(Editor’s note: Part 1 can be found here. Part 2 here.)
The next premise of the note is that, “power consumption is higher than expected.” This is simply not true. Power is measured in two ways, idle and TDP, with idle being leakage dominated, and TDP is work/usage dominated. Idle mainly measures how efficient the transistors are at doing nothing, IE how close to zero watts they can get when they are supposed to be at zero watts, and TDP is dominated by how much power it takes to switch a transistor on or off, basically change state.
Measuring idle wattage of a CPU is hard, mainly because the measurements tend to be overwhelmed with power used by other parts of the system. A laptop CPU may draw less than 1W at idle, but the screen takes closer to 10W, the hard disk 7W, and the chipset 5W. Measuring the idle wattage accurately is almost impossible for a third party reviewer, but given the numbers on the transistor performance from Intel, 22nm seems to be a lot better than 32nm at leakage and idle power use. Anecdotal evidence supports this, but it is impossible for the layman to accurately test.
TDP is much easier to test, and here, 22nm is a clear win. Why? Look at the i7-2700 series, aka 32nm 4-core Sandy Bridge CPUs. Note that they have two listed TDPs, 95W and 65W, the 130W parts are a completely different chip, Sandy Bridge-E/EP, not Sandy Bridge-no-dash-lettered-suffix. Then look at the i7-3700 series, aka 22nm 4-core Ivy Bridge CPUs. They have TDPs of 77W and 65W.
Any reviews of both chips show that the 3700 line does about 10% more work on the CPU side, about twice the work on the GPU side, and does it for more than 20% less power than the 2700 family. How anyone can conclude that this is “higher than expected”, especially when the CPUs conform to the 77W TDPs that have been listed on Intel roadmaps for 6+ months. We can’t explain it in any logical manner, all the evidence is completely contrary to that conclusion. Simply put, Ivy does not use more energy than Sandy to do the same amount of work, it quite provably uses less energy to do more work.
The note goes on to point out that Ivy Bridge/3700 processors run hotter than Sandy Bridge CPUs. This is at times completely true. The 4-core Ivy Bridge is a 160mm^2 chip with 1.4 billion transistors. Sandy Bridge 4-core chips are 216mm^2 and pack 1.16 billion transistors in to that area. That means Ivy is .74x the size of Sandy, and has 1.2x as many transistors, or is a 62% of the prior generation’s size. Theoretically, it should have been closer to 50%, but no modern process shrink gets to the theoretical limits, so Ivy is looking quite good there.
The note however points to heat as a problem, and Ivy may run hotter, but that is quite simply not a problem. Heat is a meaningless metric for CPUs, at least as far as the end user is concerned. Intel will specify an operating temperature range that a device can work at, and as long as the device is in that range, the temperature is not a problem. If you have two cars, one has a coolant range of 0-220 degrees, and usually runs at 180 degrees. If your other car is meant to run from 0-260 degrees, and normally runs at 240 degrees, is that problematic because your first car would blow up at that temperature?
Ivy Bridge uses 20% less power at TDP in more than 25% less area, so the power density goes up by a considerable amount. Due to the different types of transistor structures, planar vs ‘3D’ the basic building blocks can have fundamentally different behaviors. That said, Intel has set 22nm products to run at a higher temperature than 32nm products. As long as the temperature stays in that range, you won’t have any problems.
That doesn’t seem to matter to the note’s author however who claims that “Sandy Bridge and Ivy Bridge runs hotter than its predecessor when overclocked.” This is likely true, and Sandy does overclock better than Ivy, but there are many things far more basic that will affect overclocking ability, speed path limitation being the key one. Sandy is a completely different design to Ivy, mainly because of the transistor changes. On a logical level, it may act the same way, but physically, it is not even close.
On top of that, overclocking is, by definition, running the chip out of its expected spec. It often includes upping voltages beyond listed ‘safe’ tolerances, and doing other things that are not normally allowed by Intel. Comparing two chips run out of design spec, and then ignoring many first order explanations to the observed effect to support a plainly wrong thesis is borderline dishonest.
Ivy Bridge runs within spec. It runs hotter than its 32nm predecessor while using less energy. Heat is a design parameter, and as long as it runs with the expected temperature range, it is working as designed. This is not a problem, is not related to 22nm transistors, nor is it the sign of anything more general than a choice Intel made that works as intended.
There are several more extremely technical points brought up by the note that we will skip because the explanations are long enough to make the above seem like a bullet point. Suffice it to say SemiAccurate disagrees with them all, and thinks they are fundamentally technically incorrect. If you are interested in the technical explanation on this level, please contact the author directly.
Back to the last issue, a non-technical one concerning the 17w ‘ultrabook’ Ivy Bridge CPUs. The note says, “Finally, it makes no sense that the company would launch desktop before ultrabooks processors. We think this points to power consumption issues with the 22nm process.”
We explained the delay in March, basically one large OEM changed their order from a common bin to a far less common bin. This in turn forced Intel to make changes, and delay the introduction of those parts until the changes had worked their way through the fab, and sufficient stockpiles of product had been made to assure supply to large OEMs.
SemiAccurate’s checks have shown those CPUs are shipping in volume, well before the rumored June delays, and devices bearing those CPUs will be out in volume very shortly. If you are making Laptops, you can buy these chips now in volume, no problem. Consumers have to wait a few more weeks for the official launch date, but an enterprising end user can find them now if you know where to look. There is no problem other than the usual scramble when a massive customer does a radical order change. The ‘delay’ of 17W Ivy Bridge CPUs is not indicative of any process problem, mainly because there is no delay, and no 22nm process problems. Just because you don’t see something happening doesn’t mean that it is not actually happening, but ‘analysts’ may have more in common with ostriches than I was brought up to believe.
Some people are just unwilling to believe that 22nm is available despite there being many examples of product with ‘3rd Generation Core’ CPUs in them readily available off the shelf. There are no shortages of any launched product that any SemiAccurate channel check has been able to find, and any ‘delay’ of 17W parts is there to ensure that these bins are in full supply when launched. This is called good product planning, not bad process engineering.
While the note in question is riddled with basic factual inaccuracies, very tenuous leaps of logic, and ignores the obvious technical explanations to other observations, there is one more problematic factor, the press.
TSMC has a PR problem, basically the company has process problems and instead of engaging the press, they specifically do not comment even if it behooves them to do so. What the company does do is use a paid source of disinformation to ‘prove’ their side of the story, facts to the contrary notwithstanding. This TSMC source does not disclose their funding or ties to the foundry, and that is probably a good thing for TSMC given the source’s behavior.
This source, lets call it TSMC’s Attack Turnip, with no offence meant to the otherwise upstanding root vegetables, has latched on to the above report as ‘proof’ that Intel is having problems on 22nm. While Attack Turnip claims to be a process and semiconductor specialist, he either blatantly and dishonestly ignores the above note’s problems, or simply does not understand them. Given SemiAccurate’s reading of Turnip’s previous analyses, we believe it to be both.
When this is used as proof about TSMC’s lack of problems on 28nm, it blindly ignores the fact that TSMC is a full process node behind Intel, and when they get to 20/22nm, they will still only have planar structures. TSMC will not have FinFETs until 14nm, 4 years behind Intel, best case for TSMC. By then, Intel will likely be on 10nm and 3rd generation FinFETs, or possibly carbon nanotubes.
In the end, most mainstream sites have chosen to ignore the initial note because of obvious technical flaws. That makes those who promote it stand out, and not in a good way. TSMC is doing themselves a lot of damage in the financial community with these foolish actions as well, all they are doing is highlighting their own flaws instead of promoting their strength. TSMC does have strengths, FWIW.
It seems that the technical press is wisely ignoring the original note about Intel’s 22nm problems. Unfortunately many other less technical people are not able to sort fact from fiction, or do not have the background to understand why the note is so error ridden. Never take any technical note at face value, and always verify claims with independent sources, it isn’t as hard as it may seem.
The short story is that Intel does not have 22nm process problems, the -LP processes are ahead of schedule by at least a year, and 14nm looks to be challenging and expensive, but no more so than the historical trends. In the end, nothing to see here folks, 22nm is moving along nicely, and so should you.S|A
Latest posts by Charlie Demerjian (see all)
- Bits about AMD’s Threadripper Next lead out - Jun 8, 2023
- AMD manages to make the Ryzen 7040U look worse than it is - May 4, 2023
- Is Microsoft looking to buy silicon expertise? - Apr 3, 2023
- Intel updates roadmaps at Analyst Day 2023 - Mar 29, 2023
- Intel’s Emerald Rapids has a big secret - Mar 27, 2023