Lower power consumption
Wednesday, June 17, 2009
Power consumption has been a big issue in Yonah's creation. A lot of attention has been paid to the leakage current -- the power taken when the chip isn't doing any processing at all. Although this is typically a few billionths of a watt per transistor, with 151 million transistors it becomes a major issue. Various techniques can improve the situation, such as switching off power altogether to idle areas of the chip and changing the transistor geometry and composition to improve the insulation between its components. Intel says that it's now designing for power consumption in the same way that it has always designed for minimal chip area and most effective timing, but that deciding to do this came about almost by accident: it found big power advantages while designing for aggressive area reduction just as the market started to worry about watts.
The cores themselves have a complex series of power saving options: Active, Halt, Clock Stopped, Sleep, Deeper Sleep and Enhanced Deeper Sleep, which successively push the chip deeper into a coma from which it takes successively longer to return. Each core can slip into the appropriate mode independently of the other; dynamic power management also watches the frequency and power trade-offs for portions of the cores. The cores can finesse their frequencies as appropriate, but don't have independent voltage control; Intel says it looked at this, but that the form factors for multiple off-chip voltage regulators and the complications caused by routing split high-quality power feeds across the circuit board made this an unappetising option.
Off-chip factors haven't been ignored. More precise thermal monitoring of hotspots on the chip should mean that power isn't wasted spinning the cooling fan before it's needed. The chipset can also be safely powered down when the CPU is asleep through a process rather cutely called Dynamic Bus Parking. The chipset itself can abandon main memory to self-refresh mode during enhanced deeper sleep; it can minimise the amount of memory accessed by the display, dim the backlight automatically and compensate by lightening the LCD's pixels, and read an external sensor to reduce the brightness when the computer is being used in low light levels.
Cleverer cache
Back in the CPU, the Level 2 cache is a major part of Yonah's design. The 2MB of L2 cache is dynamically allocated between the two cores -- the busier core gets more cache. If one core is idle, the whole of the L2 cache becomes available to the other; if both cores are idle, then the cache can go into one of a number of sleep modes, including one where some or all of the data is written out to main memory and some or all of the cache is physically turned off. Improvements to the caching algorithms and deeper write buffers also reduce expensive main bus traffic, which would otherwise reduce performance and increase power consumption.
Both cores share a single bus into the cache, and data that's shared between processes on different cores can be accessed by either core without creating main bus traffic. This is unlikely to be much of an issue in typical notebook use, but is an example of an architectural feature that may become more important as the chip design filters into servers. It also compensates somewhat for the lack of a dedicated interprocessor bus.
The Core Duo processor's 2MB of Level 2 cache is dynamically allocated between the two CPU cores, which share a single bus into the cache.
Yonah has some good old-fashioned logic and arithmetic improvements too. It can bundle more instructions together for simultaneous handling -- both ordinary and SSE multimedia -- and do some operations in one micro-op that took four before. It's got deeper pipelining in the floating-point multipliers and more intelligent integer dividers that can do 32-by-32-bit divides in between a half and a quarter of the cycles required by the previous chip. Intel says that software profiling suggests that a lot of divisions will be helped by this: it might be a peculiarly obscure part of the chip's design, but the engineers are rather proud of it. Finally, there are more integer arithmetic units and ten new SSE3 instructions for media handling. In all, a fairly respectable smorgasbord of minor novelties that should lift overall performance across the board.
Yonah's designers are keen to emphasise that dual-core fits neatly into the hyperthreading story, but with higher performance than hyperthreaded chips because both threads are entirely independent. However, most of the headline performance issues with hyperthreading have been cache issues, where a wide-ranging housekeeping thread skips through large amounts of main memory to the extreme detriment of other processes. Dual core by itself doesn't help here: it could be that the improved dynamic cache allocation logic in Yonah will lead to fewer severe conflicts, but that remains to be seen
Outlook
Intel claims that the third quarter of 2006 will see two major events: the point at which dual core shipments overtake over single core, and the point at which 65nm achieves parity with 90nm. Mobile processor shipments are also increasing rapidly after languishing at around 20 percent of all processors: growth kicked off in Q2 of 2004 and is now at around 33 percent with 'not long' to go before 50 percent is reached, according to Dadi Perlmutter, general manager of Intel's mobility division, when he talked to CNET.com.au sister site ZDNet UK. He also said that the company is 'at least three quarters ahead of the competition' in 65nm, and 'one to two quarters ahead' in dual core on mobile platforms.