There's hype, and then there's reality. When it comes to leading-edge chips, the chasm between the two is rather wide. Reverse engineers tear these technologies apart for a perspective that often diverges far from commentaries offered by the chipmakers and other industry mavens.
Last year was full of hype about the impending release of 65-nm devices, particularly in the microprocessor and consumer arenas. Intel shipped its dual-core Yonah Pentium M processors to system builders in the last weeks of 2005, readying them for a product launch at the International Consumer Electronics Show (CES) this month.
AMD is less aggressive in process terms. It's aiming to launch its 65-nm devices during the latter half of this year, but the company also is ramping them up in its 300-mm Fab 36. AMD has been in the dual-core business for a while, so it's under less pressure to push out new parts. And, it's also continuously improving its 90-nm parts.
Speaking of multicore processors, IBM has been rolling out its eight-core Cell processor infrastructure. It should be shipping in volume by the third quarter of this year for the Sony PlayStation 3. Sony's launch date for the PS3 remains vague. The most recent date we've heard is early second quarter in Japan.
Given that IBM also has a three-core processor in Microsoft's Xbox 360 and a dual-core in the upcoming Nintendo Revolution, it appears to have a pretty good lock on the gaming market for the next couple of years (Fig. 1a).
IBM has been pushing the Cell in other areas. It also will be available in Mercury Computer Systems' Dual Blades, supposedly in the second quarter of 2006. This is a 90-nm chip, so its fabrication should commence as scheduled. Debugging it, however, may be another story.
We have another interest in the PS3, since we hear that the Nvidia-designed graphics processor will be from the Sony/Toshiba fab using 65-nm technology. The graphics chip from the PS2 was the first 90-nm chip in volume production, but it looks as though Intel will be the first off the mark in 2006.
Yet it's interesting to note that Matsushita has quietly announced that it will jump straight from 130 nm to 65 nm. The company is shipping DVD controllers from its new 300-mm fab.
ATI also has done well in the electronic gaming market with design wins in the Xbox 360 and the Revolution (Fig. 1b). The latter chip likely will be 90 nm. But ATI reportedly is moving to TSMC's interim 80-nm process next year, with its Radeon X1300CE (R505) and RV580 GPU, possibly launching in the first quarter. Its X1800 was late in getting to market this year, so maybe we should be looking at the second or third quarter for this launch.
Fujitsu is at the forefront as well. Its 90-nm process used for the Transmeta processor had transistor gate lengths of about 40 nm, putting the company halfway to the 65-nm node compared with other fabs (Fig. 2). It rolled out its 65-nm process last fall, but as of yet we have not identified any users contemplating this process for 2006. Sun will work with Fujitsu on its dual-core SPARC64 VI, but that will be 90 nm, evolving to the 65-nm VI+ in 2008.
In the wireless arena, Texas Instruments launched its OMAP2 processor series in 2005. TI likely will migrate to the 65-nm process currently ramping in its Dallas fabs. Qualcomm also has unveiled its first 65-nm chips for cell phones (out of TSMC), including its MSM6800, MSM6280, MSM6260, MSM6255A, and MSM6245 solutions for cdma2000 and WCDMA third-generation phones. We find it typically takes nine to 12 months to actually get the products downstream, so we would expect to see these in phones in the third or fourth quarter of 2006.
The majority of products in this segment come from the fabless sector. Xilinx, Altera, Lattice, and Actel all use foundries. Xilinx and Altera are reportedly sampling 65-nm parts out of their foundries (Xilinx with Toshiba and UMC, Altera with TSMC).
Lattice is currently at the 130-nm node with its EC, ECP-DSP, and XP series of products, with the SC 90-nm parts expected in 2006, all from Fujitsu. Even though Lattice extended its deal with Fujitsu to 65 nm, we don't expect products before 2007. Fujitsu's 130-nm process is interesting since it seems to be the only production process using Dow's SiLK spin-on low-k dielectric material.
Actel, the fourth largest FPGA vendor, seems to be targeting the high-reliability and space markets. So while the company isn't predicting any changes in the technology node, we're likely to see more space- and automotive-qualified products coming from it.
2005 could be dubbed the Year of Flash. We have witnessed incredible growth in this segment, due largely to the consumer demand for MP3 players, cell phones, and digital cameras. Flash technology also has become a driving force in process advancement.
Samsung introduced its 73-nm 4- and 8-Gbit NAND flash memories in 2005, with a cell measuring an incredible 0.0225 µm2(Fig. 3). Also, there seems to be every likelihood that we will see the 50-nm 16-Gbit chip in 2006. Samsung had 50% of the sales in the third quarter of 2005, dominating the market for the year, and it will do so again in 2006.
The flash market is so successful, all the other sizable companies are trying to get in on the act. For example, look at the recently announced Intel-Micron joint venture. Micron has a 2-Gbit part in 90-nm technology. The company likely will push to 70-nm higher densities in 2006. It's switching on flash in its 300-mm fab in Manassas, Va., which probably accounts for its impressive 400% growth in the third quarter of 2005. Even so, that made Micron the fifth player, with 3.4% of sales.
Toshiba had been the second-string NAND flash supplier in 2005, with 23% in the third quarter of 2005. In our analyses of Apple products, we have seen almost as many Toshiba dice as Samsung. Toshiba has put more emphasis on multilevel-cell (MLC) flash, with 2 bits/cell, but at the cost of more complex peripheral circuitry and strict gate engineering to address capacitive coupling issues. However, MLC is well suited for low-bit-cost, consumer digital audio applications such as the Apple iPod nano. Samsung led the market with its single-level cell (SLC) and tried to downplay MLC, but it is in the roadmap now, and the company is shipping products.
Hynix, the third player, is up to speed on the technology. The company is shipping 90-nm 4- and 8-Gbit parts, and it claims to be readying a 70-nm, 16-Gbit part for 2006. Renesas was fourth in the third quarter yet is reported to be suspending MLC 8-Gbit NAND flash production. But following an alliance with Grandis, it announced that it will be shipping 65-nm MRAM parts.
2006 will see the shift from double-data-rate (DDR) to DDR2 technology, a shrink from 90- to 80-nm processing, and the start of 1-Gbit chip shipments. At the module level, fully buffered dual-inline memory modules (DIMMs) will grab a noticeable market share.
Again, Samsung is at the leading edge in the DRAM technology race (Fig. 4). It uses an innovative recessed transistor in the cell, and this will shrink to 80-nm design rules. Also, Samsung had 30.6% of DRAM sales in the third quarter of 2005, followed by Hynix with 16.6%, then Micron at 15.3%, Infineon at 13.5%, and Elpida with 7.2% market share.
Hynix also is shipping 90-nm DRAM, and it plans to move to 80 nm. Micron is shipping 110-nm parts, but the cost pressures on DRAM mean that we will likely see 95-nm products soon. Micron has an innovative 6F2 cell instead of the usual 8F2 cell, so it can achieve high densities with a larger technology.
Infineon, the only trench-DRAM maker in the top five, has a well-publicized plan to spin off its memory group. In product terms, it is shipping 90-nm and has 70-nm parts in development. However, it seems unlikely that it will do more than sample those in 2006. Infineon has has come up with innovative technologies to increase packing density, such as square trenches and knobbly high-k capacitor dielectric.
Elpida also is in full 90-nm production and anticipating the shift to 80-nm. This will probably occur in the second half of 2006. In addition to the top five's in-house production, there is an increasing amount of outsourcing in the DRAM business, and this probably will continue to evolve. Infineon has relationships with Nanya and its Inotera joint venture, and Elpida has capacity with Powerchip and SMIC in China.
Camera phone sales are expected to reach $500 million in 2005, taking two-thirds of mobile phone sales.1 Simultaneously, we have seen the shift from 1.3- to 2-Mpixel cameras. In 2006, 3-Mpixel cameras will come on stream. The technology also is migrating from charge-coupled devices (CCDs) to CMOS, which is becoming ubiquitous in phones.
So like flash, manufacturers and foundries are offering image sensor processes, and CCD companies are getting into the CMOS business. The most notable of these is Kodak's announcements of licenses to IBM and TSMC, with reported production starts of 3- and 5-Mpixel chips at both foundries. We should see these in downstream products in 2006, such as cell phones and still cameras.
Micron is a relatively new entrant in the imager market, beginning with a 1.3-Mpixel sensor. The company has just started to ship its MT9T012 3-Mpixel chip, with a leading-edge 2.2- by 2.2- µm pixel size. Micron also offers a 5-Mpixel version that targets still cameras.
Sony is the biggest-selling imager manufacturer, mostly in CCDs, although it also is increasing its proportion of CMOS sensors. The company had a well-publicized failure problem with its CCDs in a variety of cameras from many manufacturers in 2005, so it probably will have to make up market share in 2006. It has a 2-Mpixel CMOS imager on the market today, so it cannot afford to miss the 3-Mpixel opportunity, even though there appear to be no pre-announcements.
Toshiba held fifth place for sensors in 2005, but it has just scored a design win in Nokia's N90 smartphone (Fig. 5). We expect to hear more from the company in 2006. It introduced a 3.2-Mpixel sensor in the fall of 2005, so we expect to see it downstream in 2006, perhaps in another Nokia phone.
When it comes to still cameras, CMOS is making steady progress into territory formerly owned by CCDs. We expect this trend to continue. Canon already has introduced CMOS into its high-end SLR cameras, with the notable introduction of a 12.8-Mpixel, 24- by 36-mm sensor in its EOS5D model.
In the big-dollar volume wireless communications market, a prime driver for reduced prices has been the integration of as many functions as possible into CMOS. In the frequency range from 1 to 10 GHz, one function that has resisted integration is the front-end amplifier of wireless receivers. Known as the RF amplifier or low-noise amplifier (LNA), it operates at the RF carrier frequency.
Most LNAs have been completed with silicon bipolar silicon-germanium (SiGe) transistors or Group III-V semiconductor materials such as gallium arsenide (GaAs). Compared to standard CMOS, neither SiGe nor GaAs can be regarded as high-volume, low-cost semiconductor process technologies. Therefore, integration of the LNA function onto a CMOS receiver die would remove both a cost component and a separate IC package from the system. The frequency range of interest here, roughly 1 to 10 GHz, contains such existing and future huge volume markets as IEEE 802.11 applications, WiMAX, W-USB (wireless universal serial bus), and GPS in cellular phones.
In approximate terms, the advancement of process technology from 90 to 65 nm can be expected to improve the basic unity-gain frequency response ( ft ) of NMOS transistors from around 100 to 130 GHz. Many RF designs require a transistor ft of at least 10 times the application frequency. So the technology migration from 90 to 65 nm will give CMOS circuitry significant added speed capability in the commercially important frequency range of 1 to 10 GHz.
Traditionally RF/mixed-signal processes have lagged behind the leading-edge logic processes. But the cost savings to be gained by integrating the RF front end of a cell phone into CMOS may drive these products into 65 nm quicker than expected. Wireless USB, as another applications area, operates from 3.1 to 10.6 GHz, right in the range enabled by 65-nm technology. Add in the advantage of low overall power consumption, and we have powerful drivers to accelerate the technology shift.
As an example, TSMC expects its 65-nm process to double standard-cell gates, increase speed by 50%, and reduce standby power by 20%. Broadcom, Qualcomm, and Freescale all have run early silicon through TSMC's initial 65-nm Cybershuttle, and Texas Instruments claims it has shipped 65-nm wireless sample products. While we would not expect to see 65-nm RF chips downstream in early 2006, given the cost drivers, we could see some by the end of the year.
Even though 65-nm logic product launches may dominate the media, there will be plenty of interest in the other segments of the industry. One thing is for sure?2006 will be interesting inside technology!
Reference:1. Future Image, 10/27/05