TSMC Begins Massproduction Of 7nm chips.

[da_knug]

SOURCE: https://wccftech.com/amd-7nm-vega-gpu-zen-2-cpu-mass-production-tsmc/

 

TSMC has just confirmed that they've begun Massproduction of their 7nm lithography, confirming that AMD will be contracting them to produce their Vega+ GPUs as part of their Radeon Instinct and Radeon Pro lineups. These GPUs are expected to hit market during H2, 2018.

 

AMD is also expected to use TSMC for their upcoming Zen 2, 7nm, processors. 

 

TSMC's production capacity of their 7nm lithography is expected to increase three-fold by the end for the year, this is what WCCFTech says about it:

"This means that TSMC will be able to deliver a mass total of 1.1 million units by the next year, a three-fold increase over the current year. The AMD 7nm OEM orders are expected to being production in Q4 2018 and TSMC is expected 7nm to account for up to 20% of their Q4 2018 revenue. As for other nodes, TSMC will be beginning mass production on the 7nm+ using EUV technology next year, which will further reduce power consumption by 10%, compared to 7nm."

[DrMacintosh]

RIP Nvidia, and RIP my wallet from buying all the AMD Cards  

[Okjoek]

1 minute ago, DrMacintosh said:

RIP Nvidia, and RIP my wallet from buying all the AMD Cards  

Is a node size change really going to give AMD that kind of an advantage shift?

[TheGlenlivet]

Is this new architecture or just die shrunk existing architecture?

[Taf the Ghost]

2 minutes ago, TheGlenlivet said:

Is this new architecture or just die shrunk existing architecture?

We're expecting Navi to be on GloFo's process. Vega on 7nm+ is a shrink.

[DrMacintosh]

6 minutes ago, Okjoek said:

Is a node size change really going to give AMD that kind of an advantage shift?

I don't expect AMD to release a 1080Ti killer, but I also don't care since that part of the market is irrelevant to me. 

 

I do expect AMD to continue pushing out Nvidia in the mid range and budget options. (1060 and down) And if AMD can give the resources that Vega needed at the start to their next gen GPUs*, I bet they could challenge the 1070 and win. 

 

*

 

[TheGlenlivet]

1 minute ago, Taf the Ghost said:

We're expecting Navi to be on GloFo's process. Vega on 7nm+ is a shrink.

Ah, OK thanks.

So it will be an improvement, but not a radical one.  Hopefully heat and voltage are better controlled.  It would be great if AMD could compete at the 170/1080 level with Nvidia.

[Taf the Ghost]

2 minutes ago, TheGlenlivet said:

Ah, OK thanks.

So it will be an improvement, but not a radical one.  Hopefully heat and voltage are better controlled.  It would be great if AMD could compete at the 170/1080 level with Nvidia.

Space, yield, better clocks & lower power draw, more than likely. The design works fine for the Instinct/ML stuff, so it's a good product to shrink first to a new node. "Pipe cleaner", sure, but it is a product where you can make up the cost on low-ish volume.

 

Expect Navi (Polaris replacement) and Desktop Ryzen 3rd Gen on GloFo. Epyc 2 should probably be on TSMC for the "big" die.

[Okjoek]

1 minute ago, DrMacintosh said:

I don't expect AMD to release a 1080Ti killer, but I also don't care since that part of the market is irrelevant to me. 

 

I do expect AMD to continue pushing out Nvidia in the mid range and budget options. (1060 and down) And if AMD can give the resources that Vega needed at the start to their next gen GPUs*, I bet they could challenge the 1070 and win. 

 

*

 

Yeah, that's why I got an RX 460 personally. I like the market for the cards that use ~75 watts or less. Usually nothing at that point costs more than 100-150 USD and I can get well over 60 FPS in the games that matter to me. The Vega in the APUs is so good now I wish they'd give it a better VRAM solution than DDR4. I remember hearing a rumor that AMD was working closely with the RAM producers on future DDR5 memory. I hope it will do a better job for graphics or perhaps they could just build VRAM into the package thanks to increased density or what I don't get is why VRAM can't have its own DIMM slot on the motherboard? There's probably good answers to those questions, I just don't know them.

[ravenshrike]

33 minutes ago, Aiedail said:

TSMC will be beginning mass production on the 7nm+ using EUV technology next year, which will further reduce power consumption by 10%, compared to 7nm

Should be interesting to see if TSMC encounters similar problems to Intel or if they designed everything better in the first place.

[Christophe Corazza]

I wonder if this is why we haven't seen anything new from Nvidia, maybe they have changed their plans too and shifted their focus from 12 nm to 7 nm.

 

Makes sense in many ways, Volta powers the worlds most powerful supercomputer and the GTX 1080 Ti is still waiting for the competition to turn up.

[S w a t s o n]

5 hours ago, Taf the Ghost said:

Space, yield, better clocks & lower power draw, more than likely. The design works fine for the Instinct/ML stuff, so it's a good product to shrink first to a new node. "Pipe cleaner", sure, but it is a product where you can make up the cost on low-ish volume.

 

Expect Navi (Polaris replacement) and Desktop Ryzen 3rd Gen on GloFo. Epyc 2 should probably be on TSMC for the "big" die.

Wut? I thought GloFo was doing all CPUs with TSMC doing GPUs. Why would AMD split up production of zepplin dies anyways.

[Dionyz]

8 hours ago, Aiedail said:

SOURCE: https://wccftech

............... They spew bs. Basically they are there to hype you about a product. but they don't know the date, so if no hints follow up on the date then WCCFTECH switches up their stance and say "Nvidia over produced on their GPU's" and announces another further date. Basically the hype train never stops till the product is launched 

[Mihle]

2 hours ago, Swatson said:

Wut? I thought GloFo was doing all CPUs with TSMC doing GPUs. Why would AMD split up production of zepplin dies anyways.

AMD said they would do Zen 2 on whoever of them that managed to produce 7nm chips first. It turned out to be TSMC. So at the start at least TMSC will produce for Epyc 2 and when GloFo will do for Threadripper 3 and Ryzen 3 when they are ready.

[Taf the Ghost]

2 hours ago, Swatson said:

Wut? I thought GloFo was doing all CPUs with TSMC doing GPUs. Why would AMD split up production of zepplin dies anyways.

6-8 month advantage in Time To Market. AMD's relationship with GloFo is a little wonky, but TSMC is currently in HVM while GloFo is maybe December at the earliest. 

 

In this generation of Server parts, that really, really matters for AMD. If Icelake-SP keeps slipping, AMD could be on 7nm with more cores & roughly equal IPC for over a year on Intel. That's why we expect Epyc on TSMC's node.

 

If you assume AMD is going to be running two main designs. Big & Little, with both ending up in Server SKUs by next year.

[Bananasplit_00]

Whatever happened to AMD <3 GloFo? Is that over? I thought they had a deal for much longer

[Taf the Ghost]

17 minutes ago, Bananasplit_00 said:

Whatever happened to AMD <3 GloFo? Is that over? I thought they had a deal for much longer

They do. AMD had to play millions to put designs at other foundries, and they still need to buy a certain number (a very large one) of wafers at GloFo. Server CPUs are big, big margins. It's a roughly 40 Billion USD market, the highest revenue & margin market within the x86 landscape. There's a reason to put your big 12c or 16c design on TSMC and get out there early.

[mr moose]

9 hours ago, ravenshrike said:

Should be interesting to see if TSMC encounters similar problems to Intel or if they designed everything better in the first place.

No, the issues Intel are having are due to the lithography process not the node size. Intel have said they will move to the same EUV lithography when they go to 7nm, but they will also be winding back their density targets (10nm was 2.7X more dense than their 14nm while the industry average is 2x at best).

 

 

[Taf the Ghost]

https://twitter.com/witeken/status/1010280303359873024

 

Good chance 10nm & 10nm+ might just be outright skipped. 10nm++ might be the next node from Intel, which is probably late 2020 to 2021. That's kind of why AMD would want to be out as early as possible on 7nm.

[mr moose]

24 minutes ago, Taf the Ghost said:

https://twitter.com/witeken/status/1010280303359873024

 

Good chance 10nm & 10nm+ might just be outright skipped. 10nm++ might be the next node from Intel, which is probably late 2020 to 2021. That's kind of why AMD would want to be out as early as possible on 7nm.

According to Murthy Renduchintala  Intel are not planning on quickly replacing their 10nm process because even though it has had a few issues they have managed to create huge leaps in density and speed using the process.

 

https://www.pcgamesn.com/intel-7nm-production-delays

https://www.tomshardware.com/news/intel-cpu-10nm-earnings-amd,36967.html

 

I am more interested to see where Intel's 10nm performs against AMD at 7nm. All other differences accounted for of course.

[leadeater]

52 minutes ago, mr moose said:

No, the issues Intel are having are due to the lithography process not the node size. Intel have said they will move to the same EUV lithography when they go to 7nm, but they will also be winding back their density targets (10nm was 2.7X more dense than their 14nm while the industry average is 2x at best).

It's pretty much only ever density that makes things hard, lithography can make something possible or a bit easier but if you're gunning for density then it's hard. Any fab that easily gets X node size is a really good indicator that it's not dense and therefore not all that important if we're hoping to see more cores or larger caches etc.

 

An easy 7nm means jack compared to a hard 10nm, and in some cases larger is better so yea will be interesting how 10nm plays out.

[mr moose]

38 minutes ago, leadeater said:

It's pretty much only ever density that makes things hard, lithography can make something possible or a bit easier but if you're gunning for density then it's hard. Any fab that easily gets X node size is a really good indicator that it's not dense and therefore not all that important if we're hoping to see more cores or larger caches etc.

 

An easy 7nm means jack compared to a hard 10nm, and in some cases larger is better so yea will be interesting how 10nm plays out.

I think I read that Intel's 10nm process had 34 stages in it which is where the failures where coming from (more to go wrong with less tolerance for errors), while the EUV 7nm has only 7 process stages.   Better yields from the get go but it will competing with Intel's 10nm that has 2.4X 2.7X the density of 14nm

 

EDIT: sorry, apparently the density of 10nm is actually 2.7X,  2.4X is what they are aiming for to reduce failures.

[GOTSpectrum]

11 hours ago, DrMacintosh said:

I don't expect AMD to release a 1080Ti killer, but I also don't care since that part of the market is irrelevant to me. 

 

I do expect AMD to continue pushing out Nvidia in the mid range and budget options. (1060 and down) And if AMD can give the resources that Vega needed at the start to their next gen GPUs*, I bet they could challenge the 1070 and win. 

 

*

 

If they manufactured enough Vega 64s that sold at MSRP they could have pushed nVidia out of even the 1080 level IMO. The trust is most people shouldn't care about power usage, most of the people I know who has a gaming computer has a way overspec'ed PSU anyway.

[Taf the Ghost]

9 minutes ago, mr moose said:

I think I read that Intel's 10nm process had 34 stages in it which is where the failures where coming from (more to go wrong with less tolerance for errors), while the EUV 7nm has only 7 process stages.   Better yields from the get go but it will competing with Intel's 10nm that has 2.4X the density of 14nm

Before moving to EUV, there's a lot of steps to making a "7nm" generation processor. TSMC & GloFo both got right to the limits of DUV capabilities, so the metal layer & other aspects are large generational improvements rather than completely re-engineered. Intel's 10nm has had several problems. The base metal layer (Cobalt issues) is being reworked, which is a pretty big "we screwed up" admission.

 

Now it seems to be coming out that the cell design is wrong and causes most of the manufacturing errors. If true, this node really is fundamentally broken for Intel.

[mr moose]

6 minutes ago, Taf the Ghost said:

Before moving to EUV, there's a lot of steps to making a "7nm" generation processor. TSMC & GloFo both got right to the limits of DUV capabilities, so the metal layer & other aspects are large generational improvements rather than completely re-engineered. Intel's 10nm has had several problems. The base metal layer (Cobalt issues) is being reworked, which is a pretty big "we screwed up" admission.

 

Now it seems to be coming out that the cell design is wrong and causes most of the manufacturing errors. If true, this node really is fundamentally broken for Intel.

Got some links, because most of what I have read about cobalt is either just about how it is better but nothing in depth explaining why it might be an issue.  This from the article I posted earlier:

Quote

10nm is Intel's last process based on traditional photolithography, and though Krzanich didn’t dive deep into details, he listed the lithography technique as a significant contributor to the low 10nm yields. The company will switch to EUV at 7nm. Currently, Intel's multipatterning process is generating too many yield-reducing defects to produce 10nm cost-effectively.

 

They definitely have silicon issues, that's why many speculate they hired Jim Keller.  But it remains to be seen just how serious these issues are and if they will resolve the lithography issues or not.   

 

Here is what Dick James reported:

https://electroiq.com/chipworks_real_chips_blog/2017/12/18/iedm-2017-intels-10nm-platform-process/