Intel 10nm and GF 7nm at IEDM 2017

[goldstone77]

I think it's worth mentioning Samsung with their aggressive process node path, and they are implementing EUV at 7nm. AMD's agreement with GlobalFoundries runs out in 2020, and GlobalFoundries did license Samsung's 14nm LPP design.

Samsung and GlobalFoundries, in partnership with IBM, unveiled the world’s first 5nm silicon chip based on EUV technology and GAAFETs earlier in the year. In Samsung’s recent roadmap, the company is expecting to debut GAA technology by the time it reaches 4nm, replacing FinFET which will likely be exhausted after 5nm, even with new materials. The downside of GAAFETs is that the nanowiring is more difficult to manufacture than traditional FinFETs, and the process will therefore be much more expensive.

https://www.androidauthority.com/4nm-processing-node-812959/

 

[CatMerc]

AMD does benefit from having a close partnership with GloFo. The fab being as reliant as they are on AMD's business and success means that they have every incentive to create processes in collaboration with AMD to suit their exact needs.

7nm is already an example of that, with GloFo targeting the high performance segments of the market while TSMC satisfies Apple's low power SoC requirements.

After the WSA expires, they'll only be more motivated to make processes to suit AMD, as now AMD will be more free in choosing others, while GloFo will still need their business.

 

[goldstone77]

GlobalFoundries has been riddled with problems in the past. That is the reason why the had to abandon their own 14nm process, and license Samsung's 14nm design. I think/hope GlobalFoundries has upped their game with the inclusion of IBM personal and assets. The agreement AMD has with GlobalFoundries limits AMD to GlobalFoundries to the extent that they still have to by wafers from them no matter what, but they can use other partners like Samsung and TSMC. It's an expensive partnership, and AMD will be better off and more profitable in 2020 without being beholden to GlobalFoundries. But it does create a make it or brake it situation for GlobalFoundries to perform, and they know it!

 

[Ajay]

http://www.zdnet.com/google-amp/article/iedm-2017-intel-inches-closer-to-euv-lithography/

TL : DR

Using when viable

From the article:

The exposure source, however, continues to cause issues with overall system uptime. The industry goal is 84 percent this year and 88 percent in 2018, but "we are lucky to be at 75 percent tool availability," Turkot said, due to long periods of unscheduled downtime.

AMSL will have a 250W production machine soon, but the problem for Intel, at the moment, appears to be down time on the LPP (Laser Produced Plasma) module. 25% downtime does look horrendous for a manufacturing concern.

Edit: typo

 

[LTC8K6]

It seems to say 25% down time, which is still bad.

 

[Ajay]

7HPC for desktop Ryzen CPU is a good choice. Base clocks will be at the sweet spot of the 7HPC curve (close to start of the 7HPC curve in the graph below). Single core turbo will be at the top end of the voltage/freq curve. So for stock performance efficiency will be decent. But expect that to go out the window if you want to hit 5 Ghz across all cores.

You make a fair case, but I'll believe it when I see it. I suspect that 7HPC will not have as favorable a yield curve as 7LP. Plus, I don't think AMD is doing enough volume yet to bifurcate it's enthusiast and server CPUs.

 

[Lodix]

That is news to me, but I am not aware of any other product using 10nm anyway, so this 10nm is more like an testing ground for 7nm. Which was when I thought would be TSMC using quad patterning.

Both TSMC and GF has a much larger market to help with R&D compared to the old days. So it is not surprising Intel lost / will lose the lead soon.

It is on millions of phones/tablets from Huawei and Apple.

 

[IntelUser2000]

Both TSMC and GF has a much larger market to help with R&D compared to the old days. So it is not surprising Intel lost / will lose the lead soon.

It's that, and Intel screwing up.

I've seen this happen in an extreme scale with mostly failed companies like Nokia and RIM(Now called Blackberry).

At the time when they should be putting their foot down so they do not further lose the lead, they instead made decisions that makes things worse. At least in Intel's case, their venture with mobile chips helped them with the low power laptops and ultrabooks, which probably stemmed the tide somewhat. They may have focused on it too much, because the process fallout may be due to the over-investment during the 22nm days, causing hubris, and thus a struggle with subsequent generations. That's affecting all their lines right now.

 

[CatMerc]

http://www.zdnet.com/article/iedm-2017-globalfoundries-announces-7nm-chipmaking-process/

Overall GlobalFoundries is promising a 2.8x increase in density, and either a 40 percent boost in performance or 55 percent reduction in power at the same performance. The high-performance version can deliver another 10 percent boost

HPC doesn't appear as appealing when put that way. That said Fmax should be significantly higher anyway.

 

[goldstone77]

http://www.zdnet.com/article/iedm-2017-globalfoundries-announces-7nm-chipmaking-process/


HPC doesn't appear as appealing when put that way. That said Fmax should be significantly higher anyway.

Yeah, but don't forget it's going to wrapped up in a 6 core ~$200 and 8 core for ~$329. Pretty nice

 

[Ajay]

Yeah, but don't forget it's going to wrapped up in a 6 core ~$200 and 8 core for ~$329. Pretty nice

How do you know what AMD's pricing will be?

 

[goldstone77]

How do you know what AMD's pricing will be?

That was a guess that prices would be relatively the same based on previous pricing. Of course we won't know until AMD makes an official announcement.

 

[raghu78]

http://www.zdnet.com/article/iedm-2017-globalfoundries-announces-7nm-chipmaking-process/

HPC doesn't appear as appealing when put that way. That said Fmax should be significantly higher anyway.

7HPC is all about 5+ Ghz for IBM/AMD. Obvious tradeoff - leakage power. But who cares about power when overclocking to 5 Ghz. I think we could see a 12C/24T Ryzen 3000 series with base clocks in the 3.6-4.0 Ghz range and max single core turbo of 5 Ghz.

 

[goldstone77]

7HPC is all about 5+ Ghz for IBM/AMD. Obvious tradeoff - leakage power. But who cares about power when overclocking to 5 Ghz. I think we could see a 12C/24T Ryzen 3000 series with base clocks in the 3.6-4.0 Ghz range and max single core turbo of 5 Ghz.

I think the switch to 7.5T libraries at 12nm will gives us a good idea what to expect(Fmax) with 7nm and 6T libraries. 6 core per CCX(6 core gaming cpu), and if they can stabilize IF frequency/reduce intra core latency would be huge for Ryzen, and I hope this is the path they take. We could see over 5GHz(5.3GHz maybe). We live in interesting times!

 

[Arachnotronic]

Both TSMC and GF has a much larger market to help with R&D compared to the old days. So it is not surprising Intel lost / will lose the lead soon.

GloFo's annual revenue is somewhere in the vicinity of $6 billion, Intel's is >$60 billion. The "GloFo has more R&D capability" meme is really silly because it's just not true.

Even TSMC's annual revenue is only in the neighborhood of $30 billion, or about half of Intel's. If Intel falls behind, it's not because GloFo/TSMC could outspend them, it's because Intel executed poorly.

 

[Ajay]

7HPC is all about 5+ Ghz for IBM/AMD. Obvious tradeoff - leakage power. But who cares about power when overclocking to 5 Ghz. I think we could see a 12C/24T Ryzen 3000 series with base clocks in the 3.6-4.0 Ghz range and max single core turbo of 5 Ghz.

Well power still matters when overclocking large core count CPUs. Just look at the issues with Skylake-X with over driven VRMs.

 

[raghu78]

Well power still matters when overclocking large core count CPUs. Just look at the issues with Skylake-X with over driven VRMs.

The issue with Skylake-X was poor cooling for VRMs . Motherboard designers had to rush designs as Intel pulled in the launch. The newer boards are now having much better VRM cooling.

https://www.extremetech.com/computi...mperatures-disaster-limit-skylake-x-potential

 

[beginner99]

http://www.zdnet.com/article/iedm-2017-globalfoundries-announces-7nm-chipmaking-process/

Overall GlobalFoundries is promising a 2.8x increase in density, and either a 40 percent boost in performance or 55 percent reduction in power at the same performance. The high-performance version can deliver another 10 percent boost

HPC doesn't appear as appealing when put that way. That said Fmax should be significantly higher anyway.

I mean the 40% boost in performance we know is obviously fake. 1800x has 3.6/4.0 and 40% more would be 5.0/5.6. Never going to happen. But that isn't a huge issue as even something like 4.2/4.7 would already be huge improvement and pretty bad for intel.

7HPC is all about 5+ Ghz for IBM/AMD. Obvious tradeoff - leakage power. But who cares about power when overclocking to 5 Ghz. I think we could see a 12C/24T Ryzen 3000 series with base clocks in the 3.6-4.0 Ghz range and max single core turbo of 5 Ghz.

HPC is for IBM only, obviously. AMD server chips will need to be on SOC and hence so will be all other Zen2-based chips. AMD doesn't have the cash for separate designs (and masks). Just not worth it. Better to sell lower performance for lower price. The cool thing about current design is that the can recuperate the cost of R&D by selling server chips. The consumer chips don't need to have R&D cost included in price and hence can sell at a relatively low price and still make a lot of money. If you include R&D price will go up and volume will go down. This is also against what AMD needs. they need high volume to build market share and because of the WSA.

 

[Yotsugi]

I mean the 40% boost in performance we know is obviously fake.

Fake news?
Dude that's an actual HP node.

1800x has 3.6/4.0 and 40% more would be 5.0/5.6. Never going to happen. But that isn't a huge issue as even something like 4.2/4.7 would already be huge improvement and pretty bad for intel.

L-lies? L-lies!

AMD doesn't have the cash for separate designs (and masks).

Thaaaaaaaat's why they are doing client-only Pinnacle Ridge.

 

[itsmydamnation]

Fake news?
Dude that's an actual HP node.

L-lies? L-lies!

Thaaaaaaaat's why they are doing client-only Pinnacle Ridge.

Doing one extra die every two years is very different cost model to two extra dies every two years.

 

[Gideon]

I mean the 40% boost in performance we know is obviously fake. 1800x 3.6/4.0 and 40% more would be 5.0/5.6. Never going to happen. But that isn't a huge issue as even something like 4.2/4.7 would already be huge improvement and pretty bad for intel.

(emphasis mine)

IMO the 4.0 turbo claim is often very poorly understood by people who don't own Ryzen chips.

I cannot comment on the 1800X , but in my experience, my "3.4/3.8" 1700X with Tomahawk B350 might as well be labelled as a 3.5 Ghz chip, period (and I think it's exactly the same for 1800 @ 3.7).

3.5 is it's all-core-boost and @stock it is glued to that frequency 99,9% of the time. Yes, 3.5. It doesn't matter if your temps are excellent (e.g <40°C die with my Noctua cooler). Even if you close everything and run only a single-threaded benchmark (e.g. CPU-Z or Cinebench) it will stay at 3.5 Ghz. Likewize, you can run Prime95 on all cores for hours ... it will still be 3.5 and never need to downclock base.

I have played with power-profiles, BIOS settings etc ... and I have literally seen ST turbo kick in 3 times (usually when everything is closed in windows and say, opening chrome) and even then only for for intervals of 1-2 secs max (with voltage being >1.4v, while on idle it's 0.9v and usually fluctuating around 1.2v-1.3v).

Is this my isolated experience, or is it the same with more expensive boards?

Anyway, this is why Precision Boost 2 will be such a big deal for a 12nm Ryzen. Even if the Single Core turbo clock is only, say 4.2 - 4.3, but it can actually sustain that clock in lightly-threaded games (dipping maybe 50-100Mhz at times), the actual clock speed improvement would still be closer to 15% rather than 10% compared to 1800X.

(I'm personally hoping for a 4.4+ ST-Turbo clock for top-line Ryzen @12nm, this is just an example)

TL;DR:
A 7nm 1800X successor will "only" have to hit 5.1-5.2 Ghz ST-boost for a 40% improvement, in real- world use cases, because of the limitations of Precision Boost 1.

 

[CatMerc]

I mean the 40% boost in performance we know is obviously fake. 1800x has 3.6/4.0 and 40% more would be 5.0/5.6. Never going to happen. But that isn't a huge issue as even something like 4.2/4.7 would already be huge improvement and pretty bad for intel.



HPC is for IBM only, obviously. AMD server chips will need to be on SOC and hence so will be all other Zen2-based chips. AMD doesn't have the cash for separate designs (and masks). Just not worth it. Better to sell lower performance for lower price. The cool thing about current design is that the can recuperate the cost of R&D by selling server chips. The consumer chips don't need to have R&D cost included in price and hence can sell at a relatively low price and still make a lot of money. If you include R&D price will go up and volume will go down. This is also against what AMD needs. they need high volume to build market share and because of the WSA.

The figures are at a certain voltage level. 40% doesn't necessarily mean 40% higher Fmax, just that at some point on the power/frequency curve, you can get 40% higher performance at the same power.

That said, 7nm Fmax is significantly higher than 14nm LPP, don't be surprised to see higher than 5GHz boosts.

 

[majord]

(emphasis mine)

IMO the 4.0 turbo claim is often very poorly understood by people who don't own Ryzen chips.

I cannot comment on the 1800X , but in my experience, my "3.4/3.8" 1700X with Tomahawk B350 might as well be labelled as a 3.5 Ghz chip, period (and I think it's exactly the same for 1800 @ 3.7).

3.5 is it's all-core-boost and @stock it is glued to that frequency 99,9% of the time. Yes, 3.5. It doesn't matter if your temps are excellent (e.g <40°C die with my Noctua cooler). Even if you close everything and run only a single-threaded benchmark (e.g. CPU-Z or Cinebench) it will stay at 3.5 Ghz. Likewize, you can run Prime95 on all cores for hours ... it will still be 3.5 and never need to downclock base.

I have played with power-profiles, BIOS settings etc ... and I have literally seen ST turbo kick in 3 times (usually when everything is closed in windows and say, opening chrome) and even then only for for intervals of 1-2 secs max (with voltage being >1.4v, while on idle it's 0.9v and usually fluctuating around 1.2v-1.3v).

Is this my isolated experience, or is it the same with more expensive boards?

Anyway, this is why Precision Boost 2 will be such a big deal for a 12nm Ryzen. Even if the Single Core turbo clock is only, say 4.2 - 4.3, but it can actually sustain that clock in lightly-threaded games (dipping maybe 50-100Mhz at times), the actual clock speed improvement would still be closer to 15% rather than 10% compared to 1800X.

(I'm personally hoping for a 4.4+ ST-Turbo clock for top-line Ryzen @12nm, this is just an example)

TL;DR:
A 7nm 1800X successor will "only" have to hit 5.1-5.2 Ghz ST-boost for a 40% improvement, in real- world use cases, because of the limitations of Precision Boost 1.

Your experience is very isolated - I've never experienced this with any Ryzen chip - All run at least at there max precision boost (non XFR) frequency under single core loads- the XFR frequency, sure, it can be more 'rare' to see.

perhaps you're not monitoring clocks at a fast enough interval - Try locking threads to a core for testing purpose. Otherwise it can appear like it's not running consistently as the thread bounces around cores, and boosted core switches rapidly with it.



This can further be verfied using known benchmarks. A stock 1800x wouldn't get the ~160-165 ST r15 score it does if it wasn't running at least 4Ghz. As proven by the fact if you pin it to 4Ghz (typical max 24/7oc) your ST scores barely change, if at all.

 

[Gideon]

This can further be verfied using known benchmarks. A stock 1800x wouldn't get the ~160-165 ST r15 score it does if it wasn't running at least 4Ghz. As proven by the fact if you pin it to 4Ghz (typical max 24/7oc) your ST scores barely change, if at all.

Thanks, definitely need to check that out! Good to know that i might have been wrong.

I just assumed it as even my ancient 2500K would often clock up all the way to 3.7 Ghz, even in semi-multithreaded workloads like some games (though it didn't stay there for very long obviously). And I haven't seen anything in Ryzen master, Windows task manager, in AIDA64, that MSI Afterburner overlay ... you name it.

Anyway it might be a monitoring mistake indeed, as my 1700X seems to have slightly higher ST results in CPU-Z single-thread, than the reference score (memory @ 3466 CL-16-18-18 shouldn't really change that, right?).

 

[raghu78]

I mean the 40% boost in performance we know is obviously fake. 1800x has 3.6/4.0 and 40% more would be 5.0/5.6. Never going to happen. But that isn't a huge issue as even something like 4.2/4.7 would already be huge improvement and pretty bad for intel.

The comparison is not for Fmax but at a certain point on the voltage/freq curve. GF 7LP has 2 versions -
1. High density 7SoC 6T . H240 cell . 2 fin cell. 13 Metal layers
2. High performance 7HPC 9T. H360 cell . 4 fin cell. 17 Metal layers

7SoC 6T is optimized for 3.5 Ghz. This means at those clocks you will get the best combination of perf and power efficiency.

https://www.eetimes.com/document.asp?doc_id=1330467&page_number=2&piddl_msgpage=2

The process is geared for the same kinds of premium cloud and smartphone processors as today’s 14/16nm nodes. An ARM Cortex-A72 core could run at more than 3.5 GHz in the process, the company estimates.

https://www.globalfoundries.com/sites/default/files/product-briefs/7lp-product-brief.pdf

7HPC is designed for 5+ Ghz operation . The obvious tradeoff is density and power efficiency.

HPC is for IBM only, obviously. AMD server chips will need to be on SOC and hence so will be all other Zen2-based chips. AMD doesn't have the cash for separate designs (and masks). Just not worth it. Better to sell lower performance for lower price. The cool thing about current design is that the can recuperate the cost of R&D by selling server chips. The consumer chips don't need to have R&D cost included in price and hence can sell at a relatively low price and still make a lot of money. If you include R&D price will go up and volume will go down. This is also against what AMD needs. they need high volume to build market share and because of the WSA.

AMD has already gone with a separate client die at 12LP. This chip is very likely to built using high performance libraries while the first gen Ryzen used 14LPP High density libraries. So AMD had already gone with separate dies for server and client. Going forward 7nm Rome server chips are going to built at 7SoC 6T while desktop Ryzen will use 7HPC.