU is critical to delivering high-quality analysis.”

Well-known industry analyst and editor, Jim Turley, will join the publication as editor-in-chief. In addition to serving as past editor of the Microprocessor Report, Turley has also held positions at processor companies in corporate management and engineering, hosted numerous industry conferences, authored several books and served as editor-in-chief of Embedded Systems Design magazine.

“I look forward to leading the content development of the report, and working with an expanded team of seasoned analysts,” said Jim Turley, editor-in-chief, Microprocessor Report. “Together, we can provide the in-depth research and articles that our subscribers need to identify technology trends and formulate successful product strategies.”

Tom Halfhill, the publication’s senior editor and analyst, has served nearly 10 years on the report’s editorial team. A respected authority on processors, computing and software, Halfhill has covered technology for industry-leading publications, including Byte Magazine, Maximum PC, Compute! and Game Players magazine.

The Linley Group’s team of technology analysts, including Jag Bolaria, Joe Byrne and Bob Wheeler, will also contribute to the report. Tapping a larger analyst team will broaden and deepen the coverage in the newsletter.

Subscribers can continue to access eight years of archived Microprocessor Report articles and weekly updates at www.MPRonline.com. Subscription inquiries can be sent to mpr@linleygroup.com.

About The Linley Group

The Linley Group is the industry's leading source for independent technology analysis of semiconductors for networking, communications, mobile, and wireless. The company provides strategic consulting services, more than a dozen in-depth analytical reports, and Linley Tech events focused on advanced technology topics. For insights on recent industry news subscribe to "Linley Wire" and "Linley on Mobile," the company's free newsletters.

Microprocessor Report is the leading technical publication for the microprocessor industry since it was first introduced by Michael Slater in 1987. Microprocessor Report is exclusively subscriber-supported and dedicated to providing unbiased, in-depth, and critical analysis of new high-performance microprocessor developments.

Will Independent OS Vendors Survive?

The impact of Intel's acquisition of Wind River continues to reverberate through the embedded industry. This week, Freescale announced that Enea and Green Hills are its partner of choice for real-time operating systems (RTOS). Although Wind River's VxWorks is the most popular RTOS among Freescale's customers, the Intel acquisition makes it difficult for Freescale to share roadmaps and code with Wind River. Instead, Freescale will work with Enea and Green Hills to optimize its RTOS for its new processors. The PowerPC vendor will continue to support VxWorks on its existing processors, but that OS will be a lower priority in the future.

Although any OS transition requires extra work for customers, this change comes just as Freescale is rolling out its first multicore processors. Moving from single core to multicore forces some software changes in any case, and Enea's RTOS is better optimized for multicore than is VxWorks. Customers that don't need multiple CPUs, however, will still have to move to the new OS at some point.

Similarly, Freescale is now partnering with Mentor Graphics for its Linux OS, after former partner Monta Vista was acquired by competitor Cavium. Freescale's goal is to establish a strong set of independent OS vendors for its platform. Intel, in contrast, is likely to focus on VxWorks as its primary embedded OS, in order to get some value from its $880 million acquisition. Although Intel hoped that Wind River could continue to gain revenue from other processors, Freescale's move will take away its most popular platform. As we predicted last year, the expensive acquisition appears to have done little for Intel, except perhaps to hamper the customers of Freescale and other competitors.

Since last year's acquisitions of Wind River and Monta Vista, no other OS vendors have been bought by processor companies. Freescale's announcement indicates it prefers an open ecosystem, providing its customers with a choice of operating systems and OS vendors. Other large processor vendors have yet to indicate their direction.

We see value in the business model of independent operating systems. The cost of developing and supporting the OS can be shared across several platforms rather than being borne by a single vendor. Although Intel can afford its own OS, smaller processor vendors cannot. We believe the industry will be stronger if other processor vendors restrain themselves from following Intel's path. --Linley

Linley Gwennap, principal analyst

ARM Outmuscles Atom on Benchmark

ARM—and, for that matter, MIPS—CPUs outperform Intel’s Atom, at least as measured by the CoreMark benchmark when normalized for frequency. ARM rates its Cortex-A9 at 2.9 CoreMark per MHz (CM/MHz), whereas Atom running a single instance of the benchmark achieves only 1.8 CM/MHz. In fact, all of the single-thread CPUs profiled in Table 1 outperform Atom in terms of per-clock performance.

The open question, thus, is which CPUs operate at higher frequencies. Atom is fastest running processor of the lot. The Z550 version is available at 2.0GHz, yielding, we estimate CoreMark performance of 3,673CM. ARM promises its A9 hard core will hit 2.0GHz (Vdd will have to be goosed up), putting its performance beyond that of the Z550. Even if ARM hits only 1.3GHz, its performance will exceed that of the fastest Atom. The fastest A9-based chip available now, Nvidia Tegra 2, hits 1.0GHz, putting it behind the garden-variety Atom N280. MIPS rates its 74K at 1.6GHz in a 40nm G process, putting it also ahead of Z550. We are unaware of any 74K-based chips sampling or in production operating in excess of 1GHz, however.

As with any benchmark comparison, caveats abound. A benchmark only reports how fast a given CPU in a given chip runs the specific benchmark when compiled a certain way. Any conclusion about how well any other program will run on the particular chip is a matter of interpretation, as is any comparison among processors unless conditions are identical. Although the above scores are based on use of the GCC complier, the version used varies as does the compile flags selected. The Atom scores, for example, are for CoreMark compiled with GCC 4.4.1 and O2 optimization and were 5.8% faster than runs when GCC4.3.3 were used. The Cortex-A8 score is based on GCC 4.3.3 with O3. With O2 and GCC 4.3.3, the A8 achieves only 2.2 CM/MHz. Despite the variations in compiler and optimization, the overall conclusions are clear: Cortex-A9 and MIPS 74K will outrun Atom on CoreMark.

Whether CoreMark is representative of typical applications is another question. Designed to test the integer core of a CPU, it is small enough to fit in L1 cache. Although it is not based on a specific application, it tests common functions, such as matrix manipulation to test multiply-accumulate and other math operations, linked-list manipulation to test pointer-manipulation performance, and state-machine operation to test data-dependent branches. Thus, we believe it to be representative of inner loops of general integer code.

It is far better than Dhrystone. Long used as a benchmark, Dhrystone has well-known shortcomings. The leading one is its reliance on string manipulation, which is atypical of most software and which is implemented in C libraries. The latter is critical because processor makers can (and do) create highly optimized string libraries with the sole aim of scoring high on Dhrystone. Using regular libraries, it is impossible to reproduce vendors’ scores for Dhrystone.

A further issue with Dhrystone’s dependence on string manipulation is that the x86 instruction set includes string operations, whereas MIPS and ARM instruction sets do not. Unsurprisingly, Atom does well on Dhrystone but shows weak performance on CoreMark.

Interestingly, Dhrystone scores vary more than CoreMark scores. ARM rates Cortex-A5 at about 64% of the per-cycle Dhrystone performance of Cortex-A9, yet the A5’s CoreMark performance is relatively better at 72% of that for the A9. The MIPS 24K and 74K illustrate the same phenomenon. The 24K is 70% as fast per-cycle as the 74K on Dhrystone but 92% as fast on CoreMark. We expect the realized benefits of superscalar microarchitectures, compared with their single-issue counterparts on real code to be similar to what CoreMark shows.

CoreMark supports runs with multiple CoreMark instances running simultaneously. Because CoreMark is L1-resident, performance scales perfectly linearly with the number of CPUs making comparison of multicore implementations rather uninteresting. Analyzing the results of multithread processors, however, is more enlightening. The MIPS 34K and Atom both show big throughput gains from running two instances of CoreMark simultaneously. The 34K scores 26% higher than its single-thread counterpart, the 24K. Taking advantage of the second hardware-supported thread on Atom boosts performance 56%. With 16 stages, Atom has a longer pipeline than the nine-stage 34K. Atom thus has bigger pipeline bubbles caused by loads and branch latencies for a second thread to take advantage of. A more sophisticated microarchitecture with features such as out-of-order and speculative execution and advanced branch prediction would likely show lesser gains from hardware multithreading. Code with many cache misses would show greater gains. Any gain less than 100% per thread, however, represents a slowdown in per-thread performance. A CoreMark thread on the 34K, for example, runs 37% slower than if it were to run on the 24K.

In our latest report on CPU IP, we look further at the midrange and high-end CPU cores from ARM, MIPS, and others, including both household names and obscure companies, such as IBM and Beyond Semiconductor. The report compares not just performance but also die area, power consumption, and microarchitecture features. As the above comparison highlights, levels of CPU performance that were once the province of PC processors are now available for system-on-chip designs. When combined with the latest DSP, video decoding, and graphics technology, these CPUs imbue consumer electronics and communications systems with capabilities far beyond what seemed conceivable a few years ago. --Joe

Joseph Byrne, senior analyst

Innovation Disrupts FPGA Duopoly

We estimate that 86% of the market for programmable logic devices (including FPGAs) is controlled by Xilinx and Altera. Further, this $3.5 billion market is about the same size as it was 10 years ago and is projected to grow at a low rate of 3.5%. Both of the leading vendors reap high profit margins and seem content with their current revenue levels. High-end FPGAs, which can sell for more than $1,000 each, provide much of the duopoly’s margins. Consequently, each company carefully manages its product lines to ensure that low-end products do not encroach into high-end applications.

Lattice exploited this environment by offering midrange FPGAs with some high-end capabilities but with lower prices. Lattice had some success with this strategy until the market leaders created their own midrange FPGAs. Xilinx and Altera could easily match these products because Lattice offered little technology differentiation.

Technology from two new companies, Tabula and Tier Logic, poses a greater threat to the FPGA duopoly. These startups have fundamentally changed the FPGA silicon architecture to offer midrange and high-end FPGAs at low-end prices. If Xilinx and Altera attempt to follow suit using their current architectures, they would ruin their margin structure. Instead, the market leaders are likely to use other tactics to discredit these startups, prevent them from gaining a foothold at major accounts, and increase the financial barriers for each to be successful.

In the long run, scare tactics are unlikely to work, because OEM customers would dearly like to see a strong third supplier instead of declining competitors such as Lattice. These tactics may buy time, however, for Xilinx and Altera to improve their architectures to match the lower cost offered by these innovative startups. One way or another, FPGA customers are likely to see lower prices and more fundamental innovation than they have seen in the past decade. Although these changes will reduce margins at FPGA suppliers, they are also likely to enable true growth in the FPGA market through greater displacement of the $20 billion ASIC market. --Jag

Jag Bolaria, senior analyst

Additional coverage of FPGAs can be found in our report "A Guide to FPGAs for Communications."

For analysis and news on devices for networking and communications, subscribe to our free newsletter, Linley Wire