Every once in a while, you are forced to do something that is annoying to avoid yet more annoyance. Over the last week, we have gotten so many questions about Intel’s 22nm process, we feel compelled to write this article. (Editor’s translation: Charlie is unavailable for calls as he is on an airplane to Taipei and so this is provided as a courtesy.)
The questions come from the technical consulting side of SemiAccurate, it has been deluged about potential problems with Intel’s 22nm process, 22nm ramping, and similar prophecies of impending doom. Sure enough, there was one source for all of this false information, and it was a financial report from a major analysis house. Instead of blowing over, the questions kept coming, so here is our technical consulting speech on the topic for absolutely no charge.
We shall avoid naming the note directly because SemiAccurate has not had the chance to talk to the author, but lets go down the ‘facts’ one by one. The short version is that there are a few correct assertions, a few stunningly inaccurate ones, and the overall conclusions drawn are simply laughable. The central point to the note is that Intel’s 22nm process has serious problems, and 14nm is in deep trouble too.
The first assertion is that Intel is having problems ramping its 22nm process. Why? “We now believe that the six-month delay in the Romley server launch was due to issues with Intel’s 22nm process and not a chipset problem.” This does indeed sound serious, a six month delay to Intel’s cash cow is a big problem for Intel, the only thing worse is a process issue.
Except that it is factually inaccurate on not just one, but two fundamental grounds. First, Romley is the successor to Gulftown, the server platform that used a 32nm Westmere-EP as the main CPU. Romley uses Sandy Bridge-EP as it’s main CPU. Westmere-EP and Gulftown were launched on March 16, 2010, Q1/2010 in Intel speak. Servers, like most other Intel parts, are usually on a one year cadence for replacement. So when did Romley and Sandy Bridge-EP launch? Was it March of 2011 as a one year cadence from Westmere-EP would suggest? No, it was March 6, 2012, or Q1/2012 in Intel speak. That is ten days off of 2 years, no 18 months as a six month delay would indicate.
If you want to be pedantic, you can point out that Sandy Bridge-E, the single socket desktop variant of Sandy Bridge-EP did indeed launch on November 14, 2011, or Q4/2011 in Intel speak. That is much closer to a six month delay vs an expected March 2011 launch. One problem with this line of logic is that the desktop (not code named Westmere-E at the time there was no -E line, just -EP) launched in i7-970 guise in Q3 of 2010. Whatever way you look at it, Romley/Sandy Bridge-E/Sandy Bridge-EP is a year late, not six months.
It did have chipset problems, as we reported before launch. If you doubt that, look at the Sandy Bridge-E specs, the chip has PCIe2 while Sandy Bridge-EP, a later stepping of the same chip, has PCIe3. The chipset connects over a link that is essentially PCIe’s physical layer, so you can’t get more direct than that as proof of chipset issues.
Romley is NOT Ivy Bridge, it is Sandy Bridge-EP
Then there is the biggest problem with the premise ‘Romley delay = 22nm problems’ premise, Romley, AKA Sandy Bridge-EP is not 22nm, it is 32nm. Really, look at the roadmap picture above, then take a look at this link, specifically the process listed. If there is some brilliant leap of logic that ties a one year, not six month, delay of a 32nm platform to problems with 22nm, I am not seeing it. On top of that, SemiAccurate has had direct conversations with the CPU, chipset, and platform architects of Romley, and they all indicated that the delay over an expected Q1/2011 launch had nothing to do with process problems, the very complex architecture of the platform just took a long time to finish. We have no explanation for the basic mistake of calling the Romley CPU a 22nm part, it simply is not, never was, and never will be built on that process.S|A
Latest posts by Charlie Demerjian (see all)
- Intel kills off the 10nm process - Oct 22, 2018
- 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