Application processors duel

Convergence of functionality in handset designs is constantly raising the bar for value-added features, such as cameras, games, music playback and video capturing. It is estimated that the average cost of a handset will drop from about $125 today to $110 by 2008. In contrast, the average cost of the ICs within the handset will increase from approximately $40 today to $44.50 by 2008, as consumers demand higher functionality at a low price. A versatile application processor is key to achieving a small form factor, cost efficiency and ease of integration of new functions.

But application processors are limited in what they can accomplish, mainly because they cannot increase the system clock frequencies or speed up data movement to the memory. To compensate for these limitations, and to fulfill more advanced functionality, an increase in design and manufacturing process complexity is necessary. This has become a good indicator of the capabilities and maturity of the manufacturer. Texas Instruments, Qualcomm, STMicroelectronics, Samsung, Intel and Freescale Semiconductor are some of the application processor providers fighting for their slice of a $5.8 billion market.

The two main contenders for this market are Qualcomm and Texas Instruments. Both have found significant traction in the handset market. But Semiconductor Insights' Handset Design Win Subscription service indicates that Texas Instruments has been receiving more design wins, due in part to its wide range of products and nonproprietary coding, which gives it a larger OEM base than Qualcomm.

In contrast, Qualcomm uses a proprietary licensing technology that requires programmers to learn the system and has a narrower product offering. Texas Instruments is being designed into Motorola, Sony Ericsson, Sharp, Panasonic and Nokia handsets, while Qualcomm can be found in Audiovox and Samsung handsets. The only identified handset OEM that used application processors from both companies is LG.

This article will compare and contrast some of the structural differences (package type, die size, process technology, memory) and application features of the application processors offered by Qualcomm and Texas Instruments.

The focus will be on the 2003 parts from both Texas Instruments and Qualcomm, with additional information from 2002 parts as a comparison when it is relevant. From Texas Instruments we investigated the Omap1612 and a Nokia-packaged application processor. The competing Qualcomm parts are the MSM6250 and the MSM6300. All of the devices were created using a copper CMOS 0.13-micron process technology and were measured to have a minimum transistor gate length of 0.11 micron. In addition, all integrate an ARM9 processor and internal DSP.

The devices are created using the same process, so the die size measurement indicates the technical efficiency of the device with some comparable features. Qualcomm has the smallest die size with a measurement of 47.61 mm2. The Texas Instruments Omap1612 has the second smallest die size at 60.82. This provides an advantage to Qualcomm, since it has the potential to create more die from a wafer. But it has lost this advantage with the newer MSM6250, which has a die size of 72.25 mm2. This is likely due to the extra features included in the MSM6250, such as Qvideophone MPEG-4 two-way point-to-point video phone, USB host (on the go) and QVM Java J2ME hardware acceleration.

The next comparison is the number of layers required to create the processors. Texas Instruments requires only five metal layers, while Qualcomm requires six. Additional layers generally cause an increase in the power consumption of the device, which may affect Qualcomm's overall performance. And even when comparing the width and pitch of these layers, Texas Instruments, which required fewer layers, generally measured a lower width and pitch.

From a packaging perspective, the Texas Instruments processors are found within BGA packages with a relatively small form factor, measuring 144 mm2. Qualcomm has changed from using a 169-mm2 chip-scale package, which offers a smaller profile due to die positioning, to a larger-sized 196-mm2 BGA package. This is likely due to the fragile nature of the chip-scale package and the increased pin count of the MSM6250. Since these processors are being used in high-end handsets, the smaller package will take up less board area, providing an advantage for Texas Instruments.

While the die size gives a measure of advancing process technologies, it does not necessarily mean that the die is being utilized efficiently from a functional perspective. To compare the die area differences of a specific functional block, we compare the different amounts of die that the ARM cores use. The amounts vary somewhat, but in general, Qualcomm has ARM cores around one-third smaller than those in the comparable Texas Instruments devices. This is due to the enhanced features that Texas Instruments includes in its ARM core, such as a 16-kbyte instruction cache, 8-kbyte data cache, support for 32-bit and 16-bit instruction sets, data and program memory management units (MMUs), two 64-entry translation look-aside buffers for MMUs and a 17-word write buffer. The smaller size, and thus lower cost, of the ARM core structure gives an advantage to Qualcomm, but this weighs against the added features provided by Texas Instruments, which are designed to provide additional functionality to the system.

Another comparable function is the DSP core. The Omap1612 uses a single Texas Instruments TMS320C55x DSP, while Qualcomm's processors contain dual QDSP4000 DSPs. Even with the Qualcomm parts using two DSPs, the total die area occupied by the DSPs is approximately five times smaller than the Texas Instruments parts. But the Qualcomm DSPs are assumed to have less memory.

A comparison can also be made between the two companies with respect to the amount of embedded memory available within each device. In this area, Texas Instruments provides significantly larger amounts of memory, which can be used to increase the functionality of the device by decreasing the number of times the processor has to poll the memory to manipulate information. The devices from Texas Instruments offer significant advantages both in the amount of memory provided and the area that the memory takes on the die. This is interesting, since the application processors are using the same process technology. It means that one embedded memory technology is more efficient in terms of bit/mm2, highlighting a clear advantage for Texas Instruments.

The final consideration is the cost of the devices. A cost estimate is created based on information regarding the process, foundry, number of layers, wells, silicide and passivation, as well as certain Semiconductor Insights assumptions based on our experience in analyzing semiconductor devices. From these details, process flow can be predicted to determine the number of masks it takes to create a wafer. Semiconductor Insights has developed a good understanding of wafer process fabrication, allowing for an accurate cost estimate with a reduced margin of error.

After analyzing all of the data necessary to estimate a total cost, Texas Instruments shows a clear advantage for both 200-mm and 300-mm wafer sizes, with a cost less than $9.50 at a 200-mm wafer and under $7 at a 300-mm wafer. This is compared with Qualcomm's total costs of more than $10 at a 200-mm wafer size and close to $12 for a 300-mm wafer. Qualcomm's total cost at 300 mm increased because TSMC, when this cost analysis was completed, had not yet reached full production capacity at this size. Moreover, since Qualcomm is a fabless semiconductor company, it may be paying additional fees that Texas Instruments would not be required to pay.

Peter Di Paolo (peterd@semiconductor.com), technology manager for communications at Semiconductor Insights.http://www.eet.com

Click image to enlarge