TSMC Completes 5 nm Design Infrastructure, Paving the Way for Silicon Advancement

[lexluthermiester]

Cloud will never replace local computing, not unless Google, MS, Amazon start subsidizing their cloud.

Even then most people don't care. I'm not storing anything in the cloud and I'm certainly not going to operate fully that way. Ever. Lot's of people share that sentiment.

cut down vega die (60 cu) at 1750mhz with 300w tdp
full vega die (64 cu) at 1670mhz with 345w tdp
not even close to being good enough for a 14nm-7nm shrink

Those numbers also make a lot of assumptions and generalizations. It's far more complicated than that, and you know this.

 

[notb]

We've been through this in other threads.. Never gonna happen! While the cloud has it's uses, it can not and will not replace standard operation methodologies for the common/general user.

Indeed, we've been through this in other threads and I believe not much has changed.
I'll be frank but as polite as I can.

You seem like a nice guy (piracy stuff put aside). I know you're building/tuning PCs and you know a lot about it (more than me for sure).
You have absolutely no idea about enterprise computing. It's obvious to me and every forum member in this business will sense it as well.
It also seems you have limited understanding of how casual PC users do stuff today (you don't believe in cloud, you're not very enthusiastic about laptops etc).

Learn something about datacenters and cloud if you really want to. Or keep doing what you're good at. I don't care.

 

[cucker tarlson]

Those numbers also make a lot of assumptions and generalizations. It's far more complicated than that, and you know this.

is it ?
R7 perf per wat numbers aren't that much better than Vega,still behind 1080Ti

Spoiler

and it's mostly due to massive bandwidth increase,the process itself left most of us unimpressed.
I'm a simple guy,I look at the basics mostly and prefer numbers to intricate theories
I don't really jump on hype trains either.Like, I do understand the high expectaions for Zen 3000,but does early 7nm really produce that much better parts than refined 14/16nm ? And do we really need two ccx's and two dies for a six core cpu in 2019 ?

 

[R0H1T]

Learn something about datacenters and cloud if you really want to. Or keep doing what you're good at. I don't care.

There's a few things you need to clear up yourself, what sort of cloud you're talking about? Moving data, lots of data, serving customers around the globe - yes that's where AWS, Azure, Google work wonders. Need to crunch numbers, rendering, sensitive code or proprietary stuff - nope, lot of that is done locally & for good reason. If you really think cloud was where everything was headed, how about you look at AMD & Intel's retail (HEDT included) numbers? If not then check them out, not everything is about multi billion dollar companies.

 

[lexluthermiester]

(piracy stuff put aside)

?

You have absolutely no idea about enterprise computing.

That's an assumption. But again, I'm not talking about enterprise and business needs where the cloud has tremendous application potential. I'm talking about the general public in reference to cloud computing/gaming as you stated earlier.

It also seems you have limited understanding of how casual PC users do stuff today (you don't believe in cloud, you're not very enthusiastic about laptops etc).

Not at all. I sell/service more laptops than desktops. I understand my client base very well. Most people have no interest in cloud computing/gaming/storage.

However, we're a bit off-topic here. Let's agree to disagree and rope it in a bit.

is it ?
R7 perf per wat numbers aren't that much better than Vega,still behind 1080Ti

Spoiler

and it's mostly due to massive bandwidth increase,the process itself left most of us unimpressed.
I'm a simple guy,I look at the basics mostly and prefer numbers to intricate theories
I don't really jump on hype trains either.Like, I do understand the high expectations for Zen 3000,but does early 7nm really produce that much better parts than refined 14/16nm ? And do we really need two ccx's and two dies for a six core cpu in 2019 ?

While your point is easy to see, the jump to 7nm was and remains a difficult one. The performance improvements offered by 7nm are compelling enough and improve enough to be considerable. For Turing to jump to 7nm would be a serious advantage as it would offer that much more performance to that GPU platform, to say nothing of a jump to 5nm.

 

[notb]

Technically they aren't but they are competing with AMD+TSMC as a whole & previously TSMC+ARM in the mobile space. Guess what they lost the latter with billions of dollars down the drain, even when the competition was using inferior nodes.

I'm not sure about that. Yes, an "Atom smartphone" idea was a failure, but I'm not sure if they had high hopes.
Intel's tablet/ultrabook platform is doing perfectly fine.

they aren't even second best in most other categories!

I'd argue that Intel is at the forefront of WiFi, IoT, AI (car AI in particular) and non-volatile memory of all sorts.
They may not be first or second in some products, but they're in top3 of almost everything computer-related. That's not a bad position.

The amount of time & money spent in promoting most of these products is staggering, in some cases they bribed, threatened even contra revenued the competition. Intel is a lot of things but the one thing they're not is a (tech) leader that you'd look upto.

Intel is a huge company, but they still want to grow. They're trying new things all the time. Some will fail, some will stick. You can't expect them to be best at everything they try.
By comparison, AMD has the "comfort" of focusing on CPUs. But if they ever reach those 30-50% they used to have some time ago, they'll need to look into other possibilities as well. In fact exactly that happened more than a decade ago, when they bought ATI.

Look at Samsung. They're a chaebol, they make countless things in very different industries.
At some point Samsung decided they'll make mirrorless cameras. And many agree that what they came up with was the most advanced gear at that time. And you know what? They quit. It didn't match their profit expectations.

The fact that Intel (or any other company) leaves a particular business doesn't mean they failed to deliver a good product. It means it doesn't provide the financial figures they require.

Cloud will never replace local computing, not unless Google, MS, Amazon start subsidizing their cloud.

"Cloud" is not a paid service that Google, MS and Amazon sell you. Cloud is a concept. You can set your private cloud on your private server.

 

[cucker tarlson]

?

That's an assumption. But again, I'm not talking about enterprise and business needs where the cloud has tremendous application potential. I'm talking about the general public in reference to cloud computing/gaming as you stated earlier.

Not at all. I sell/service more laptops than desktops. I understand my client base very well. Most people have no interest in cloud computing/gaming/storage.

However, we're a bit off-topic here. Let's agree to disagree and rope it in a bit.

While your point is easy to see, the jump to 7nm was and remains a difficult one. The performance improvements offered by 7nm are compelling enough and improve enough to be considerable. For Turing to jump to 7nm would be a serious advantage as it would offer that much more performance to that GPU platform, to say nothing of a jump to 5nm.

we'll see,they may have reserved the best 7nm so far for ryzen.

 

[M2B]

current 7nm is just the prelude to 7nm EUV that's the real deal.Nvidia is waiting for 7nm EUV (7nm+) cause current 7nm that we see on RVII and Zen 3000 is not really that much of a game changer.
TSMC's 12nm (refined 16nm) that it's supplying for turing works almost as good as 7nm on R7.You can see R7 clocks pretty badly given how vega liquid could do 1700mhz stock on 14nm and R7 needs water cooling capable of almost 500W to reach 2100mhz, that 2080Ti comes close to doing on air on refined 16nm.
process number is not the whole story.

The 7nm EUV is just a different solution to reduce the manufacturing costs, as far as I know.
It's not necessarily more efficient than the current Non-EUV process.
And keep in mind Nvidia did gain a lot more performance by going from 28nm to 16nm than AMD did so Radeon VII might not be the best example.

 

[Space Lynx]

Before we get maxed out on 5nm, almost all computing activities will be moved to cloud - gaming included (at least for the people that can accept the idea).
And you'll be looking like a dinosaur with your ever-growing chiplet CPU.

Incorrect. Even Google Stadia is causing deaths to be due to latency issues, the problem of latency will never be solved unless the entire nation moves to advanced fiber optics point to point across entire nation, and that will never happen, even Google Fiber is cancelling some cities it had planned and pulling out of others due to failure.

Have fun with your 60 fps 4k on Google Stadia I will be doing 165hz 1440p thanks, take care now.

rumors are 3700x hits 5ghz boost. if true and I can OC that to be true for all 8 cores... Intel might even lose the IPC battle.

 

[moproblems99]

Let's review...

So, you're telling me that AMD is going to skip 7nm EUV, the process that is basically for these higher power chips, and go straight to 5nm all in a year? How long were stuck on 12/14/16 and all of a sudden we are going from 12 to 5 in one generation?

Risk Production for 7nm started around April 2017. It is now April 2019 and we still don't have meaningful 7nm. I don't know if I really count VII, it is 7nm but it may as well not have been.

So if we take that 7nm Ryzen launches at Computex and 5nm risk started in January, two and half more years puts us at June 2021. Do you think AMD is going to sit on the 3000 series until 2021? And of course 5nm is going to be perfect on the first try. I mean it isn't like Intel is having problems or anything.

I don't.

cut down vega die (60 cu) at 1750mhz with 300w tdp
full vega die (64 cu) at 1670mhz with 345w tdp
not a game changer.

Keep in mind that clocks aren't solely determined by node but also by architecture.

 

[comtek]

But the interesting part is what I've read over time places the physical limit of a Si transistor at or around 5nm. The next decade will get really interesting.

Organic Microprocessors.
We feed them with Lamb Steak and Orange Juice Instead of electricity.

 

[TheoneandonlyMrK]

So they have the backbone and dev tools in place to make something called 5Nm now, in initial risk production status, Once you have actually finished a design using these process guide and control tools.

That's 9 months before we hear how good this is doing at risk production is'nt it?

Probably not going to rule the world this year , possibly for 2020 h2 phones of the top end variety , If it all goes well.

And personally i expect cloud uptake for gaming to be big but , the bells have tolled for pc gameing before and the cutting edge can still only be had one way ,PC.

it will shrug off this assault (cloud) for a great many years yet imho, since too many things are not quite there universally ie bandwidth.

 

[Shatun_Bear]

current 7nm is just the prelude to 7nm EUV that's the real deal.Nvidia is waiting for 7nm EUV (7nm+) cause current 7nm that we see on RVII and Zen 3000 is not really that much of a game changer.
TSMC's 12nm (refined 16nm) that it's supplying for turing works almost as good as 7nm on R7.You can see R7 clocks pretty badly given how vega liquid could do 1700mhz stock on 14nm and R7 needs water cooling capable of almost 500W to reach 2100mhz, that 2080Ti comes close to doing on air on refined 16nm.
process number is not the whole story.

Nope so much wrong in this post. The first gen 7nm is 'the real deal' in terms of performance gains/power savings over TSMC's 12nm+ process. 7nm EUV doesn't bring such a jump it's just laying the groundwork for EUV to be more extensively used in 5nm, which is more like a true EUV process, which is what you think 7nm EUV is.

 

[cucker tarlson]

Nope so much wrong in this post. The first gen 7nm is 'the real deal' in terms of performance gains/power savings over TSMC's 12nm+ process. 7nm EUV doesn't bring such a jump it's just laying the groundwork for EUV to be more extensively used in 5nm, which is more like a true EUV process, which is what you think 7nm EUV is.

okay,but could you please learn what the -1 button is for ?

 

[RH92]

current 7nm is just the prelude to 7nm EUV that's the real deal.Nvidia is waiting for 7nm EUV (7nm+) cause current 7nm that we see on RVII and Zen 3000 is not really that much of a game changer.
TSMC's 12nm (refined 16nm) that it's supplying for turing works almost as good as 7nm on R7.You can see R7 clocks pretty badly given how vega liquid could do 1700mhz stock on 14nm and R7 needs water cooling capable of almost 500W to reach 2100mhz, that 2080Ti comes close to doing on air on refined 16nm.
process number is not the whole story.

You can't compare 7nm with 12/16nm just by comparing R7 and 2080Ti clocks you just can't since the architectural differences are huge and play a big role in clock speeds .

 

[Shatun_Bear]

okay,but could you please learn what the -1 button is for ?

That's not me downvoting your posts, it's the words in them that's responsible for that.

Listen, you'll see how much of a 'real deal' TSMC's first gen of 7nm is when our fave Nvidia uses it for their next generation of graphics cards. It will blow Turing out the water believe me. Radeon 7 should not be used as a gauge of its merits, as that is basically a tweaked 7nm Pro card (M160) repurposed for gaming.

 

[2901BitSlice]

Also, don't unify TSMC and AMD. TSMC is just a supplier - they sell to the highest bidder. If Intel decides to become a TSMC client, there will be no supply left for AMD.

It would be a Hot Day on Pluto before TSMC takes business from Intel at any price. The risk of Intel lifting TSMC's Technology to back engineer their own manufacturing process is very high. Why would TSMC throw away their trade secrets ?

 

[bug]

Quantum Guard? Do you realize how funny that sounds?

I do. I meant it that way

>5nm
That's a marketing name of the process. Actual sizes proposed for TSMC's 5nmprocess is 44nm for transistor gate pitch and 32nm for the interconnect. This is nowhere the sizes where quantum mechanics start getting in the way.

I know, that why I said things will get interesting in the next decade. We haven't hit that limit yet, we may do so in 10 years.

 

[efikkan]

Not really, they will just move to a chiplet design, well AMD already, have you seen pictures of Ryzen 3700? There is room for another chip on the die. When everyone is on 5nm and maxed out, they will just make the die bigger and add more chiplets and scale it.

IPC will probably be dead though.

AMD have themselves stated that they expect clocks to drop with the new nodes, and nothing so far indicates that either TSMC, Samsung or Intel at their new or upcoming nodes will reach 6-7 GHz. My expectations is that we are near or at the peak of clock speeds with the current type of semiconductor technology.

The new nodes will make it possible to squeeze more cores into small dies, but it will come at a cost. Firstly, the node efficiency gains alone are not enough to maintain clock speeds, and current Coffee Lake and Zen+ CPUs are already in throttle territory. The other problem with more cores is diminishing returns, as most non-server workloads are synchronized. This means that in reality adding many more cores doesn't really compensate for slower cores, in fact, with ever-increasing core count core speed becomes more important to maintain good performance scaling.

So, with clock speeds stalled once again, the future of performance scaling is pretty much depending on what we commonly refer to as "IPC". Intel's upcoming Sunny Cove/Ice Lake and Golden Cove(2021) will both feature IPC gains. The other big area of improvement is SIMD, namely AVX. Ice Lake will bring AVX-512 to the mainstream, and while Zen 2 will bring an appreciated doubling of AVX2 performance, it still lacks AVX-512 for now. AVX is in generally underutilized, which is sad, since AVX and multithreading combined scales incredible well for both vendors.

The 7nm EUV is just a different solution to reduce the manufacturing costs, as far as I know.

It's not necessarily more efficient than the current Non-EUV process.

EUV is expected to increase production speed and improve yields, perhaps even giving a small performance increase too.

Both Intel and TSMC has struggled a lot on their new nodes. It's been a year since Intel started shipping their first 10nm parts, and the volumes have been really low, and the yields I guess are embarrassingly low.
TSMC have been somewhat more successful with their "7nm" HPC node producing medium sized chips, but still unable to ship anything in volumes. I'm really curious how this is going to pan out for Zen 2, which luckily have their "chiplet" design, but I still wonder what kind of tricks TSMC may have up their sleeve, since shipping "hundreds" of chips this time is not going to cut it.

And keep in mind Nvidia did gain a lot more performance by going from 28nm to 16nm than AMD did so Radeon VII might not be the best example.

I think the point was that Vega 64 => Radeon VII is the same architecture on two different nodes, and with a roughly comparable configuration.
Maxwell => Pascal featured some architectural improvements and a large increase in shader processors, which is something Radeon VII did not.

 

[Vayra86]

And for many years, when Intel was the only party doing advanced CPUs, that might have been true. Not because of Intel's greatness. Semiconductor and processor technology is not limited by some corporations' R&D. What you can buy today is basically what our civilization is capable of at given moment.
GPUs and ARM were few years behind on tech, because they didn't need to be on the edge. They kept using well known, cheaper fab node that has been around for longer.

ARM: because no one knew how to use its potential.
GPUs: because they were used for gaming and no one cared. I mean: we had some GPUs that gave us some fps in some games. No benchmark. We really didn't know if that's the limit of this tech.
But then Nvidia started improving performance by 20% yearly. With CPUs we're getting ~5% yearly because of tech limits. This means gaming GPUs were many years behind.

Maybe they will, maybe they won't. It's not that important. Intel is making their own CPUs, so they aren't competing with TSMC.
Intel did make a 10nm product as a showcase (a tiny CPU for laptops) before TSMC launched 7nm. They have the tech. It just wasn't profitable.


I don't see this "copying". MCM is a very old idea, which both Intel and AMD (among many other companies) utilized over the years.
Now, what Intel does in marketing (calling competition's product "glued") is something totally separate from what they do in engineering. It's better to make marketing mistakes and good products than other way around.
And MCM is a huge compromise - something that should be seen as the last resort. So yes, Intel tries to avoid it at all cost, but at this moment they didn't manage to compete with EPYC without it. When they move to smaller node with good yield, maybe MCM won't be needed anymore.


I'm not sure what you meant here (you gave few answers but haven't partitioned my post). The cloud part? I meant exactly what I is written there.
Computing will be covered by cloud in 5 years tops. By "covered" I mean: you won't need a high performing PC at all, for any task.
Today you still need to do some things locally - gaming being the obvious example. But I'm sure you've noticed we're getting awfully close.

And of course cloud will always be priced to compete with intermittent hardware use. So if you game for 2-3h a week, cloud should be cheaper. But if you run a computing node 24/7, hardware will remain the cheaper option.

This is 20/20 hindsight speak, and really has little value, sorry mate. GPU was never 'behind'.

What we have today is not what our civilization is capable of, it is what is economically feasible. Totally different approach, but far more realistic if you ask me. You act like CPU development is carried along in great science endeavours like CERN - its not. Its a completely different ball game with a completely different dynamic. This is mostly true because CPU/GPU performance is scalable and its a mass consumer product.

Node shrinks are one primary example of an economical problem. Its only worth it if you can sell it. But we can do it no problem, it just takes a huge amount of time and resources. For GPU, the reason it developed as it did was because the demand changed over time.

 

[stimpy88]

I would bet that this "5nm" process is still worse than Intel's 10nm process. But I guess it has availability going for it! Samsung, TSMC et al have always been very fast and loose on what they base their xxNM claims on, as a lower number always looks great on the marketing bumpf.

However, I would say that these fabs have almost caught up with Intel.

 

[TheGuruStud]

I would bet that this "5nm" process is still worse than Intel's 10nm process. But I guess it has availability going for it! Samsung, TSMC et al have always been very fast and loose on what they base their xxNM claims on, as a lower number always looks great on the marketing bumpf.

However, I would say that these fabs have almost caught up with Intel.

Yeah, when 10nm is worse than 14 lol. Good one.

 

[bug]

Yeah, when 10nm is worse than 14 lol. Good one.

I believe he's referring to how TSMC and Samsung count their nm to make it look like they have a lead on Intel. And they do technically, because they can draw smaller features. But when you take a step back and look at the bigger picture, they're not really offering higher transistor density, which is what matters at the end of the day.
However, all of that is purely academic, 7nm you can buy trumps 10nm you can't, no matter the actual density

 

[Gungar]

But the interesting part is what I've read over time places the physical limit of a Si transistor at or around 5nm. The next decade will get really interesting.

The process node is called 5nm, but nothing is made in 5nm. Just marketing names.

 

[Fx]

"Cloud" is not a paid service that Google, MS and Amazon sell you. Cloud is a concept. You can set your private cloud on your private server.

No. Cloud is indeed understood (as a whole) that it is external computing services and storage provided by external businesses. Yes, you can set up your servers to perform as a self-serving, on-premises cloud, but that does not redefine what we refer to as the cloud.