I wish the industry would move to transistors per mm or something instead of bongus "nm" numbers that are meaningless - 91 to 171 MTr/mm^2 even sounds more impressive then whatever 7 to 5 nm means.
It depends on the mix of circuits. Memory don't follow same design rules as logic so can scale quite well. Logic density may well be limited more by power dissipation than by design rules. In other words you can pack some circuits a lot denser than others.
I do agree that it is frustrating for someone from the outside to understand exactly what kind of scaling one can expect.
CPU purchase cost matters too. The electricity cost of one 100W CPU might be around 500 USD per year in a datacenter. Assuming 5 years useful lifetime that's 2500 USD per CPU. But cheaper CPUs can easily cost 1000 USD less, so even if they're 20% less efficient they may still come out ahead.
A 100 watt CPU in any modern data center would be at an incremental cost of well under $40.
Data centers get cheap bulk power, cooling is mostly passive now (evaporative) and backup power systems are also a lot cheaper now compared to the compute cost than in the past.
https://semiengineering.com/5nm-vs-3nm/ TSMC’s 5nm technology is 15% faster with 30% lower power than 7nm. A second version of 5nm, due out next year, is 7% faster. Both versions also will use EUV.
Seeker released a very insightful feature about the ASML Extreme UV / soft X-ray (13.5 nanometers) photolithography system that is being referred to in the article.
It means that you can produce the same design (IP or intellectual property) on both processes. The denser process should then result in a better frequency or power performance without a redesign. Many processes do require a redesign because of different latency or more redundancy/error checking required and so on.
N5s being optimised for high performance computing off from the start (Ratber than only mobile) look like just the kind of thing an Arm MacBook would use.
Is it just me or the competition on the CPU market is about to get really interesting? My understanding is that if Intel loses the edge on fab-tech then the door is open for all kinds of new entrants. (I could be wrong though.)
Who knows, perhaps even Intel will use TSMC's process at some point.
The CPU market has been interesting for a couple of years. Intel effectively already lost the edge on fab-tech in terms of what they are able to produce. Forget about TSMC Xnm vs Intel Ynm... the numbers are meaningless. Look at the performance per watt of AMD's latest offerings vs Intel's. They are behind today.[1]
That said, in a lot of ways the really interesting stuff is likely to happen on the older nodes. There's really nothing enabling about the latest fab node... it's just bragging rights for the highest absolute performance and/or performance per watt for CPUs which most applications don't need and can't afford. ARM/RISC-V/whatever doesn't need 5nm to be feasible... in fact, those 'other' products often aren't viable on the latest nodes due to cost unless you're a market leader with huge volumes to amortize the cost over.
[1] Sure, they still offer the absolute top performance per thread... by a bit. But look at the price they're paying from a perf/watt standpoint to get there. I seem to recall AMD being ridiculed not too many years ago for doing this...
What makes you think new entrants? Just curious. On the fab side definitely not... there's really only 3 fabs competing on the bleeding edge side of things, and it costs way too much to open another. On the 'chip designer' side maybe, but I'm not sure why Intel's performance affects whether or not someone jumps in. I basically see a convergence coming, with ARM stuff getting better performance, and x86 stuff using less power...so there will be an interesting crossroads of sorts to see which direction things go.
The new entrants would be on the design side.
If Intel no longer holds the keys to best fab-tech in the world then all bets are off.
(I'm assuming that Intel doesn't allow competitors to use their fabs; and that Intel does indeed have the best fab-tech at the moment; maybe I'm wrong.)
Even a fabless semiconductor company (e.g. nVidia, Apple, heck, AMD) needs to spend hundreds of millions to tape out a leading-edge IC. It's a capital-intensive business, and would present quite a risky bet for any upstart competitors. Also, patents.
It would be fascinating for another major fab to arise. It's not impossible to imagine how either, a large chip designer with piles of cash could decide they want to get into fab, or a niche fab could slowly move into it?
Not going to happen unless there is some absolute paradigm shift in technology. Not many people throw around 20 billion dollars on upstarts.
The problem is that the paradigm shift technology we've read about multiple times has always been overcome by more investment in standard silicon processes.
Economically we're long past the point of contraction - there has been a ridiculous pace of mergers in the semiconductor industry to have the network effects to stay current. The effects are worse at the top, and we are down to three companies even in the race.
Basically, no, we'll never have another fab entrant who is a major player.
> My understanding is that if Intel loses the edge on fab-tech then the door is open for all kinds of new entrants. (I could be wrong though.)
Doubt it. There are only two players in the x86 market.
I haven't seen anything compelling for personal computer and server space yet to take over AMD or Intel. I believe AMD is going to gain some market share with their new products.
That means "same power", in the context of a comparison of some other metric. So (from the article) "15% higher performance at iso-power" means that for the same power draw, a chip made with the new process has 15% better performance (clock speed) than one made with the old process. Likewise for iso-performance: new chip can get the same perf with less power.
Assuming you are referring to this:
"N5 provides around 15% higher performance at the iso-power or, alternatively, up to 30% power reduction at iso-performance."
They are just saying that for the same power usage as previous tech, they get 15% more performance, or 30% power reduction while achieving same performance as previously acheivable."
The prefix "iso" means "same"/"equal". E.g. isobars (same pressure) on a weather map, isotopes (same place) on the periodic table, isomorphic (same structure), etc.
Here iso-power means comparing performance for the same power usage, and iso-performance means comparing power usage for the same performance.
If I’m reading this correctly... the expected soon to be released AMD 7nm mobile chips (early 2020) will potentially consume 50% less power than their current generation?
With 5nm further improving that 7nm another 30% reduced power?
If so, we’re talking 25-35% of the current power draw for the same speed in the next 18-24 months?
I have to be wrong here... I know they use a design which doesn’t get 100% of the Benefits due to the “chiplet” arch, but still.
Fair - I’ll edit, I did the math wrong (I was looking at each iteration N7P, N6, N5, N5P). However, judging from their graph not all paths should be multiplied.
Depends on who you ask, yield is definitely not as good as TSMC, but Samsung like to use brute force approach to literally everything so I assume it will be sorted out soon. The downside to this approach is margin but profits isn't important to Samsung Foundry ( yet ).
