13 ms·
Daily reminder: Intel is having various technical problems and arguably made some bad business decisions, but remember, semiconductor manufacturing processes ar
by segfaultbuserr 7y ago
Daily reminder: Intel is having various technical problems and arguably made some bad business decisions, but remember, semiconductor manufacturing processes are no longer directly comparable using the "x nanometer" number. A comparison should mention their technical differences.
"x nanometer" has became a technically meaningless trademark solely exists for indicating new generations of process for marketing since ~2009's introduction of FinFET. It has no actual relation to gate length, metal pitch or gate pitch. When some people say "5 nm" is the early-stage of nanotechnology, remember that the gate length actually stays at 25 nm since 2009.
GlobalFoundries' 7 nm process is similar to Intel's 10 nm process. TSMC and Samsung's 10 nm processes are only slightly denser than Intel's 14 nm in transistor density. They are actually much closer to Intel's 14 nm process than they are to Intel's 10 nm process. I'm not saying that Intel is 100% faithful, the 14 nm in the original ITRS roadmap was described by Intel as "10 nm". Also, DRAM chips are yet another different can of worms, the process "number" for DRAM chips is not directly comparable with CPU's "number" as well.
As conventional Moore's Law is reaching the end, the semiconductor firms spent a lot of efforts to make things as confusing as possible. I'm not a semiconductor engineer, I'm just an ordinary programmer, I don't fully understand the funny business going on since 2009, any correction is welcomed.
In other words, Intel still stays on the state-of-art of semiconductor manufacturing, in line with GlobalFoundries and TSMC, not significantly better or worse, in contrary to what the "x nanometer" number would make you to believe, but they do have a lot of production issues.
- throwaway2048 7y agoGlobal foundries' 7nm process was similar to Intel's 10nm, however Intel heavily revised their 10nm process to be approximately 1.3x less dense.
- segfaultbuserr 7y agoYou seem to know more about recent news better than me. I heard that Intel's 10 nm delay was mainly the result of its overambitious engineering and beliefs that they could do 10 nm via existing DUV without the troubles of moving to EUV, but the yield proved to be extremely low and basically a disaster. Is it true? What is their reenginnered solution in addition to making it 1.3x less dense?
- Guthur 7y agoI assume both the parents are referring to TSMC, global foundries stepped out of 7nm to concentrate on existing processes
- dannyw 7y agoOn the other hand, the top AMD Ryzen (3950X, $749) is considerably faster than the top Intel CPU (10980XE, $1000). https://pcper.com/2019/10/ryzen-9-3950x-benchmark-i9-10980xe/ https://pcper.com/2019/10/ryzen-9-3950x-benchmark-i9-10980xe... Technically meaningless, including thinking that Intel is superior to AMD. They are losing the desktop wars hard to AMD with Ryzen, and starting to lose ground on the server front.
- segfaultbuserr 7y ago> including thinking that Intel is superior to AMD I never stated that. Edit: I updated the original comment to clarify that. > They are losing the desktop wars hard to AMD with Ryzen, and starting to lose ground on the server front. I never purchased any new Intel's CPUs (or AMD's CPU, for that matter) since Ivy Bridge. If I have to build a new x86 PC today, I'll definitely buy AMD.
- iCarrot 7y agoI'm curious, do you still use the old Ivy Bridge or have moved on to something else non-x86?
- segfaultbuserr 7y agoSorry if it is disappointing: I'm still using the original Ivy Bridge machine ;-) I found it's a fairy capable machine for day-to-day uses, even building a relatively large project is not a problem. There is a performance hit after Meltdown/Spectre, especially when you disable hyperthreading. But yes, I'm actively looking forward to try a non-x86 platform as my desktop, and I'm watching for this PowerPC laptop [0]. Another option is the MNT Reform ARM laptop that offers an interesting, hackable design [1]. For desktops, HoneyComb LX2K is an ARM motherboard that offers workstation-level specs [2]. But the first two projects are still in development. HoneyComb looks good but the 550 USD price tag requires me to consider carefully (the same money may buy me more interesting hardware in the near-future soon). For POWER, the Raptor Talos II workstation [3] is the first choice comes to mind, but it's POWER, the 3000+ USD price tag is not for the light-hearted. For RISC-V, it costs around 2000+ USD (1000 USD for the CPU board [4], another 1000 USD for the PCI-E board [5]) to build a desktop from the current development system. I'm still waiting for an affordable RISC-V SoC that can be used as a 400-500 USD desktop. I can pay 2000 USD if I decide it's the One True machine that can be used as my daily driver, but it's obviously not the case for now. So... waiting, waiting, waiting, and decision, decision, decision... [0] https://www.powerpc-notebook.org/en/ https://www.powerpc-notebook.org/en/ [1] https://news.ycombinator.com/item?id=21231031 https://news.ycombinator.com/item?id=21231031 [2] https://www.solid-run.com/nxp-lx2160a-family/honeycomb-workstation/ https://www.solid-run.com/nxp-lx2160a-family/honeycomb-works... [3] https://www.raptorcs.com/ https://www.raptorcs.com/ [4] https://www.sifive.com/boards/hifive-unleashed https://www.sifive.com/boards/hifive-unleashed [5] https://www.crowdsupply.com/microsemi/hifive-unleashed-expansion-board https://www.crowdsupply.com/microsemi/hifive-unleashed-expan...
- pkaye 7y agoI wonder if the scale could be expressed in the area required to implement a given amount of logic and SRAM. This would make is easier to compare things.
- rrss 7y agoSRAM cell area is a very common benchmark for processes.
- permatech 7y agoIf feature size doesn't matter why has EUV been so critical?
- elihu 7y agoMy (incomplete) understanding is that current lithography relies heavily on multipatterning, which takes a long time. With EUV, you can have very small features with fewer lithography steps which improves factory throughput. That's more of an economic benefit for chip manufacturers than a tangible benefit for a user of the end product, but it means being able to buy cheaper chips with more gates (in addition to any power or performance improvements that can still be had these days).
- wtallis 7y agoFYI, GlobalFoundries canceled their 7nm process over a year ago. The leading-edge semiconductor fabrication race is converging on just Intel vs. TSMC, with Samsung still relevant in certain markets.
- mrfusion 7y agoCould we rate them on number of transistors instead? Is that number going up?
- Veedrac 7y agoYes, at the same rate it always has, to a T. https://fuse.wikichip.org/news/2207/tsmc-starts-5-nanometer-risk-production/ https://fuse.wikichip.org/news/2207/tsmc-starts-5-nanometer-...
- mrfusion 7y agoSo something is shrinking? It’s just not at the nm’s they claim?
- brennanpeterson 7y agoThere is the fin pitch, the gate pitch, the metal pitch, and (related) the track height (which is a design density number). Fin pitch shrinks a little, gate pitch a little, track lots (Intel used to use 9-12, moving to 6). There is also space between device boundaries, which also can be shrunk. The 5nm is the gate half pitch, which is no longer a sufficient description. All else being equal, simple shrink would shrink everything, and that gate half pitch would be useful. That hasn't been true for 15 years. Density comes at the cost of performance and yield, so you balance different aspects to get the right combination. Or _a_ right combination for your design.
- kbumsik 7y ago> GlobalFoundries' 7 nm process is similar to Intel's 10 nm process. TSMC and Samsung's 10 nm processes are only slightly denser than Intel's 14 nm in transistor density. They are actually much closer to Intel's 14 nm process than they are to Intel's 10 nm process. Where can I get the source? I often see someone saying "X's process is better than Y's" but they never leave any links. I would like to know any credible source on the semiconductor manufacturing.
- segfaultbuserr 7y agoThe main source is WikiChip. It appears that the wiki is edited by the people who work in the industry and has a lot of information that seems to be fairly reliable. Unfortunately, it rarely has in-line citation to its primary sources, but still a good source. Read, * 14 nm lithography process https://en.wikichip.org/wiki/14_nm_lithography_process https://en.wikichip.org/wiki/14_nm_lithography_process * 10 nm lithography process https://en.wikichip.org/wiki/10_nm_lithography_process https://en.wikichip.org/wiki/10_nm_lithography_process Other sources include: * Life at 10nm. (Or is it 7nm?) And 3nm Views on Advanced Silicon Platforms https://www.eejournal.com/article/life-at-10nm-or-is-it-7nm-and-3nm/ https://www.eejournal.com/article/life-at-10nm-or-is-it-7nm-... * A Brief History of Process Node Evolution https://www.design-reuse.com/articles/43316/a-brief-history-of-process-node-evolution.html https://www.design-reuse.com/articles/43316/a-brief-history-... * 14nm, 7nm, 5nm: How low can CMOS go? It depends if you ask the engineers or the economists… https://www.extremetech.com/computing/184946-14nm-7nm-5nm-how-low-can-cmos-go-it-depends-if-you-ask-the-engineers-or-the-economists https://www.extremetech.com/computing/184946-14nm-7nm-5nm-ho... * Is Intel Really Starting To Lose Its Process Lead? 7nm Node Slated For Release in 2022 https://wccftech.com/intel-losing-process-lead-analysis-7nm-2022/ https://wccftech.com/intel-losing-process-lead-analysis-7nm-... You get dig deeper using these links. And don't ask me why I'm just copy & paste Wikipedia citations. I added those citations...
- grandmczeb 7y agoChipwiki has some numbers[1] sourced from industry presentations, but unfortunately foundries usually prohibit publishing direct comparisons with reference designs. As a result there's a lot of "folk knowledge" that is commonly known by insiders but doesn't have a public source. [1] E.g. their 7nm page https://en.wikichip.org/wiki/7_nm_lithography_process https://en.wikichip.org/wiki/7_nm_lithography_process
- baybal2 7y ago> When some people say "5 nm" is the early-stage of nanotechnology, remember that the gate length actually stays at 25 nm since 2009 Single exposure limit is 40nm for ArF and 30 for EUV. The cost of those 10nm were ~$18B spent over 2 decades, and it will go up much further. 3400 is rumored to costs $360M a pop in early production. 157nm fluorine laser can actually get to 25nm with a single exposure, and pure argon even further. 25nm is better than EUV at the moment, but it was an industry wide decision that it was better to stop beating a dead horse, and work on a first in-vacuum process as it has the biggest potential for further improvement, and opens a road to less painful transition to soft X-ray and particle beam litho.
- discordance 7y agoIs there still a relationship in regards to reduction of power consumption with smaller feature sizes?
- wmf 7y agoYes, smaller transistors are more power efficient.
- imtringued 7y agoThe number of transistors grows faster than the power efficiency though, so the percentage of unused silicon (aka dark silicon) increases as you go smaller but keep the die size the same.
- xwolfi 7y agoI guess the best the pleb like us can use as a metric is TDP/core frequency/number of core ? The TDP for me was always a sign of progress in density for same core count and frequency.
- ComputerGuru 7y agoNope. Some manufacturers exclude “turbo” power usage from the TDP, others don’t.
- aoeusnth1 7y agoIntel excludes turbo power usage, while AMD doesn't. Intel's TDP is really its base load.
- paulmd 7y agoAs of the 2000 series, AMD uses the same model as Intel. Their "105W" processors will draw up to 141.75W while boosting, their 65W processors will draw up to 88W while boosting, etc etc. It is, of course, always more of a generalization than a precise measurement, but AMD and Intel are now generalizing roughly the same thing. Both brands will exceed specified TDP while boosting.
- moonbug 7y agonot so for epyc.
- jiggawatts 7y agoLike you said, I'm starting to ignore the "nm" units and just think of these new processes in terms of generation id numbers. The thing is, I don't really care that they're fudging the numbers for marketing, because at the end of the day the end result for me as a consumer is awesome. High-end products like my iPhone X, AirPods, NVIDIA 2080 RTX, or the upcoming AMD Threadripper 2 chips are simply mind-blowing to me. The ridiculously huge transistor density has put crazy amounts of computer power in my hands and I absolutely love it. I'm constantly excited for what the future will bring, and from what I can it'll just keep on getting ever more spectacular. If some foundry dumps $20B of investment and man-centuries of cutting edge research into technology this awesome, they can call it whatever they want.
- gameswithgo 7y agoand on top of all of that, transistor density no longer works are a rough approximation of performance either. clock rate may get worse, with a more dense node, heat issues may not let them actually put more transistors in the same area despite the smaller pieces, etc etc etc. as usual, the only way to actually judge a cpu's tech is how many frames of doom it can render per second ;) if any pedant responds to this about how modern doom is mostly a function of GPU I will kill a kitten.
- vocatus_gate 7y agoIncidentally, there's a great write-up by Digital Foundry on why Crysis still melts modern CPUs, and is still CPU-limited. https://www.eurogamer.net/articles/digitalfoundry-2018-why-crysis-still-melts-the-fastest-gaming-pcs-10-years-later https://www.eurogamer.net/articles/digitalfoundry-2018-why-c... "Not a trick, not an illusion. Yes, this 2007 game is running at under 40fps on the fastest CPUs money can buy."
- danmaz74 7y agoWould transistor density be a better metric? Or that wouldn't be significant either?
- lmilcin 7y agoDensity is no longer meaningful. I have stopped looking at the process node or stated specs. There is so much architectural differences that the only sane way is to run YOUR workload and compare performance/TCO.
- Shorel 7y ago"they do have a lot of production issues" means they still sell every single processor they make. New computers are about 80% Intel and 20% AMD. Until this is reversed Intel will still be the kings. I remember fondly the age of Athlon 64 processors. AMD was a winner. On paper. Intel still beat them in the long run. And I say this as an AMD fanboy.
- rasz 7y agohttps://wccftech.com/amd-outselling-intel-by-more-than-double-analyzing-5-year-historical-sales-at-mindfactory-de/ https://wccftech.com/amd-outselling-intel-by-more-than-doubl...
- Nokinside 7y agoMindfactory is retailer and very small part of the market. Revenue 2018: Intel $70.8 billion AMD $6.5 billion
- Nursie 7y agoIn the consumer space AMD appear to now be outselling Intel, in some markets this is reportedly by a considerable amount. This may be down to the Zen 2 release going well, and it's going to be interesting to see if it's sustained. I say this as a fanboy of neither really, and owner of both! (I'm really a fanboy of the re-emergence of serious competition in the space 8C/16T for ~$300? Yes please...)
- abrugsch 7y ago>(I'm really a fanboy of the re-emergence of serious competition in the space 8C/16T for ~$300? Yes please...) 100% this. The lack of any credible competition in the consumer space for intel really let them get complacent. I jumped to a 2500K sandy bridge as soon as it was apparent it was a game changer. I didn't keep it in the end as I wasn't really using it, so let my father in law have it for his video editing rig. Only now can I answer the question "Is it worth upgrading yet?" with a qualified YES! I personally have no AMD/intel bias (just happenstance that when I've needed to build a PC, intel had the best price/performing/overclockable part at the time), but I welcome the competition to not have a single supplier stagnate the market with 4C/4T consumer parts for nearly a decade... (currently have a R5 1600 and looking forward to a large range of upgrade options within the AM4 socket - though I am aware the forward compatability may not last much longer - but at least I'm paying attention to the market again!)
- bogomipz 7y ago>""x nanometer" has became a technically meaningless trademark solely exists for indicating new generations of process for marketing since ~2009's introduction of FinFET." What is the connection between the introduction of the FinFET and the nm process number becoming a meaningless distinction? Might you or anyone else elaborate? Thanks.