Intel's 7nm is broken, company announces delay until 2022, 2023

[CarlBar]

4 hours ago, leadeater said:

Mobile CPUs actually sell the most by far, server is just high margin but much lower volume. If you control the laptop market then you control the majority of the entire PC market, if you also control the laptop market then you will control the desktop market as businesses stick to same CPU architecture brands and consumers will just buy what the popular thing is.

 

Nobody really cares about desktops, what's in them or any other such things. Just look at all the marketing and press that exists, product sames, talking points, near as much everything is laptops and mobile devices, unless you play games you have zero interest in desktops entirely.

 

Edit:

Per quarter unit shipments:

• Laptops: over 40 million
• Desktops: under/about 20 million
• Servers: under 4 million

 

I was under the impression that servers made up for fewer customers by the much larger number of CPU's per customer they shipped. That said thanks for clarifying, i couldn't find any hard data on it when i tried a google search to confirm.

 

Doesn't change the fact AMD can't begin to ship enough CPU's to go around. You got any useful data on that as well as i'm going off fairly old, (in PC terms,, it's about a couple of years IRL), info there as well and google was similarly unhelpful.

[leadeater]

1 hour ago, CarlBar said:

Doesn't change the fact AMD can't begin to ship enough CPU's to go around. You got any useful data on that as well as i'm going off fairly old, (in PC terms,, it's about a couple of years IRL), info there as well and google was similarly unhelpful.

Not really, revenue is easier to find but shipment volume is much harder. Personally I'd say AMD wouldn't represent more than 20% of volume for laptops or desktops and the number would be lower as you go back the years, 20% is probably rather high too. Either way Intel owns and operates a lot of fabs with a lot of output and they struggle to supply the ~200mil odd CPUs across the market they need to and it would be near as much impossible for AMD to supply 50% of that currently, TSMC would have to have enough 7nm fab capacity for that plus other customers and GloFo would need to be able to produce the IODs which would be the most likely to be able to out of the two (lack of demand and GloFo 12/14nm is actually exceedingly good).

[mr moose]

28 minutes ago, leadeater said:

 Either way Intel owns and operates a lot of fabs with a lot of output and they struggle to supply the ~200mil odd CPUs across the market they need to and it would be near as much impossible for AMD to supply 50% of that currently,

I wish people would pay more attention to small details like that.

[leadeater]

5 minutes ago, mr moose said:

I wish people would pay more attention to small details like that.

Not sure ability to produce and supply product is a small detail, but I know what you are saying 

[5x5]

12 hours ago, CarlBar said:

 

Guess what. The majority of the desktop money is being made on OEM's, not people selling individual CPU's. And Desktop is a tiny percentage of the market, the real money is in server space. It' also where the really high volumes are.

 

TSMC also just flat out cannot manufacture enough CPU's for AMD to meet global demand. A heck of a lot of people worldwide, especially in the server space are going to be forced to go inel for quite some time to come simply because AMD can't supply everyone.

Good news then, AMD are able to double their capacity with Huawei getting shafted if what you're saying was true

[5x5]

21 minutes ago, Jack_of_all_Trades said:

 

ah shit, never realized that TSMC is in a way a bottleneck if AMD actually catches up with Intel(or surpasses) in terms of sales.

AMD would then have to outbid other clients. But with Huawei out of the picture, that won't be an issue soon

[leadeater]

6 minutes ago, 5x5 said:

AMD would then have to outbid other clients. But with Huawei out of the picture, that won't be an issue soon

It's going to take more than just Huawei to be able to provide that kind of supply, TSMC has a lot of customers and produces so many other end products than just CPUs and GPUs.

 

Quote

Included in the report is a ranking of the 25 largest wafer capacity leaders in terms of monthly installed capacity in 200mm-equivalents as of December 2019. The world’s top-five wafer capacity leaders each had capacity of more than 1,000,000 wafer starts per month (Figure 1). Combined capacity of the top five companies represented 53% of total global wafer capacity at the end of 2019.  In contrast, the top five capacity leaders in 2009 held 36% of worldwide capacity.  Capacity at other semiconductor leaders, including Intel (817K wafers/month), UMC (753K wafers/month), GlobalFoundries, Texas Instruments, and STMicro, fell off rapidly from the top five.

 

 

 

TSMC global output is only 3 times Intel but like Intel and all of fab companies you need the capacity in the production node for the product. TSMC 7nm wafer capacity is about 1.1 million and that needs to supply all customers.

 

 

 

Quote

Apple Daily reports that TSMC’s 7nm wafer capacity will be 110,000 wafer starts per month (WPM) in the first half of 2020. TSMC’s Top 5 would consist of Apple, HiSilicon (who is starting to ship to external customers), Qualcomm, AMD and MediaTek.

 

Quote

In numbers, AMD would consume 21% share of TSMC’s 7nm wafer capacity with 30,000 WPM. HiSilicon and Qualcomm would be responsible for 17-18% share and MediaTek for 14%. This would leave 29% of TSMC’s 7nm capacity to the rest of its customers.

 

[xentropa]

I don't think servers is viable market for x86 AMD or Intel CPUs in the long run.  Servers and supercomputers will eventually shift to ASICs or dedicated architectures like the Google TPU.  

[cj09beira]

20 minutes ago, xentropa said:

I don't think servers is viable market for x86 AMD or Intel CPUs in the long run.  Servers and supercomputers will eventually shift to ASICs or dedicated architectures like the Google TPU.  

i dont see that happening at all, servers do a lot of different tasks, from website hosting to database management, to number crunching of various different kinds, to serving as a storage pool interface, to virtual machine hosting, there will always be a great market for general computing servers, and there x86 is more than good enough

[leadeater]

56 minutes ago, xentropa said:

I don't think servers is viable market for x86 AMD or Intel CPUs in the long run.  Servers and supercomputers will eventually shift to ASICs or dedicated architectures like the Google TPU.  

Unlikely, FPGAs and ASICs/GPUs etc have been around for a long time but they are either limited in utility and life span or have very poor cost benefit. In either case they need a CPU to function at all and connect to the system at which point because these are necessary it doesn't make sense to not as best utilize that resource as you can. For every ~4 GPUs there is a CPU and you can make similar statements about other ASICs and FPGAs. Plus this is almost exclusive to the supercomputer market which actually isn't that big overall, just cool to talk about.

 

Basically every business needs traditional server capacity be that purchased and operated on-prem or as a service from a provider and right now and as far as anyone can predict in to the future those are and need to be x86 based. Until companies can purchase software like finance and payroll software, asset management, customer and service relationship management for non x86 platforms nothing is ever going to change and these companies and many other software like it are historically slow to move on anything and extremely picky on exactly was is and is not supported. Our call center software vendor still insists on certain functions to be run on physical hardware not as virtual machines so they are nearly 2 decades behind what most would consider standard industry.

 

None of the above benefits at all from ASICs or FPGAs either, you don't need a 6s drag car to drive to work, which will overheat and die not moving in traffic. Cool and sexy generally isn't actually what is most popular, which is the Toyota Corolla and when is the last time anyone ever talked about those cars like we do about GPUs and CPUs.

[Kisai]

8 hours ago, CarlBar said:

 

I was under the impression that servers made up for fewer customers by the much larger number of CPU's per customer they shipped. That said thanks for clarifying, i couldn't find any hard data on it when i tried a google search to confirm.

 

Doesn't change the fact AMD can't begin to ship enough CPU's to go around. You got any useful data on that as well as i'm going off fairly old, (in PC terms,, it's about a couple of years IRL), info there as well and google was similarly unhelpful.

Dell and HP are the major OEM's, and if you look at their product range, which would have been engineers at least a year in advance if not two, Intel offerings outnumber AMD offerings, at least 4:1. If suddenly there was a rush to AMD parts, HP and Dell would have problems filling those orders just like retail offerings had problems filling orders for desktop Ryzen.

 

Intel deciding to use TSMC for some offerings, will likely end up being with their laptop parts first. Because you can't sell a 10nm laptop part while your competition is AMD selling a 5nm part and Windows/Apple deciding to use 5nm ARM parts as well. It just makes no sense to even engineer a laptop using that 10nm chip.

 

Intel should have:

a) focused on getting that smaller node up rather than sit on on 14nm for 3+ cycles. I t might have been fine to sit on it for the 6000 and 7000 parts, but then they sat on it for the 8000, 9000 10000 and are still planning on sitting on it for 11th gen parts. That's just insane.

Note there's still 14nm CPU's, laptop CPU's in 2021 listed. That just should not be a thing, nobody is going to want to buy those if the battery life of the laptop is half that of competition, and thus HP, Dell, Lenovo, Acer, Asus, MSI, Clevo, etc would unlikely use these parts in anything if AMD is offering Ryzen 4000 parts on 7nm THIS year and 5nm next year.

 

https://www.pcgamer.com/amd-zen-4-specific-5nm-enhanced-node/

Quote

In a news story from Chainnews (via @chiakokhua), ostensibly about the cut in 5nm and 7nm production orders from Huawei, the piece also notes that TSMC isn't worried about this drop because Apple has taken up the slack. It's asking for a whole bunch of extra chip wafers, potentially competing with AMD for 5nm demand at the end of this year.

 

But the piece also goes on to say that: "TSMC is said to have developed a 5nm enhanced version of its process specifically for AMD, which has a capacity requirement of no less than 20,000 12-inch wafers per month."

Keep in mind that's from a TW chinese-language source and english speakers tend to mis-interpret things.

 

b) Intel should have just jettisoned the current skylake/core brand after the 9th gen. There are enough bugs in these CPU's from poor validation on Intel's part that data centers would very much reconsider Intel Xeon CPU purchases if AMD CPU parts have less security-related bugs in them. Bring out another brand based on an entirely new design that gets rid of hyperthreading, uses big.little cores, whatever, just stop selling us on a brand that immediately brings up stories of unfixable, exploitable bugs.

 

https://www.zdnet.com/article/intel-csme-bug-is-worse-than-previously-thought/

Quote

Researchers say a full patch requires replacing hardware. Only the latest Intel 10th generation CPUs are not affected.

 

For end user desktops, and even laptops, bugs like these are not typically encountered, and even if it is, it's more likely to be abused by the owner of the device on purpose (eg to break DRM.) Hence Intel really should have just phased all the 14nm skylake-family parts out.

 

If this were any other business, there would have been recalls and refunds/lawsuits for this kind of behavior. 

[5x5]

22 hours ago, pas008 said:

-snip-

Man, chill. Intel are not selling out nor are they selling well. A shortage doesn't mean they suddenly can't produce 1/10th the amount AMD are selling, let alone producing. As we stand, Zen 2 is outselling Intel 11:1 across at least five markets. It's nowhere near equal in the others as we have data from many retailers confirmed people are looking at Ryzen more than Core i

Edited July 26, 2020 by colonel_mortis
Snip unconstructive posts that have been removed

[wall03]

On 7/24/2020 at 2:00 PM, straight_stewie said:

last node shrink anyone does

I bet you that AMD and Apple and Nvidia will push for lower and lower nodes. I expect to at least reach 1nm, while Intel doesn't ever go past 7nm without going bankrupt lol.

 

Intel be like
 

Spoiler

zuckerburg watching every step of Intel

 

 

 

corporate office be like

 

... wheres the money

 

 

[Helpful Tech Witch]

8 hours ago, leadeater said:

Either way Intel owns and operates a lot of fabs

Not for long!

[wall03]

6 hours ago, leadeater said:

In numbers, AMD would consume 21% share of TSMC’s 7nm wafer capacity with 30,000 WPM. HiSilicon and Qualcomm would be responsible for 17-18% share and MediaTek for 14%. This would leave 29% of TSMC’s 7nm capacity to the rest of its customers.

that means that Nvidia and Apple combined (I am leaving 8% for "other") are around equal to AMD.

 

wOaH I didn't know AMD was doing this well..

[Sauron]

surprised_pikachu.jpg

[Tech_Dreamer]

Quote

Intel's 7nm is broken..

You know what else is broken??

 

[leadeater]

4 hours ago, 5x5 said:

Man, chill. Intel are not selling out nor are they selling well. A shortage doesn't mean they suddenly can't produce 1/10th the amount AMD are selling, let alone producing. As we stand, Zen 2 is outselling Intel 11:1 across at least five markets. It's nowhere near equal in the others as we have data from many retailers confirmed people are looking at Ryzen more than Core i

All that data is from parts sales for DIY builders, that's only a very small percentage of the market. I wouldn't read in to that too much as a single OEM like Dell is shifting something like 100-1000 times the volume compared to people buying a CPU and motherboard and building their own thing.

 

For a good year Intel wasn't able to supply Xeons of the most popular models which delayed our orders through HPE, we had the same problem for our HP desktops with 9700 and 9900 non K CPUs them in them.

https://www.theregister.com/2020/01/20/intel_hpe_xeon_shortage/

 

Intel CPU shortages hasn't been a big problem for like the last 6 months and that was fixed mostly by Intel being able to make more, the problem was actually in not enough assembly fab output. They were able to make the chips fast enough but not package them in to actual CPUs.

[straight_stewie]

9 hours ago, wall03 said:

I bet you that AMD and Apple and Nvidia will push for lower and lower nodes. I expect to at least reach 1nm, while Intel doesn't ever go past 7nm without going bankrupt lol.

I don't think that you fully understand the physics at play here.

5nm is likely as small as it gets for electricity based devices.

This is not a business thing, it is a physics thing.

[hishnash]

52 minutes ago, straight_stewie said:

This is not a business thing, it is a physics thing.

They can go smaller still the issue is you need to increase the frequency of the lazer light you are using (aka reduce the wave length) doing so also however also makes it much harder to make these machines... they are already into the deep ultraviolet spectrum as some point soon they will be entering the x-ray section of the spectrum when this happens the cost of making the machines will go up massively making cor-herant x-ray lasers is not something we are good at yet.. 

[straight_stewie]

13 minutes ago, hishnash said:

They can go smaller still the issue is you need to increase the frequency of the lazer light you are using (aka reduce the wave length) doing so also however also makes it much harder to make these machines... they are already into the deep ultraviolet spectrum as some point soon they will be entering the x-ray section of the spectrum when this happens the cost of making the machines will go up massively making cor-herant x-ray lasers is not something we are good at yet..

Being able to have small enough lithography isn't the only concern here.

You also have to deal with quantum physics at that scale: You can make transistors as small as you want, but it's all meaningless unless they only change state when you want them to. You additionally have doping issues: The mask can be as small as you want, if you can't control the thickness of your dope layers as accurately as you can control the masking, you haven't actually gotten anywhere. Beyond that there are thermal issues: We are already struggling with heat density issues, getting smaller only makes that more difficult.

Another technology companies are looking into to try to get below 5nm is called Directed Self Assembly. But it still doesn't solve the quantum tunneling or heat density issues.

[5x5]

8 hours ago, leadeater said:

All that data is from parts sales for DIY builders, that's only a very small percentage of the market. I wouldn't read in to that too much as a single OEM like Dell is shifting something like 100-1000 times the volume compared to people buying a CPU and motherboard and building their own thing.

 

For a good year Intel wasn't able to supply Xeons of the most popular models which delayed our orders through HPE, we had the same problem for our HP desktops with 9700 and 9900 non K CPUs them in them.

https://www.theregister.com/2020/01/20/intel_hpe_xeon_shortage/

 

Intel CPU shortages hasn't been a big problem for like the last 6 months and that was fixed mostly by Intel being able to make more, the problem was actually in not enough assembly fab output. They were able to make the chips fast enough but not package them in to actual CPUs.

Interestingly enough I think it was Dell themselves who commented in an ama that interest in their intel machines dropped a LOT and they're barely able to meet Ryzen demand. That's why they added a lot more Zen based systems on offer and cut down a few intel based ones they had as they were in the red with them.

[leadeater]

3 hours ago, 5x5 said:

Interestingly enough I think it was Dell themselves who commented in an ama that interest in their intel machines dropped a LOT and they're barely able to meet Ryzen demand. That's why they added a lot more Zen based systems on offer and cut down a few intel based ones they had as they were in the red with them.

Yea I would expect that to be a thing now, takes a rather long time to get any traction in the OEM market from customers. A lot of that is just device life-cycles and also in part to the fact CPUs really not mattering that much at all, I could do my job equally effectively with either so I don't have a specific reason as an employee to ask for one or the other.

 

The current Ryzen mobile CPUs really will change the narrative as there is more tangible benefits on that platform like better battery life and significantly better performance enabling the ability to do things that were not possible or were prohibitive before. That's just how much better Ryzen mobile is now.

[DuckDodgers]

 

[straight_stewie]

On 7/26/2020 at 12:35 AM, gabrielcarvfer said:

TFETS are already a thing, take advantage of quantum tunneling to work, and should be able to be scaled down a bit more.

All true. What you left out is that, in 10 years of active development, nobody has figured out how to successfully package them into even VLSI chips. In plain English, after ten years of concerted development effort, TFETs are still not even remotely close to being production ready in large scale processors.

Still, barring that, it's unlikely that even in the best case TFETs would get down to 1nM. 1nM is roughly as long as two silicon atoms placed as closely together as possible. There are other substrates and dopants, but they aren't really that much better in this regard. For reference, 1nM is only roughly 10 hydrogen atoms placed as closely together as possible.

Geometrically speaking, each transistor would need either 3x3x3 or 3x2x2 atoms each, depending on the configuration.

 

Whatever we like to think, we are certainly very near the end of the road in terms of the minimum possible transistor size.

 

On 7/26/2020 at 12:48 AM, gabrielcarvfer said:

Some working on nanotubes to exchange heat using liquids (don't remember if vapor-chamber like or not, but I believe high pressure inside layered glass isn't a bright idea).

I read recently about some experiments involving circulating a cooling fluid through the die itself. I forget whether it was effective or not, but certainly interesting.