Two weeks ago, Globalfoundries announced a ‘7nm’ process, but does it live up to the name? SemiAccurate takes a look at what is being promised and a clear picture is revealed.
If you remember our look at what the TSMC ’10nm’ node delivers and how it lives up to the name, you will see that that node is, umm, not 10nm in name, spirit, or anything else technical or otherwise. As reader Etienne Robert pointed out, the nm in the process names of late no longer stands for nanometer, it now stands for nanomarketing. Judging by what TSMC did at ’10’ and GF is doing at ‘7’, we think that these number are utterly meaningless and will soon become a quarterly marketing race to the bottom.
Although games around process numbers have been going on for a while, much of this started out at the 14/16nm node. If you recall both TSMC and the Samsung/GF processes were nothing more than their 20nm planar nodes with FinFET transistors slapped in, the back-end of line (BEoL) was the same. This meant the shrink delivered by the process was something like zero. Rather than take the honest road and call it 20nm FinFET or something similar, they called it 14 to intone a full shrink and the technical merit that it carried.
A Samsung engineer that SemiAccurate spoke to on the ‘shrink’ said that the most aggressive shrinks going from 20 to 14nm would be about 15% if the designer was well versed, had the right tools, and really wanted to spend the effort to extract every last mm^2 from the design. It was strongly suggested that the bulk of the designs ‘shrunk’ in this way could see about a 7% shrink with enough effort put in,
One thing to bear in mind that these are area shrinks, that means the liner dimensions shrink by the square root of that number so a 15% area reduction would be ~8%. This is well within, below actually, the rule of thumb number of 15% or so from a processes introduction to final updates. Given that, the ‘shrink’ from 20nm to 14/16nm is likely zero or a regression, it would have happened anyway without the ‘shrink’.
As SemiAccurate pointed out with the upcoming 10nm (We are sick of the single quotes thing, you get the point by now) processes, they aren’t. They aren’t even what a real 14/16nm process should have been and never will be. The only one offering something that lives up to the name with technical advances is Intel and their 14nm process, but given the latest set of messaging, they are hopping on the BS train too.
So back to GF and the 7nm process. What is it? Lets start with the money quote from the official press release, “GLOBALFOUNDRIES’ new 7nm FinFET technology is expected to deliver more than twice the logic density and a 30 percent performance boost compared to today’s 16/14nm foundry FinFET offerings.” We will assume that the use of “and” when saying twice the logic density and a 30% performance boost means what it says, and is likely the case in a real shrink. More than twice probably means low 50% range, if the first number was a 6 you can bet they would have called that out. So it looks like a real, honest, full node shrink. This is a good thing, right?
It would be if it wasn’t for the marketing trolls who are taking a technical specification and discarding it for headlines. Why do we say that? A real 50% shrink, aka twice the logic density is really tough, and 7nm is half 14nm right? Linearly yes but that is two full nodes when you look at the square of the linear shrink which is what the numbers represent. 14^2 = 196, 7^2 = 49. 49/196 = 1/4 or .25 which is another node down from 7nm. In GF marketspeak, that would likely be a 3.5nm node, call it 3nm to make it sound better. It is a joke, not a real 7nm process.
Setting aside the reality that it probably can’t actually draw anything <40nm, it does still get worse. Why? Because the 14nm base process they are comparing it to is really 20nm by any sane technical measure. Their 14nm was just a mature 20nm process with a new transistor, maybe a hair smaller but not close to what a real 14nm process should have been or even a half node from 20nm. What GF is saying is that their 7nm process is a real, true, honest shrink but not from a real, honest, true 14nm process, it is from a real, honest, and true 20nm process. What do we know that as? 14nm.
In the end the industry is slinging the BS harder and harder, it is not just Globalfoundries doing it, everyone has joined in the game. 14/16nm was not a shrink from 20 for anyone bar Intel but they all called it one anyway. 10nm is barely a half node from that but 7nm looks like a true full node. The only think they won’t point out is from what, 14nm ~= 20nm so 7nm = a real 14nm process for the engineers. I would complain about the BS marketing but the bad guys have won and now they are dragging the good guys into the mud. Sad times. S|A
Latest posts by Charlie Demerjian (see all)
- HyperX ships it’s 60 millionth enthusiast memory module - Oct 15, 2018
- Bittware/Nallatech water cools 300W of Xilinx FPGA - Oct 12, 2018
- More on Intel’s 10nm process problems - Sep 17, 2018
- Intel puts out another 14nm 2020 server platform - Sep 11, 2018
- Why Can’t Intel Supply Enough 14nm Xeons? - Sep 10, 2018