>How times have changed! I still vividly remember Intel being a full node ahead of everyone else with their release of 22nm node. For example, see this article from 2012:
Vividly remember, I was reading this and assuming it is going to be way back.... 22nm and 2012 really wasn't that long ago :D
Intel has had the processing node leads on CPU since the 90s. I vividly remember Pentium were always at least one year ahead of AMD. That was the Pentium, Pentium MMX .... Not quite sure if it was the case in 386 / 486 era.
I remember 180nm was already in AMD Athlon era, so it was likely in 2xx nm or 3xx nm.
AMD was actually briefly ahead twice before, once in early Athlon times and a few years later with the first AMD64 ("x86-64" or as Intel called it "EM64T") CPU with on-die memory controller.
Those were microarchitectural advantages. I don't think AMD was ahead on manufacturing process shrinks. A quick check shows that the P3 and Athlon went from 250nm to 180nm at the same time.
I don't know whether the Intel foundry accepts outside customers and to what extent the outside customers have influence on Intel, but TSMC benefits from being single mindedly focused on chip fabrication and has the support of a wide array of customers, financially and technically. Apple, Nvida, AMD, Qualcom, Micron, most of the phone manufacturers, and any other chip consumers all give orders to TSMC. All these pour money, design and production requirement into TSMC. TSMC simply is more focused with the resource they have, while Intel is being distracted with other things.
Sure TSMC is more focused but Intel has dramatically greater resources. Their R&D budget is ~5x TSMC's and ~10x AMD's. You have to be pretty badly distracted to fail to produce superior products with such an enormous budget advantage.
Intel technically accepts outside fab contracts but they haven't had any customers since they bought out Altera, their previous only customer.
Intel's budget is not just for chip fabrication, they have a very wide array of product areas. Chip foundry is very cost intensive, in the range of 10's billions. I'm not surprised TSMC spends way more on fabrication than Intel does.
According to the numbers of the post above, if they allocate more than 20% of their R&D budget to process improvements and more than 10% to their architecture improvements, they match or exceed their two rivals, leaving 70% to be allocated to other tracks, which I find rather excessive all things considered.
I remember a comment I made a long time ago, in the beginning of the Browser Wars: Microsoft had more people assigned to design their icons than NetScape had employees. NetScape's fate was sealed from the get go and their resilience was commendable.
And yet, in the end, Microsoft still lost that war.
I think the key difference here is also competition. If Intel's foundry can't get 10nm right then Intel is SoL until that's resolved. Whereas AMD was able to ditch GF for TSMC as soon as it became clear that TSMC had the superior process.
This has me really excited for the future. One question I have now is how this compares to Intel’s roadmap. I’ve read comparisons like “TSMC 7nm is equivalent to Intel 10nm” but never seen any further discussion that goes deeper into how that’s true.
To summarize the article: the density at which Intel can pack the transistors on a chip is high enough even with a larger transistor size, still means Intel has more transistors in the same chip size. The advantage of the 7nm TSMC transistor is that they are faster and more power-efficient.
Yes, although there was an element of luck that Intel 10 nm failed at the same time that TSMC 7 nm came out smoothly. The opposite situation happened back at 32 nm and could happen again.
It is derived through quantum mechanics and via the Plank constants and black hole physics.
Basically any more information squeezed together would turn into a black hole.
Though ordinary (non quantum) thermodynamics limits information density well before this bound.
As to what happens in the next decades with the good old Silicon stuff. Well, nobody knows just yet, but 3nm is what anyone realistically talks about, and just sci-fi tech after that.
Current density is ~100M (1E8) transistors/mm^2, Assuming transistors with a side of ~10 atoms (and bond length 0.5nm) that means we could get ~4E12 transistors/mm^2. This is almost certainly over optimistic but still it shows that we are orders of magnitude away from limits due to "atomic nature" of matter.
Right now these are planar devices. If you stack you can get orders of magnitude more. That's why you can get absurd memory capacities in things like nitrogen defect diamonds [0].
If you need to move charge then you get nasty joule heating, but if you only move spin density then this causes almost no heating (that's the promise of spintronics)
Those aren’t forgone conclusions. That’s just guesswork on a whiteboard roadmap. At 10nm we are already at a stage where quantum effects start to dominate, and it’s not clear how much further these classical effects can scale. 7nm was “a process too far” for Intel, and it is a bit of a miracle that TSMC managed to pull it off.
5nm isn't just on a whiteboard, it's in risk production already at TSMC. Volume production is slated for Q1 2020, less than a year away, and TSMC tend to be accurate with their estimates.
I hope you're right, but it's hard to imagine they aren't going to hit an unexpected wall at some point. The question is where is the wall? Basically you can never predict and unknown unknown.
5nm started risk production were direct from TSMC report. They are expecting even faster ramp than 7nm. And TSMC has a history of being very conservative.
Their approach has been iteration rather than a leap like Intel did. And you can bet it will arrive on scheduled for next year's Apple iPhone.
There are roadmap, tools, technique for 3nm ( As explained in the Article ) and even 2nm. So none of these are pipe dreams. As long as these customers can keep paying top dollar to be on leading node, TSMC seems to have no problem with innovating. The question is when will these clients slow down and stop paying every year because leading node is too expensive. Basically the cost of designing leading node is doubling every two year. So we are seeing something like the inverse of Moore's Law.
I bet yes, just the top dog clients will have to change. See, semiconductor companies might look like quite formidable, but the Internet companies like Facebook, Amazon, Google actually have more money then even them. They will pay.
For consumer electronics, there is genuinely no benefits now going to smaller nodes, but for something like a "single chip supercomputer" type products, made with the most over the top semi tech, there is no better client than them.
For them, they are basically turning joules into ad clicks, and thus money, much akin to that bitcoin thing, which up until few years ago was one of the biggest semiconductor consumer globally.
> Internet companies like Facebook, Amazon, Google actually have more money then even them. They will pay.
I agree. And this is where the market dynamics changes, it will be interesting when the cost / economics model changes how things will unfold. Especially Amazon as they are already designing their own ARM CPU for Cloud Services.
The performance of an individual transistor actually went down in 14>10>7 transition, but since you can pack more of them, the net effect is gain for companies with huge chips
Yes, but I think the cost of the chip (and cost per transistor on the chip) is starting to go back up again with each die shrink, rather than down every die shrink.
I'm not saying they're pipe dreams. I'm saying things go wrong sometimes. How many companies have never missed a deadline? I'm not even saying that to be argumentative - I am genuinely curious.
It sounds like you're still not taking seriously the fact that 5nm isn't just planned, but already well into the execution phase. It's far enough along that the worst case scenario at this point would be for TSMC's 5nm to be much more usable than Intel's first attempt at 10nm. The horizon for really serious unforeseeable problems is well past 5nm already.
>You could have said the same things about Intel and 7nm a few years ago.
Intel's 10nm and 7nm never went into Risk Production, Intel could lie about their progress because they were the only user of their Fab. And they never gave a precise definition, progress of their node, as they are not a Foundry business. Compared to TSMC Hundred of Customers relies on them being open and transparent with their progress. As long as Intel could continue to sell their Chips, it really doesn't matter which node they are on. Which is the reason Why Intel Custom Foundry never took off, they have held information to closely to themselves, and didn't treat their customer as parter in the likes of Samsung and TSMC.
That’s how you make the chip. I’m not doubting that you can get manufacturing smaller, for a while[0]. But will it still work? You’re getting into the domain where electrons start tunneling across transistors and such. There is a wall that we will hit where lithography can take us no further—we’ll need atomic precision to break through that wall. It’s just a question of where that wall is...
[0]: Higher frequency UV light is more tightly focused, but also imparts more energy. At some point the blasting power of a single photon does damage on a scale larger than the wavelengths involved, at which point EUV will take you no further.
Transistors at 7 & 5 are nowhere close to having serious “quantum” issues—that regime is thousands of times smaller than current gate design. Remember that the “size” is a marketing term that really relates most closely to “effective” wire pitch & gate density with respect to processes in the hundreds-of-nanometer scale. The absolute smallest functioning design I’ve read about is IBM’s 7-atom gate. That gate is something like 100thousand times smaller (volumetrically) than current production.
Also, current HW design is really inefficient. Think about all the crud between some front-end developer and the server powering the webpage on the backend, vs a ‘native’ app. That’s a good analogy to current HW design and real total-badass low level HW design.
I've heard about TSMC working on so-called "3nm", so while I don't know whether it approaches physical limits (those nanometers are just marketing speak after all), it's certain that we have another 10-15 years of progress ahead of us.
AMD was lucky, but it is also being run by a highly accomplished semiconductor physicist, so some things are perhaps a little less luck than one might think.
I would vote for Lisa to be the CEO of the year. It's also fun to know she is a relative of NVidia boss that once worked at AMD and they both are from Taiwan, home of TSMC and who knows if they have family inside as well...
"It is worth mentioning that another Taiwanese semiconductor elite, Nvidia founder Huang Renxun (Jensen Huang) also came from Tainan, and has a relative relationship with Su Zifeng (Lisa Su). Su Zifeng’s grandfather and Huang Renxun’s mother are brothers and sisters"
AMD has to choose a process years before that process goes into production, so there's a lot of risk that production could get delayed or yields could be poor. AMD could choose TSMC then discover later that Samsung would have been better.
I don't know that it was a lack of money so much as the cost/benefit ratio.
We like to think of all the big chips from Intel, AMD, Nvidia, etc or of bleeding edge mobile processors from Apple or Qualcomm. The reality is that these are a small fraction of the total market for chips. The vast majority of chips are made at 28nm or larger because the cost of production is so much lower.
Supply and demand is another big factor. As supply increases, prices will drop down to bare minimums and if costs can't be reduced enough, there will be expensive fabs sitting around doing nothing. Even if GF has enough money to design and make the fab, that doesn't mean the demand is high enough for them to charge enough to recover their costs.
Instead of spending that money on 7nm, they can spend it refining their 22, 14, and 12nm lines to improve efficiency and lower production costs. With TSMC or Samsung pouring those billions into 7 and 5nm, there's a good chance at GF pulling quite a bit ahead in the slightly older nodes. Once they have done this, those bulk chip makers using 28nm will start moving to smaller nodes. There's quite a tidy sum of profit to be reaped here. Given how every consumer has dozens to hundreds of these small chips for every cutting-edge chip in their home, appliances, or car, there is cumulatively a huge savings to be had in product and energy costs.
They had already developed the node, or in any case, IBM had (and they purchased the IBM fab research division). But they didn't have the capital to scale out. They decided it would be more profitable and less risky to stay on 14nm and serve the long tail of customers that don't use leading edge nodes.
In today's world of geopolitics, restricting access to decent silicon processes can have a massive knock on on other industries.
If Taiwan/China have better silicon manufacturing capabilities now, they can restrict export of those chips, to effectively capture the phone/server markets too...
Could be a big part in the nail in the coffin for USA tech dominance.
Uh, why are you lumping Taiwan in with China? They're 2 separate governing entities that aren't particularly friendly, and have missiles pointed at each other.
Absolutely astounding numbers in there as someone who doesn't know the details. N5 at a hundred and seventy million transistors per square millimeter. That's mindblowing.
I hate when acronyms prevent people from understanding important parts of an article. Here you go.
TSMC = "Taiwan Semiconductor Manufacturing Company, Limited, also known as Taiwan Semiconductor, is the world's largest dedicated independent semiconductor foundry" (from Wikipedia).
You are expected to know what TSMC is if you are reading this... People wouldn't expand AMD to Advanced Micro Devices and then explain it develops CPU (Central Processing Unit, in case you didn't know!), would they?
You're expected to know what they are, but I couldn't have given you the full version of either. It's pretty irrelevant to the conversation, more trivia than anything.
If you led an article with "Taiwan Semiconductor Manufacturing Company" my brain would chunk uncomfortably undigested stomach units out of orifaces before it understood.
Do you think you are in the majority? If so - I need to readjust my head.
How times have changed! I still vividly remember Intel being a full node ahead of everyone else with their release of 22nm node. For example, see this article from 2012: https://www.extremetech.com/computing/127987-deliberate-exce...
This too shall pass.