FAKE NEWS from EETimes AMD Beats Intel, Nvidia to 7 nm

[Daniel Nenni]

And this just in from Rick Merritt at EETimes:

AMD Beats Intel, Nvidia to 7 nm

SAN FRANCISCO — Advanced Micro Devices launched its first 7-nm CPU and GPU at the lucrative target of the data center. It showed working chips that delivered comparable performance to Intel’s 14-nm Xeon and Nvidia’s 12-nm Volta.

AMD Beats Intel, Nvidia to 7 nm | EE Times

But:

AMD has yet to reveal many details about the new chips and their performance. However, analysts are generally bullish that the company will be able to continue a significant comeback since it launched its first Zen-based chips on a 14-nm process in late 2016.

Here is the funny part, AMD was late to 7nm in comparison to NVIDIA who is an intimate TSMC collaborator. I can assure you that NVIDIA is ahead of AMD with 7nm silicon, absolutely. The other problem is that Intel is already showing and delivering 10nm CPUs and Intel 10m is greater than or equal to TSMC 7nm in density.

So yes, AMD is the first to claim being the first to 7nm, blah blah blah.

Here is a new tagline: AMD the desperation driven company!

View attachment 22732

 

[Jozo035]

These processors are already sampling to selected customers for several months and i can assure you that they are in much better state than these officially released Intel's 10nm CPUs. This is something i can 100% confirm and Intel will NOT have anything to match them in at least next 2 years (regardless 14 or 10nm).

On the other hand Intel's 10nm launch was more something like paper-launch, you can't really buy these parts either. (there are options to import it from east in limited quantities) Plus they are clearly inferior to 14nm parts in performance and power (batter life) and die size is also fishy and makes proclaimed shrink questionable (14nm CPU+iGPU was 82mm2, 10nm is 72mm2). It looks like releasing early engineering sample...

So currently from what i know (and it is not based on data from EEtimes or AMD PR) AMD is really ahead of Intel.

Regarding Nvidia i must agree. Collaboration level, relationship, design resources... everything is on another level. Only reason why they are not talking about future nodes is because their products are selling well and nobody threatening their position.

 

[Guardsmanbob]

What am I missing here? .. your post refutes nothing in the article, the claim is AMD launched 7nm GPU silicon first, which they did, 3 weeks ago.

Your claim Intel is delivering 10nm silicon, but it is now clear the limited release was PR more than a result of a cost or performance advantage from the new node, and you hold up density as an Intel advantage, but you more than any person here should know CPU's on Intel nodes are far lower than some theoretical density Intel claims, in fact Zen1 is almost twice the density of Intels 14nm CPU's. (zen is 25MT/mm^2 .. Intel 14nm for CPU is ~15MTR/MM^2).

With AMD further disentangling IO ´from logic and pushing scaling harder Intel is going to be even further behind.

You set up your own straw-man by defining 'behind ahead' to mean something entirely different, then attack the article for not using your definition and claim it is fake news.

I used to visit this place weekly, but now its just a collection of incoherent forum rants.

 

[Daniel Nenni]

What am I missing here? .. your post refutes nothing in the article, the claim is AMD launched 7nm GPU silicon first, which they did, 3 weeks ago.

Your claim Intel is delivering 10nm silicon, but it is now clear the limited release was PR more than a result of a cost or performance advantage from the new node, and you hold up density as an Intel advantage, but you more than any person here should know CPU's on Intel nodes are far lower than some theoretical density Intel claims, in fact Zen1 is almost twice the density of Intels 14nm CPU's. (zen is 25MT/mm^2 .. Intel 14nm for CPU is ~15MTR/MM^2).

With AMD further disentangling IO ´from logic and pushing scaling harder Intel is going to be even further behind.

You set up your own straw-man by defining 'behind ahead' to mean something entirely different, then attack the article for not using your definition and claim it is fake news.

I used to visit this place weekly, but now its just a collection of incoherent forum rants.

I hang with the people that actually do the work at AMD, Intel, Nvidia, not the PR people. I also hang with the ecosystem (EDA/IP) and the foundries which is where the rubber meets the road. So if you base your opinions on Google searches and press releases we may never agree, which is fine, but I will never accuse you of incoherent rants out of respect, even if you are not deserving.

In regards to density, I was speaking of process density not die density. The Intel 10nm process is denser than TSMC 7nm and that is a fact. How many transistors you can pack into a chip depends highly on the type of transistors. There is no way to do an apples-to-apples die density comparison unless you account for memory and AMS. If you find two identical digital blocks on two different chips you may come close. Published papers use SRAM blocks on test chips for density brags but that same block may not have made it into production due to yield concerns. I worked for an SRAM company so this is first hand experience.

I agree with the previous post in regards to Nvidia not pre announcing which also applies to Intel:

"Only reason why they are not talking about future nodes is because their products are selling well and nobody threatening their position."

When I mention press releases and supposed leaks of information to the people inside AMD and Intel they roll their eyes at the thought that people take that stuff seriously.

 

[raghu78]

I hang with the people that actually do the work at AMD, Intel, Nvidia, not the PR people. I also hang with the ecosystem (EDA/IP) and the foundries which is where the rubber meets the road. So if you base your opinions on Google searches and press releases we may never agree, which is fine, but I will never accuse you of incoherent rants out of respect, even if you are not deserving.

In regards to density, I was speaking of process density not die density. The Intel 10nm process is denser than TSMC 7nm and that is a fact. How many transistors you can pack into a chip depends highly on the type of transistors. There is no way to do an apples-to-apples die density comparison unless you account for memory and AMS. If you find two identical digital blocks on two different chips you may come close. Published papers use SRAM blocks on test chips for density brags but that same block may not have made it into production due to yield concerns. I worked for an SRAM company so this is first hand experience.

I agree with the previous post in regards to Nvidia not pre announcing which also applies to Intel:

"Only reason why they are not talking about future nodes is because their products are selling well and nobody threatening their position."

When I mention press releases and supposed leaks of information to the people inside AMD and Intel they roll their eyes at the thought that people take that stuff seriously.

Daniel
I do not agree with your statements. AMD is leading both Intel and Nvidia wrt moving to the leading edge. AMD has launched the 7nm Vega 20 based Radeon Instinct MI60 this quarter. AMD are on track to launch 7nm Rome and Navi in Q2 2019 and 7nm Ryzen in Q3 2019. Intel 10nm is not available in high volume and the Cannonlake (CNL) chips are not going to see continued production after the first batch of chips which were produced but had the entire integrated GPU disabled. Its laughable that a tiny 70 sq mm CNL die had more than half of it disabled. The process on which CNL was released is broken fundamentally. Intel is making changes to 10nm to get the yields up. Its rumoured that Contact over Active Gate is out of the upcoming 10nm process. COAG was the main reason for the defective iGPU on CNL. Intel 10nm based client systems are expected to launch for holiday 2019. But thats very much at the ultra low power end like Intel Core M launch back in Nov 2014.

Intel is planning another 14nm generation for desktop called Comet Lake in late 2019 with 10c/20t.
Intel's Comet Lake-S Rumored to Pack 10 Cores, Debut on 14nm - ExtremeTech

Intel's next gen 14nm server parts are going to launch in late Q1 or early Q2 2019. Intel's claims do not match up with reality. Cascade Lake is not a 2018 product but a 2019 product. Similarly Cooper Lake is a 2020 product and ICL will be a 2021 product.

Intel Partner Discloses Cascade Lake Xeon Scalable Launch Window

The way I see it AMD will have a 15-18 month lead to next gen node on desktop and server compared to Intel. On notebooks AMD will not have a lead as the first 7nm APUs will arrive in H1 2020 at the same time as 10nm Icelake for mainstream notebooks.

Nvidia has just launched Turing on 12nm and the earliest I see a 7nm Geforce GPU is late 2019. AMD's Navi products are likely to have atleast a 6 month time to market lead. AMD skipped a full generation on 12nm in favour of an aggressive transition to 7nm Navi given their lesser resources. Navi looks to be a big improvement in power efficiency.

AMD Navi release timeframe leaked, power consumption set to be "surprising" | OC3D News

Lisa Su also stated that they are confident of competing at the high end of the GPU market.

AMD Chief Lisa Su Says the Trade War Is an Opportunity - Barron's

Suzanne Plummer from AMD hinted that they are bringing the best design principles adopted by the Zen CPU core team and some of the leadership like Sam Naffziger and Suzanne Plummer have taken up key leadership roles at RTG to bring a bigger leap in graphics performance and efficiency.

YouTube ( video at 3:20)

https://hardforum.com/threads/raja-koduri-radeon-technologies-boss-leaves-amd.1947651/

I think Intel has been well and truly outsmarted on server design with chiplets by AMD. TSMC now leads the industry on leading edge manufacturing. TSMC will enable AMD to take as much server and PC share as possible using 7nm products. Intel has atleast 3-4 tough years before they can return to the top. Nvidia is in a better and stronger position and will continue to lead but AMD will cut the performance and efficiency gap between them and Nvidia significantly. The way I see it AMD will continue to grow at a strong rate in 2019-2020 and into the future. Intel will lose huge share in servers in 2020-2022. AMD's execution is steady and combined with TSMC's leadership process node execution to 5nm in 2020 and 3nm in 2022 I think the prospects are bright for AMD and grim for Intel. Nvidia vs AMD will also be a closer battle than in the past generations and thats going to be great for consumers.

 

[benb]

AMD stock price has pulled back along with almost all other semis. I think AMD is generally on a path to prosperity. They are unique in one respect, as the only x86 fabless vendor. But it's a frothy stock that seems to have pump and dump media. I think that's what Daniel is pointing his guns at, the long biased pumpers. Technical merits aside.

Broadening the topic somewhat, Apple, Nvidia, and many other stock names seem to own certain media outlets and use them strategically. As traditional media and their traditional ethics have receded, this is what you get. Maybe it's what you get in a pretty dramatic sell-off too. Fear erodes ethics.

 

[IanD]

"The other problem is that Intel is already showing and delivering 10nm CPUs and Intel 10m is greater than or equal to TSMC 7nm in density."

"The Intel 10nm process is denser than TSMC 7nm and that is a fact."

Both are literally true. But both ignore the inconvenient fact that Intel's current 10nm process isn't manufacturable with acceptable yield and never will be in its current form, and that their launched 10nm CPUs are only there so Intel can tell market analysts that "10nm is in production" -- again literally true, but not in mass production in any meaningful way for any useful product.

The "reworked" Intel 10nm process has had to give up several of the features that gave it the much-vaunted higher density and performance (COAG, SAQP interconnect, maybe cobalt interconnect...) which means it's now lower density (and maybe lower performance?) than TSMC 7nm as well as (at least) a year later to market.

Meanwhile TSMC are churning out tens of millions of 7nm APs for the likes of Apple and HiSilicon, with many other big customers in the pipeline (AMD, Qualcomm, Broadcom, MediaTek, Nvidia, Xilinx...). Most objective observers would say that this means they have a big lead over Intel as far as having a usable mass production 7nm (10nm for Intel) process is concerned, as opposed to a paper product.

Intel's other problem is that they stuck with their traditional big expensive monolithic die approach for Xeon; the resulting low yields for ~700mm2 14nm 28-core die are no problem when you haven't got any competition and can sell big server CPUs for ~$10000 each, but are a huge problem when your competitor is using ~80mm2 7nm 8-core CPU chiplets with yields (and die costs) at smartphone levels.

Given that this is exactly the same policy that Nvidia have been following with GPUs, you have to wonder whether AMD can also pull off the same trick here. No doubt Nvidia will say that stitched-together chiplets can't possibly work for GPUs -- which is exactly what Intel said for CPUs before they were proved wrong...

 

[lefty]

Both are literally true. But both ignore the inconvenient fact that Intel's current 10nm process isn't manufacturable with acceptable yield and never will be in its current form, and that their launched 10nm CPUs are only there so Intel can tell market analysts that "10nm is in production" -- again literally true, but not in mass production in any meaningful way for any useful product.

The "reworked" Intel 10nm process has had to give up several of the features that gave it the much-vaunted higher density and performance (COAG, SAQP interconnect, maybe cobalt interconnect...) which means it's now lower density (and maybe lower performance?) than TSMC 7nm as well as (at least) a year later to market.

I reckon around end of 2016 Intel realised that Cannonlake 10nm was a dud and they started working on Icelake 10nm. This is why Icelake is delayed to 2019. Icelake 10nm is a different node to Cannonlake 10nm - the notable difference is that the M1 pitch is 40nm instead of 36nm. The 36nm pitch of Cannonlake was it's undoing, simply not achievable with decent yields without EUV.
Icelake will debute some time in 2019 - mainly smaller die SKUs like laptops. It wont clock as high as the current 14nm++ process, being the first iteration (similar to what happened with Broadwell)

 

[Daniel Nenni]

AMD stock price has pulled back along with almost all other semis. I think AMD is generally on a path to prosperity. They are unique in one respect, as the only x86 fabless vendor. But it's a frothy stock that seems to have pump and dump media. I think that's what Daniel is pointing his guns at, the long biased pumpers. Technical merits aside.

Yes, and personally I am glad to see AMD is back on track to compete with Intel. The problem I have had with AMD is that their marketing outpaces engineering (to put it nicely). This has been the case for many years and clearly it will continue under Lisa Su. Competing with Intel is no small matter and Intel management is not above marketing slight of hand but it was my hope that AMD would change courses and bring some technical integrity back to the microprocessor competition. Especially now that AMD uses TSMC where we it is much more difficult to play chip marketing games. Unfortunately, in the era of fake news, that does not seem probable or even possible for AMD.

 

[Daniel Nenni]

The "reworked" Intel 10nm process has had to give up several of the features that gave it the much-vaunted higher density and performance (COAG, SAQP interconnect, maybe cobalt interconnect...) which means it's now lower density (and maybe lower performance?) than TSMC 7nm as well as (at least) a year later to market.

True, and here is the funny part, there were stories about meetings inside Intel on whether or not to rename the "improved" Intel 10nm process 7nm. They stuck with 10nm probably for the reasons you stated above. It is certainly my hope Intel 7nm is much more successful than 10nm. It would be a serious shame to see Intel go fabless.

 

[IanD]

It would be a shame to see Intel go fabless, but this might be the only sustainable long-term position for them.

Each new TSMC fab (plus the huge cost of process development) is paid for by AMD, and Apple, and HiSilicon, and Xilinx, and Broadcom, and Qualcomm, and Mediatek, and eventually many others. Each new Intel fab is paid for by Intel only, now they've junked their custom foundry program. Given the relative revenue of Intel vs. the entire rest of the industry, it's difficult to see how they'll be able to continue justifying the investment, especially if they have to start competing with AMD on wafer prices for CPUs -- they no longer have the luxury of being pretty much a single source who didn't really care about wafer prices because of huge product gross margins.

Also the dominant fab volume driver (massive rapid volume ramp-up needed to drive yields up) is now smartphone APs, not desktop/laptop/server CPUs like it used to be. And having a more expensive/faster process compared to cheaper/slower foundry processes only works if you can get it better/faster by throwing money at it, which also doesn't seem to be working for Intel any more.

So all the factors that used to justify Intel having their own fab are either disappearing or have already vanished...

 

[Portland]

Intel fabless comment came out of no where. I wonder what Mr. Nenni is thinking.

This threat went full circle trashing amd and then hyping the company.

 

[simoncc2]

Intel fabless comment came out of no where. I wonder what Mr. Nenni is thinking. This threat went full circle trashing amd and then hyping the company.

the slide below can be easily updated. I don't think Intel will be forced to go fabless, but the execution of the 10nm, and especially the delay of the execution, reminds me of other unsuccessful stories like GF, UMC and SMIC , IBM ...

https://i.imgur.com/xjTkGKS.png

 

[raghu78]

It would be a shame to see Intel go fabless, but this might be the only sustainable long-term position for them.

Each new TSMC fab (plus the huge cost of process development) is paid for by AMD, and Apple, and HiSilicon, and Xilinx, and Broadcom, and Qualcomm, and Mediatek, and eventually many others. Each new Intel fab is paid for by Intel only, now they've junked their custom foundry program. Given the relative revenue of Intel vs. the entire rest of the industry, it's difficult to see how they'll be able to continue justifying the investment, especially if they have to start competing with AMD on wafer prices for CPUs -- they no longer have the luxury of being pretty much a single source who didn't really care about wafer prices because of huge product gross margins.

Also the dominant fab volume driver (massive rapid volume ramp-up needed to drive yields up) is now smartphone APs, not desktop/laptop/server CPUs like it used to be. And having a more expensive/faster process compared to cheaper/slower foundry processes only works if you can get it better/faster by throwing money at it, which also doesn't seem to be working for Intel any more.

So all the factors that used to justify Intel having their own fab are either disappearing or have already vanished...

Leading edge node development is now a economic problem first and then a technical problem. If a foundry or IDM does not have the volume they will have to stop sooner or later. GF is the latest in the long list of casualties over the past 2 decades. TSMC leading edge node cost is literally shared by the entire fabless semiconductor industry which includes Apple, Qualcomm, Nvidia, AMD, Broadcom, Mediatek, Huawei, Xilinx and others. TSMC enables every Intel competitor from AMD, Nvidia to Xilinx and even IBM (if the recent news of IBM choosing TSMC turn out true). Whats more important TSMC 7nm will enable all of these companies to gain share from Intel as their 10nm products are badly delayed. TSMC is powering ahead with N7+ with EUV in 2019 nd N5 in 2020. Intel 10nm server ICL will go up against TSMC N7+ based Milan. In 2021 TSMC 5nm based high performance products from AMD, Nvidia and Xilinx will make things even more difficult for Intel. TSMC is unlikely to let its process node lead slip away. I see TSMC gaining significant silicon share at Intel's expense in high performance computing. TSMC is going to fight Intel hard for every wafer sold at the leading edge. At some point there will be only one winner left as the economics get even more difficult at 3nm and below . My prediction is TSMC will be the one left standing after the dust has settled in this epic contest at the leading edge.

 

[Portland]

I think it will come down to architecture and software as long as the node size is competitive. Intel has to do twice the work design as well as make and amd can focus on design.

Semiwiki always focuses on the node size but never really the architecture, software or what the consumer wants.

 

[IanD]

I think it will come down to architecture and software as long as the node size is competitive. Intel has to do twice the work design as well as make and amd can focus on design.

Semiwiki always focuses on the node size but never really the architecture, software or what the consumer wants.

This statement ignores cost. If TSMC have several times the volume of Intel their cost per wafer will be lower because the process development cost is shared out over more wafers (and more customers). Their yield is also likely to be higher, because nothing helps get yield up more than pushing big volumes through a fab as early as possible in its life, and nowadays smartphone APs (which Intel doesn't have and has no prospect of getting) are the vehicle for this. Put these together and Intel/Altera will be unable to compete on price with AMD/Xilinx -- and cost *does* matter unless you have an overwhelming performance lead, which Intel used to have but have lost.

In the "old days" Intel had by far the highest wafer prices in the industry because their near-monopoly in high-margin products meant they could afford a "gold-plated" process, and it didn't matter if TSMC yields were higher and costs were lower because there were no competing products. Lots of people write as if having (by a small margin) the highest single-core performance for desktop CPUs for gamers is what drives CPU strategy nowadays, but it simply ain't so because this market is small -- for all the big-volume markets (which drive process development) cost and power efficiency and all-core speed are far more important than headline 5GHz turbo-core clock rates.

Also nobody except geeks cares about architecture, only results, and software will run just as well on non-Intel CPUs -- in AMDs case exactly the same software. What the consumer wants only matters if the demand is big enough, and CPUs for gaming are a small and shrinking market nowadays, the CPU priority order for the fabs is CPUs for smartphones (#1 priority by far), servers/HPC (#2 priority), laptops (#3 priority), volume desktops (#4 priority) -- gamers are way down the list. The fab customers (who sell products -- which includes Intel) may care more about gamers because they can sell CPUs to them for high prices so long as they can claim "5GHz!!!" (or whatever), but in the future these are not going to drive process and CPU developments.

The basic problem for Intel is that the markets and industry have moved on from the "golden old days" but their strategy hasn't changed to reflect this -- or at least, they haven't succeeded in changing it (e.g. getting into mobile). Nokia made exactly the same mistake, because of being king of the hill for so long (which had been a huge success for them) they stayed doing the same old stuff while the markets changed and walked away from them...

 

[Portland]

Good points but lets be honest most people would choose intel over amd.

My guess is that intel likes to modify their equipment and what they had didn't have the firepower and once they modify to have the firepower then they had problems with the product. TSMC and Intel both use ASML but one has denser chips with better performance. Intel could of had retention problems that slowed them down too.

 

[IanD]

Good points but lets be honest most people would choose intel over amd.

My guess is that intel likes to modify their equipment and what they had didn't have the firepower and once they modify to have the firepower then they had problems with the product. TSMC and Intel both use ASML but one has denser chips with better performance. Intel could of had retention problems that slowed them down too.

If by "most people" you mean "most PC gamers like me" this has been (and might still be) true, though even that is changing since Ryzen appeared.

But PC gamers are not "most people" as far as the foundries are concerned, they just don't buy enough silicon to keep the foundries happy, and it's the foundries who decide what processes to develop nowadays. The customers who want much bigger volumes care about price/density/efficiency, not "Intel Inside" -- if AMD can beat Intel on these measures, they'll get the business.

 

[hkwint]

In regards to density, I was speaking of process density not die density. The Intel 10nm process is denser than TSMC 7nm and that is a fact.

Says nothing. IMEC probably also has a process which is denser, as well as GloFo.

WHICH 10nm process are you talking about?

1) 1274? <-- The failed process of the 8121U that is an utter yield-failure and not going in volume production?
2) 1274.7?
3) 1274.12?

I just read today, 2) and 3) are the ones which will be going in mass production for Sunny Cove on Ice Lake, am I right?

So if the IMEC, GloFo and Intel 1) have better density than TSMC 7nm: That's what _I_ call desingeneous INTC-pumping and FAKE NEWS!!!! That never can be your intention, absolutely!


I also assume, both 2) and 3) will have "lower density" when it comes to MPP than does 1)?

Are 2) and 3) for Sunny Cove backpedaled to 40nm MPP without cobalt now? Anybody has 'shareable' info on this?

Ed: Do I understand correctly, originally, 36MPP was announced for 1), but now Scotten Jones refers to 44nm M2P? Or is that something else?
--> Intel 7nm on track

 

[Portland]

I'm not a gamer. For office work I would go with intel for now, it's more reliable.