Arthur T Knackerbracket has processed the following story:
The Commercial Times reports that TSMC's upcoming N2 2nm semiconductors will cost $30,000 per wafer, a roughly 66% increase over the company's 3nm chips. Future nodes are expected to be even more expensive and likely reserved for the largest manufacturers.
TSMC has justified these price increases by citing the massive cost of building 2nm fabrication plants, which can reach up to $725 million. According to United Daily News, major players such as Apple, AMD, Qualcomm, Broadcom, and Nvidia are expected to place orders before the end of the year despite the higher prices, potentially bringing TSMC's 2nm Arizona fab to full capacity.
Unsurprisingly, Apple is getting first dibs. The A20 processor in next year's iPhone 18 Pro is expected to be the first chip based on TSMC's N2 process. Intel's Nova Lake processors, targeting desktops and possibly high-end laptops, are also slated to use N2 and are expected to launch next year.
Earlier reports indicated that yield rates for TSMC's 2nm process reached 60% last year and have since improved. New data suggests that 256Mb SRAM yield rates now exceed 90%. Trial production is likely already underway, with mass production scheduled to begin later this year.
With tape-outs for 2nm-based designs surpassing previous nodes at the same development stage, TSMC aims to produce tens of thousands of wafers by the end of 2025.
TSMC also plans to follow N2 with N2P and N2X in the second half of next year. N2P is expected to offer an 18% performance boost over N3E at the same power level and 36% greater energy efficiency at the same speed, along with significantly higher logic density. N2X, slated for mass production in 2027, will increase maximum clock frequencies by 10%.
As semiconductor geometries continue to shrink, power leakage becomes a major concern. TSMC's 2nm nodes will address this issue with gate-all-around (GAA) transistor architectures, enabling more precise control of electrical currents.
Beyond 2nm lies the Angstrom era, where TSMC will implement backside power delivery to further enhance performance. Future process nodes like A16 (1.6nm) and A14 (1.4nm) could cost up to $45,000 per wafer.
Meanwhile, Intel is aiming to outpace TSMC's roadmap. The company recently began risk production of its A18 node, which also features gate-all-around and backside power delivery. These chips are expected to debut later this year in Intel's upcoming laptop CPUs, codenamed Panther Lake.
(Score: 2) by VLM on Friday June 06, @06:03PM (2 children)
When they don't spec them, metrics don't mean much.
Most digital designs are repetitive and you can knock out rows/columns of memory or just sell half-capacity chips or sell a 6-core CPU with 2 dead cores as a 100% yield 4-core cpu
You pretty much can't do this with analog designs so people like to dump on analog foundries because everything is bigger yet has lower yields; well what do you expect, you can't just substitute in a different main output transistor LOL if the primary fails testing. Or if the input transistor exceeds noise limits, well, I guess its junk.
Other parts of chip design involve way more serious engineering and less marketing, like DRC rules. (similar to automated design checks for PCBs but for chip layouts, lots of minimum metalization spacing rules)
The one thing I've found that even made up "yield" metrics do usually indicate, is if the process can make it up to "meh 20%" then its been a tradition since Moores Law, or maybe before, that they can improve that to 90% given enough time and money, but if they can't initially get it up "meh 20%" on the first few runs, well, that line/process/plant is toast and they'll never be able to fix it may as well pivot or come up with a different plan. No comment about Intel's 18A process, LOL. Last I heard during the winter that Intel had made it up to 25% yield so if they don't make it up to 90% eventually, that'll be ... unusual, but it could kill the company. I would be surprised if A18 fails, once they get above 20% historically success seems "guaranteed". Of course hindsight is 20-20, past performance is no guarantee of future success, etc. 18A is a direct competitor to the process in the linked article and is a bit more ambitious and historically less successful but it "seems" to have turned the corner as of late winter; haven't heard anything reputable, positive or negative, since. Personally I think N2 and A18 are both going to make it although Intel is going to take a substantial financial hit on A18 taking longer than expected; I have no dog in the fight having no investment either short or long in either (at this moment) although I keep a really close eye on this field.
I find it intellectually interesting to keep up with energy and semiconductor financial news because with an engineering background its possible to see thru/past a lot of liberal arts financial analyst / marketing BS and historically I've been pretty good at it $$$. Petrochem is another field where a little engineering goes a long way when reading the 10-K and 10-Q reports. As a non-engineering perspective, pre-SOX 10-Qs always had a veneer of bullshit that is absent in the post SOX era and they're now about as trustworthy as 10-K reports.
Non-corporate access to foundries has been going backwards over recent decades, I think MOSIS put their once-free university program on permanent hold, for example. And its been all digital since at least the early 90s. However I've always thought it would be funny as a piece of performance art to submit a 60s style hand made tape mask to a service like MOSIS just to see what happens, like as a Youtube channel stunt. In the old days they used to make PCBs using tape masks so thats how they did ICs before 1980s. And here we are in 2025 where the handover from designers to foundry gnomes is still called "tape-out", which I think is funny. It would be like insisting on calling my modern 4K LCD a "cathode ray tube" out of tradition or calling my source code files "card decks" just to be retro as F.
(Score: 2) by VLM on Friday June 06, @06:20PM
Oh another thing I forget to mention is WRT the whole geopolitical struggle thing, A18 is 100% "made in the USA" Intel's sites in Orgeon, AZ, NM whereas famously TSMC's N2 is 80% Taiwan and 20% ... of all places, Phoenix AZ.
With respect to tarriffs and sabre rattling and general TDS and political nonsense I don't think "TPTB" will let Intel's A18 process fail, money will be shit at them until yields exceed 90% all american made. Of course, again, TPTB let Bear Stearns fail back in '08 and that was unthinkable until they flushed 'em, so in theory although it seems unthinkable, Intel could still get axed ... but I just really don't think so.
There's a lot of talk about how Apple can't make iphones in the USA because everything, include TSMC, is in Asia, however, A18 is technologically superior to N2 and is only USA-made, so I am pretty close to betting actual money that eventually the next Apple iWTF will ship with American-made chips from Intels A18 process. Currently Apple is a seemingly life-long customer of TSMC but I think that is unlikely to continue forever. They can do the fashion boutique thing for awhile but if Android phones are vastly technologically more capable, longer battery life, cheaper, then they're going to have to switch to Intel from TSMC to keep up at some point...
The point of all this is A18 is "not failing" or "succeeding slowly and expensively" so it would be a major disaster if the simpler lower performance N2 was not extremely successful, so in a way, there's no real news here. It would be incredibly bad news for TSMC if their simpler process were not wildly successful at this time. I think they're going to get passed up, hard, by Intel "pretty soon", so they have to sell as much N2 as possible while they still can.
(Score: 2) by bzipitidoo on Saturday June 07, @07:01PM
I find I had lost sight of just how very much bull there is in the world. Even with a mental note to always, always check for bull, I still find myself a little too credulous.
Maybe benchmarks are the best measure of performance? But those can be gamed. And it turns out that like with IQ, a single performance metric is too simplistic. At single threaded math, Intel is faster than AMD. But AMD performs faster than Intel at multitasking. Maybe what would be good is something more like baseball's penchant for stats for everything.
The performance of a computer system depends on a ton of factors. Like whether you have parasitic software stealing CPU cycles. All the CPU performance in the world won't do you much good if 90% of it is being taken by malware. Bugs too. Also, I get the impression that software today is full of slop, crap algorithms, bloat, and so forth. I begin to wonder if apps these days are just call after call to library functions, with enough wildly wasteful ones to noticeably clog up the computer. Like, to get the high value in a long list of values, would an app programmer call upon Quicksort, then take the value at the end of the list the sort returns?