Editor's note: The author of this viewpoint is the chief scientist at MonolithIC 3D Inc., an IP that markets technology for fabricating three-dimensional ICs which differs from Intel's tri-gate technology.
Rivalries between companies have a charm of their own. For many years, Intel v. AMD was the talk of the town, and then it became Microsoft v. Google. The most interesting rivalry today is, of course, Intel v. ARM. After Intel's tri-gate transistor announcement, I was therefore not surprised to see these news articles:
At the moment, it certainly looks as though ARM will go planar at the 22-nm node, while Intel will go tri-gate. To judge if tri-gate offers Intel a significant advantage, a few questions need to be answered:
• Intel announced that their 22-nm tri-gate transistor consumed 50 percent lower power when compared to their 32-nm planar transistor. But what are the power savings for a 22-nm tri-gate transistor when compared to a 22-nm planar transistor?
• How much chip power can one save by using a 22-nm tri-gate transistor in a microprocessor instead of a 22-nm planar transistor? Is it 10 percent? Or is it 30 percent? Or is it 50 percent?
It is not difficult to get estimates for these. Let’s take a look.
Figure 1: Transistor I-V characteristics shown in Intel’s press announcement.
Intel showed some detailed transistor I-V curves in their press briefing. These are reproduced in Fig. 1. Using this data, you get the information shown in Fig. 2. You’ll notice that the 22-nm tri-gate transistor can give a 140mV supply voltage reduction compared to the 22-nm planar transistor. The 22-nm tri-gate transistor also provides a 50 percent power reduction compared to the 32-nm planar transistor, but this advantage drops to 19 percent when compared to the 22-nm planar transistor.
Figure 2: Transistor-level comparisons of tri-gate and planar transistors.
I then used IntSim, an open-source IC simulator, to estimate benefits of tri-gate transistors at the chip level. IntSim has models that describe various aspects of a modern-day chip, and its results show a good fit to actual data from past Intel microprocessors. For more details, please refer to Fig. 3 and the original paper about IntSim at the 2007 International Conference on Computer-Aided Design (ICCAD). A GUI-based version of IntSim is available at this link (for free use).
Figure 3: IntSim, a chip simulator, was used for this analysis.
For this study, I considered a 1-GHz mobile logic core built with either (1) 22-nm planar transistors, or (2) 22-nm tri-gate transistors. Since Intel presented only relative numbers for transistor performance, I took numbers from the International Technology Roadmap for Semiconductors (ITRS) and scaled them based on Fig. 2.
Figure 4: Power savings estimated with IntSim.
Results from IntSim are shown in Fig. 4.
• The 140mV supply voltage reduction enabled by the tri-gate device is useful, since it saves both clock and wire power. Fig. 4 indicates a 28 percent reduction in clock power and wire power.
• Transistor drive resistance, which is proportional to the ratio of supply voltage to drive current, is reduced with the tri-gate transistor. Hence, it is easier to drive wire capacitance and gates can be made smaller for the same performance target. This, coupled with the transistor power benefits shown in Fig. 2, gives a 28 percent power reduction for logic gates in the design.
• Repeater power goes down by 32 percent due to the better-quality transistors.
Overall, ~28 percent power reduction can be obtained by using a tri-gate transistor instead of a planar transistor for a 22-nm microprocessor core. This is quite significant. Kudos to Intel’s technology team!
Implications to the Intel-ARM tussle
As an engineer, I really like the fact that Intel is taking the tri-gate transistor into manufacturing—it is a fantastic technical achievement.
But will it play an important role in the Intel-ARM tussle? I don't think so. Let me explain why.
There are several variables involved in the Intel-ARM equation:
• ARM has momentum in the mobile space. I've learned, in my past, that displacing a technology or product entrenched in the marketplace is very difficult.
• Intel's x86 architecture is CISC. Advanced RISC machines, better known as ARM, use a RISC architecture. RISC architectures have historically given higher performance per watt than CISC in the mobile space. Can x86 bridge this gap?
• Intel chips are one process generation ahead of ARM chips. This is a valuable advantage.
• ARM chips are typically made in low-cost fabs in the Far East. For example, fabs in Taiwan are known to be 20-50 percent cheaper than fabs in the U.S., at the same technology node. This is due to additional government incentives, tax breaks, lower building costs and lower labor costs. All these years, having fabs in the U.S. did not impact Intel, since its only competition was AMD, which itself had not-so-cheap fabs in Europe. But while competing with ARM, low-cost fabs could be important.
• There are far more suppliers of ARM chips than x86 ones. Customers like competition as it keeps prices down.
• Who will be the first to "productize" other breakthrough technologies, such as 3-D stacking of DRAM atop logic, and monolithic 3-D? Intel or ARM? Some of these technologies could provide more benefits to mobile chip power, performance and die size than tri-gate.
• Intel has a lot more resources than players in the ARM world - this could be useful, especially in this day and age when designs cost $100M, fabs cost $5B and process R&D costs $1B.
• Will Microsoft execute on its goal to get Windows 8 on ARM? How soon and how well will it execute? This is a crucial business issue for laptops and netbooks, but may not impact the smart-phone world much.
Bottom line: The tri-gate transistor is a major engineering achievement, but I doubt if it will play a key role in the Intel-ARM tussle. In my opinion, many of the points mentioned above are more important.
Dr. Deepak Sekar is the chief scientist of MonolithIC 3D Inc. Author or co-author of a book, an invited book chapter, 15 publications and 55 issued or pending patents, he serves as a program committee co-chair at the International Interconnect Technology Conference.
What I do not see from Fig.1is that at what drain voltages (Vd) the three Id/Vg curves are taken. Assuming that the upper Tri-gate blue curve and lower blue curve are taken at Vd = 1V and Vd = 50 mV, respectively, the DIBL is roughly estimated to be ~110 mV/V and the subthreshold swing (SS) is ~85 mV/decade. For FDTri-gate transistors DIBL, SS and Vt change are expected smaller. The Vt seen here is induced by large DIBL meaning the build-in source barrier is lowered by the drain voltage, resulting in lower Vt and high leakage current. Notice that ~ 0.2mA/um difference between the saturation current (Idsat) as shown by upper blue curve compared with the Idlinear shown by lower blue curve at Vd = 1V for Tri-gate are unusually small. Intel shows 32nm planar Idlinear as indicated by black curve is virtually the same as the Tri-gate Idlinear at Vd = 1V, but dose not show the corresponding Idsat for 32nm planar transistor in Fig.1. However, based on Fig. 3 of Intel’s published IEDM 2008-p943, the 32nm planar Idsat at Vd = 1V is 1.55mA/um that is significantly larger than the Idsat of 1.0mA/um for FDtri-gate transistor. Similarly, TSMC 22nm FinFET built in bulk Si substrate published in IEDM10-p600 shows Idsat equal to 1.2mA/um at Vd = 1V which is significantly smaller than Intel’s planar 32nm Idsat of 1.55mA/um.Therefore, both Intel FDTri-gate and TSMC FinFET at 22nm show the same transistor performance degradation significantly worse than Intel’s planar 32nm, and may need additional fins added in order to boost the transistor performance.
I expect the planar 22/20nm will show significantly higher transistor performance, but higher DIBL and sub-threshold leakage compared to Tri-gate and FinFET. Intel should publish its FDTri-gate device data as done for its planar 32nm to avoid the guess work
Actually it's all very simple: Intel's problem is a that it is stuck in its own paradigma. They have grown 40years+ in the paradigm that x86 is the best architecture and that the IT world belongs to x86. Intel firmly believes this will never ever change, and behaves accordingly. But, in the IT ecosystem appeared a new beast called ARM, with new characteristics, two of the key ones being power efficiency and large licensing base. Now this new beast is starting to challenge the x86 lion, little by little depriving the lion from its "gazelle meat". Would Intel accept that x86 is a thing of the pas and start working on ARM solutions? Otellini would tell you beating his chest with his fist: Never! And that's company ego. And the bigger the ego, the harder the fall...
I wish Windows 8 for ARM had an x86 emulator in order to be able to run all the x86 legacy programs...(today Window 8 for ARM does not support backward compatibility with x86) but it seems more likely that in a couple of years most typically needed programs will have android/iOS versions of them and windows will progressively sink together with Intel. Welcome to the GoogApplarm era!
Very good blog Deepak. I think it would help to split the criteria into quantitative and qualitative ones. The former can be used to make straightforward analytical techniques as you did with power consumption. The latter are more subjective in nature. I personally cannot dismiss the ~28% power advantage of Intel's trigate technology. On the other hand, the 20-50% cost advantage of ARM is also a major advantage. The real advantage however will not come from bare technology alone but from a clear advantage in the final product. The question is then: what are the key comparative advantages when it comes to products, be it in the server or mobile arena? Are there any thresholds in terms of product cost, battery life/power etc. which would make a product a success of a failure? We should then work backwards from these to see what are the corresponding thresholds in terms of bare technology.
@semi111 : you've nailed it. Intel has a cash cow in their process technology, they're addicted to it. So their single trick is more bovine than equine. Those high margins make them vulnerable to disruption. Something that Andy Grove understood very well, but seems Otellini does not.
I agree with the author that Tri Gate is a major engineering achievement-though the idea has been around many years, but putting it in production still a big deal (BTW it took the industry 15 years to implement HKMG). Intel always leading the industry with their transistor technology like USJ, selective epi S/D, HKMG etc. All these great technical advancement keep Intel's Margin high in their core business the Microprocessor and especially the server market (Some people call Intel one trick pony).
On the other hand ARM has done wonderful by putting a great platform for low power using a plain vanilla process with a great success in the low power applications. Please check how many Ipad were sold with A4 or A5 and not an Intel's chip.
@Jay: I read David Kanter's article just now... it looks like guesswork to me. And he guesses 10-20%, not the 28% number I calculate above. Like I said, I know some people who worked on the tri-gate/finfet project at Intel, and they tell me they aren't aware of a scientific paper on this subject.
@iam_girish: You would choose Vdd and Vt based on your activity profile, with the result that leakage is not more than 30% of the total power... pls. check this classic paper on the subject: http://portal.acm.org/citation.cfm?id=368755 But based on the calculations above, tri-gate alone may not be enough to make up for 50% dynamic power difference and 10x leakage power difference between two chips...
I have heard that Intel processors leakage power contribution is ~10x that of ARM, and active power consumption (switching) is ~50% higher than ARM.
This tri-gate transistor technology seem to reduce only active power. Most of the mobile devices are in active mode only for 10-20% of the time, 80% of the time they are in standby/sleep - limiting leakage is key.
Power gating using high VT devices or back biasing etc.. or better architecture with low leakage/high Vt transistor is key for success in mobile industry.
I saw an estimate of 10-20% in David Kanter's take on Intel's 22nm 3D transistors and assumed there must be literature support for it:
Intel's 22nm Tri-Gate Transistors
"Even taking Intel’s estimates conservatively, that suggests a performance/watt advantage of 10-20% for power optimized chips versus a planar 22nm process."
I asked some buddies of mine at Intel if there is public literature quantifying chip power savings of their tri-gate transistor vs. their planar transistor. They said no. My colleagues in the transistor community haven't seen any of these papers either. Would appreciate your sharing these references with me, if they exist.
Reg. ARM RISC vs. Intel CISC, it is a whole new analysis. Will try to write about it on my blog one of these days...
David Patterson, known for his pioneering research that led to RAID, clusters and more, is part of a team at UC Berkeley that recently made its RISC-V processor architecture an open source hardware offering. We talk with Patterson and one of his colleagues behind the effort about the opportunities they see, what new kinds of designs they hope to enable and what it means for today’s commercial processor giants such as Intel, ARM and Imagination Technologies.