I used to be a yield engineer for a semiconductor fab. It was not uncommon to destroy a "boat" of 25 wafers from a drifting process or human error. That would be a loss of $1m in a flash. As a yield engineer the stress was insane. You'd tweak a parameter expecting it to improve yield and instead it would loose 5% (where breakeven point is 5%) and you wouldn't know until 1000s of wafers. The most underpaid job in the world. Nobody pats you on the back when you make the yield go up, but if yield drops due to an equipment issue you had nothing to do with your ass is grass. And everyone hates you because all you do is beat up people for hurting yield. I did it for 8 years. 8 years too many.
@@davidrave563 yes they do, there are test structures in between each die that they regularly test, but they don't catch everything ... And we don't know what those limits should be anyway until the product matures ... It's a lot of educated guesswork. Many test structures you can't test until a series of 5 steps (out of 100 total) have completed, and if the error is on the first step, you could process 20 boats of wafers until the deviation is detected.
It's not just manufacturing cost, there's research and development costs also. Through the process, there's developing, making and testing a new technology, like the 7 nm size transistors, then there's cpu design(circuit, logic), prototypes and testing. You may go through several iterations.
You never get true 100% yields. This is why Intel has bins. Same chips post process locked to lower clock speeds or features disabled due to defects. So maybe a small handful of chips from the wafer are full featured at full clock rate. Many full featured at lower ranges of clock rates. Some at low clock rates with se features disabled And a handful of duds. There will also be some area of the wafer that is unusable due to the chips being square and the wafer being round.
The raw material costs are much less than half the base cost of production. When I worked in auto manufacturing supplier sector, we made a plastic part for $8 and less than $3 was raw materials, sold it to customer for $13 which they put on cars but sold the OEM replacement for over $80, this is common in all manufacturing. I guess I should also add the customer paid for 2 sensors, the R&D and the quarter million dollar mold for manufacturing which is why they had such a mark up.
The best thing is when the stuff comes from the same company as the cheaper stuff and they are exactly the same aside from labeling. My father worked in the 90s for a Company that sold everything that had to do with Carbrakes (or something like that) and sometimes they got the stuff that was supposed to be shipped to Mercedes but was cause of an error send to them,
@@mrn234 in my city,we have a warehouse from the top ball bearings manufacturer. almost in front,there is a Caterpillar warehouse dealer. you can buy the caterpillar branded bag original part ball bearing for 120, or you can go to the ball bearing actual manufacturer in front and buy the same for 30. same ball bearing, same stamped code,different package, huge markup "for USA made quality"
@@xiro6 You can also buy most of the stuff on Amazon directly from the Chinese manufacturers via AliExpress at 50% off... Amazon often do not even bother importing and stocking the goods themselves - they just pass your order on to the Chinese manufacturer and you get the product directly from China, in the usual packaging as when you order from AliExpress. And you can't get it anywhere else, because no one outside of China is still manufacturing certain parts.
@@captainobvious-CH Can you tell me which suppliers in China would have some very low-priced gamer laptops? i need one so i can study, to run some heavy applications.. and as my national currency is tremendously devalued, it is extremely expensive and difficult to buy the full price.
Yeah. The cerebras chip as an example has probably very low number of units sold, so you have to spread the cost of not only R&D but also tooling (like making the mask) over small number of sold units. Guess what has an even higher profit margin, Apps or movies or games or music sold digitally.
@@noergelstein I'm sure that's true however with Apps and games you have a huge upfront cost and very high risk. Chances are with a CPU or GPU you can count on a certain amount of sales however if a game flops you have invested millions in creating that game and the losses can be staggering.
@@joer8854 Still doesn't follow though; IF R&D comes up with a good chip then yeah it'll move, but that part can bomb just the same as a software project. They're also gambling that the market they were planning to sell to when they started developing the chip still exists and hasn't been upended by a competitor or rendered obsolete by a different technology.
@@FireStormOOO_ So what gpu sold less than 100 pieces. There are tons of games that sold less than 100 copies that the developer invested massive amounts of money. You can sell 0 copies or millions. It's gambling with having only a very vague idea of the odds. Entire studios live and die on a single game. If Intel never sells a single discrete GPU they will still continue to exist.
The construction of a single fab can cost billions depending on the wafer size and process. Bosch - 300mm wafer - 130nm process - 2019 ≈ 1bn Intel Fab 28 - 300mm wafer - 10nm process - 2018 ≈ 5bn TSMC Fab 18 - 300mm wafer - 7nm process - 2018 ≈ 17bn
However the fabs wouldn't be built if there was no profit in it!! With the kinds of money being involved you now know why governments sponsor and protect these endeavours. These are economies in their own rights!!
I just hope he moves on the S-tier content so we don't get stuck with A++++++++++++++++ content for years. Sorry, I know that horse is dead enough already but couldn't help it
I just found this channel and how the hell does it not have more subscribers. This is top notch silicon level reporting compared to most other "techtubers"
Yes, they have big „margins“, if you don’t calculate in research and planning. It’s not only staff costs but also experimenting and that sums up drastically making not too extreme margins especially when you also realize not all projects make it.
@@TechTechPotato baked are crispiest. They're also ground down dehydrated potato mash, so they dry you out and they take a lot of potato to make. Potatoes are mostly water, after all, and they'll be trying to return to being mostly water.
Its like determining Van Goghs paintings by the material cost. The R&D, logistics and marketing expenses will drive down the profit per processor way more than the material and production.
Omg thanks for this. I can't believe this is Anandtech. I've been an avid reader of your reviews in your wesbite since 2000. Way back to the Pentium 3/AMD Duron/Thunderbird era. Thanks.
Silicon substrates are just one raw material, there are lots of additional high purity chemicals needed. These and all the other additional costs add up to a lot, and take that 20x ratio down to 0.6. If TSMC had $50B revenue and made 12M wafers in a year, average price is $4k.
@@benjaminfacouchere2395 I mixed what I was talking about a little there. One was that the raw material cost for a processed wafer isn't just the silicon. That was TSMC's cost. The other idea was how does a $6k produced wafer that can be turned into $100k of retail parts lead to only 50% profit overall for AMD or Intel.
Intel does billions of dollars of business with TSMC. 5% of TSMC's total revenue comes from intel. Used to be 8%. Intel has been a tsmc partner for over 2 decades.
As Dr Sophie Wilson mentions here, the problem is partly dark silicon that needs to be left blank so the chips don't get too toasty, this ties back to your later video about shapes. If the chips could be a different shape some of these concerns could be alleviated. One of the apparent solutions would be stretching to the reticle limit and seeing how wide they can go before it melts itself on the highest available density.
In the 1990’s I was a R&D technologist and worked on a 5 year creative research and development agreement called a “CRADA” between Intel and the Department of Energy (DOE) for the development of EUV technology to expose integrated circuit wafers shrinking the light wavelength and increasing wafer yields. A multifaceted approach included the EUV source, lasers and focusing mirrors, ultra clean vacuum systems for coating (sputtering) to reduce wafer defects and the wafer handling between process steps. A new interferometer technology was developed by Dr. Gary Sommargren to measure multilayered mirror shapes prior to coating that ultimately made the EUV project successful, not diminishing the excellent work of the entire team of Physicists, mechanical and electronic engineers and their support by talented technologists. We helped change the world just a little. The was done at the Lawrence Livermore National Laboratory. It’s now the Lawrence Livermore National Security.
Kudos for featuring Sophie Wilson who, I believe, wrote the BBC BASIC interpreter and then, for an encore, designed the original ARM instruction set. And wasn't it Clive Sinclair who once said that wafer scale integrated circuits were the future?
An interesting deep dive would be ASML, the most important tech company you never heared of. They actually make the machines that make the chips. The high tech stuff they do is out of this world.
This is fantastic information. The high markups is what's been driving the entire industry. Us consumers, we don't really care all that much - either as individual or professional clients, we just want better, faster, more economic processors and if paying a bit more will accelerate the development, that's fine.
The most important part of this: There are plenty of budgets that could be cut in a crisis (marketing etc), cause the actual production itself is only 5-10% of the cost. Hence silicon is here to stay, it's more then profitable enough.
@9:47 Price did increase due to material cost went up. Hwei basically buy all that it can before the sanction. But I think when the price went up by 10-20% (shipping cost etc), they didn't increase the price for their stable/long-term consumer to maintain the relationship. New orders from new consumer however weren't so lucky to need to wait in line and still pay the full cost
Fascinating stuff, and yeah I would love to see a video on the whole silicon ingot stuff etc, the fact that chips are ultimately custom made crystals is almost alien-like in tech as much as a kind of beautiful.
Would certainly like to see more videos on the whole end-to-end semiconductor development and manufacturing process. One thing that has intrigued me is how far ahead are Intel and AMD working? When they release a Zen 3 to the public do they already have Zen 4's up and running in the labs? It feels that when you are interviewing senior industry figures, such as Forrest Norrod, they probably have a lot more information in their heads as it relates to future products that they have to be careful to not divulge.
Absolutely, AMD has multiple teams working on new processors so that each team is (roughly) finished a year or two after one another Remember it takes YEARS before a CPU is ready for production, Zen 3 processor design probably started around Zen 1 launch
So according to Sophie Wilson we haven't been on Moore's original law (cost per transistor halving every 2 years) since before 2003. I wonder how long it actually remained true. Of course, the more modern version of Moore's Law is about how many transistors it makes sense to put on a chip - and that follows Moore's Law somewhat accurately. I wonder how large Cererbras TAM is. There can't be a large number of potential users, but I as an outsider probably underestimates how many organisations that could use their systems.
@@PWingert1966 I doubt that. The key thing with Gallium Nitride seems to be increased energy efficiency to enable higher performance. I haven't heard anyone pointing at it as a way to reduce the costs of the chip.
There are a lot more costs than the raw wafer and package! Fab depreciation is huge. Yield can’t be assumed to be so high (esp with larger die). Other consumables include gas, slurry, sputter targets, power, water, etc. and don’t forget to amortize the mask set cost too.
Be aware it is common to quote wafer prices in 200mm wafer equivalents. Ie the actual 300mm wafer price is about 2.25x the 200mmequivalent price (add some edge effect correction to that)
The size of the die makes yield trouble go up superlinearly. If there are three "bad spots" on the wafer, at 600 chips per wafer, that's 0.5%, but at 60 chips per wafer, that's 5%. Then again, tooling, design, and testing, and the risk taken on by starting that process, is a significant cost. As is the depreciation and interest on those 100 million dollar steppers lithography machines...
15:40 A factor in cost not listed here that's very important is testing. That includes quality, functionality and performance testing to make sure customers receive a part that works properly and as advertised. As well as binning to sort assembled packages according to their performance to begin with. This testing comprises a few major steps, one of which is what I do for a living.
@@hansdietrich83 You mean a cellphone? Jk, not hating on you vinyl lovers, I understand the reasons for it. Just commenting about how amazing it is that we can store that amount of music in our pocket nowadays.
For small runs you can't neglect the tooling costs for masks. More complex designs also need more masks per wafer and more layers and steps in the fabrication.
This video gives me a much better understanding of Fabs and wafer production, but I'm still blown away by the price. I thought I would be much more expensive!! Also i think the most important factor for the price is the supply/demand mechanism Thanks
Yeah, as you said the cost for a fab is huge. The lifetime of a node is limited, so the fabs need constant upgrading. The chip designs are also extremely expensive, and it takes a large amount of experts around 5 years to design a gpu or cpu
Nice, looking forward to reading up on Cerebras. Thinking of buying stocks when they finally do an IPO. I remember last year being hyped on AMD and TSMC and thinking why were their stock so low. Now they boomed but so did a lot of other tech plays. AI and semiconductor companies are the future.
My job is literally to manage and reduce wafer cost for one of the big semi companies. Interesting video to see what info you have and what you don't ;)
you also need to factor in building & equipment depreciation impact on the wafer cost. You could have one company with different Fabs with different depreciation profiles (time delays in launch between fabs) on the same node with completely different wafer costs at the same volumes. Also geo impact of freight on raw materials, labour costs, layout with different fabs can be significant deltas
Thank you, now I understand where prices going - to the Moon) just because Intel and AMD can say "Deficit" and rise margin to x25 per wafer) And special thanks to Steve from Gamers Nexus for posting a link to your video. Definitely cost to subscribe)
The breakdown of costs of materials is all well and good but there is that one huge elephant in the room... R&D, cost of maintaining equipment and so on... These costs account for most of the actual spending for a certain company and not the materials themselves
Very interesting. It would have been great to mention an estimate of some of the fixed costs, particularly the cost of mask sets. My understanding is that these have become really expensive, and amortization of their costs would likely make up a significant part of the cost of lower volume chips, like the Cerbras offering.
Graet video I always thought one silicon cost from a wafer will sell like the whole wafer For example if company sell x or y cpu that cost x or y it means the whole wafer cost x or y 9900k sold for 500$ bucks that means whole wafer of 9900k worth 500$
you need profits to pay for r&d, pay for salaries, pay for business strategies, trial and error, some cash on hand to survive bad times, money for legal litigations, insurance, etc etc. usually the min. is 30-40% but the ideal is 80% for a growth company as the risk is extremely high.
"What's your minimum specification?" Not entirely sure how to answer this, but I'm currently on an AMD 5800XT with 32GB (4x8GB) DDR 3600 CL16, and an AMD 6900XT.
Don't forget to factor in ASP (Average Selling Price) of the items. Not every processor die is sold as the top SKU, so not every die makes the same money.
Great presentation. One thing you failed to mention is the NRE costs of going to wafer production. Reticules are damn expensive. This is a massive expense when you need 50+ reticules. Please discuss this too. Maybe do a video on the whole reticule design and development process ?
Ian, would you be able to elaborate on the image at 1:32? What is the reason behind the fact that they can't flip some of those yellow chips that are being printed vertically (specifically the ones on the 3 and 9 o'clock) which would make them viable healthy chips?
There's a small ring around the processor that's called the 'edge loss'. You can't create processors right at the edge, and the edge loss is different for different factories/nodes and such. The yellow die would be green if the edge loss was lower, that's really what it means. The calculator already moves the mask/die location to maximize the number of green dies. There's a link to the site in the description if you want to play around with the tool.
@@TechTechPotato Thank you for the explanation, but I've heard you explain that in a previous video. That wasn't my question. I was specifically asking why those 3 and 9 oclock chips can't be flipped from horizontal to vertical to fit them inside the edge dead zone. Something to do with manufacturing?
ua-cam.com/video/skUCP2f4HIM/v-deo.html Presumably that would require rotating the mask or wafer just to add a few more dies? I'd assume that'd add non-trivial complexity to the lithography machine, and it would only be something that could be leveraged by a subset of products using very rectangular dies.
@@Hostilenemy As well as the lithography problems hughJ mentioned, it would also introduce more complexity when cutting the wafer into individual dies, where it might not be possible (or impractical) to cut and hold an irregular arrangement like that.
It's worth mentioning that the Cerebras chip isn't the most expensive component in the box, by at least an order of magnitude (I don't think the amounts have been publicly released).
Does the cost of wafers go down over time? I think i read its $10000 a wafer for 7nm tsmc, google says $9,346, do you know if prices over time go down as the tech gets older.
its important to understand that price is not related to cost, but to perceived value to the customer. People are happy to pay a lot for chips because of the improvement they offer to their lives and businesses. It is also important but also very difficult to recognise that a major part of the cost of the manufacturing process is the research, development, manufacture and depreciation of the machines that make the machines.
They need to make a huge amount of profit in order to continue R & D and to make sure they don't go broke if and when the orders stop for a period of time. There's also the cost of running everything else. Sure, they take advantage of their monopoly, but would you really prefer the possibility of them going out of business? Or for one of the other companies like Apple, Samsung, Microsoft, Intel, etc, to own them?
All well and good talking about manufacturing costs, but how much is spent on r&d, management of it, publicity, marketing, and supporting roles, etc..that costs an ARM AMD an inteLEG too.
How many 8086s could fit on a 7nm 12 inch platter? That said I'm a fan of the 186, 386 486 686 and Pentium 4. 386 was the last processor without intentional hardware backdoors.
Very interesting video about chip manufacturing! 04:44 - Millions of wafers from a single ingot? That's one big ingot that stretches from here to the ISS :)
I wonder if the rise of cost per gate is because they had to go from planar to 3d gates and also wonder when the switch to GAAFET happens the price will receive an ulterior unusual spike. Anyway great video, thanks to it now i can watch and learn more stuff from Dr. Sophie Wilson's presentations.
Great content as always, one question: Dr Su always points out that margin on console-SoCs increase over the cycle, does this factor in a decrease of wafer costs and if so can AMD expect the cost of TSMC 7nm wafers to decrease any time soon?
At the beginning of the cycle all the CapEx is done, which doesn't factor into future production. But usually over a processor cycle, between AMD and TSMC, there are additional efficiencies that can be made. A slightly better process step that gives better yield, or perhaps a tweak that saves 50 millivolts. Or in the past, a redesign for a smaller process node that ends up being cheaper.
The cost of TSMC 7nm wafers obviously isn't going to decrease any time soon when they can't get close to meeting demand at current prices and production is hobbled by ongoing water shortages.
As a non-native English-speaker, I had no idea, what an "Ingot" was. But today I learned, that "I" "stretch" "silicon seed" "into" an ingot, and that "the size" of "my" "ingot" depends on "the speed with which "i pull it out"..? This one(!) sentence will now haunt me for the rest of my life. Engineering can come in huge packages and is hard most of the time, I guess.I had no idea how much I did wrong in this regard. I did, however, know what a "Fap" was.
make a video on how cerebras gets around defects, that would be interesting, are they just including in chip redundancy or is this something more sophisicated?
Nothing beats Haddock and chips, deep fried in beef dripping. Now given my age, its chips deep fried in rapeseed oil and cod given where i live Haddock is just not available.
Could you do one on how assembling chiplets works? Such as the different methods and the future methods they are looking at? What I guess without any knowledge is they use a masking technique and they just mask two chips together, but it is a guess.
super interesting thanks dr potato
BRUH WHAT THE FUCK NO WAY 🤣🤣🤣🤣😂
Here to serve! 📶
it's the evans boi
@@TechTechPotato serve on a silver wafer? :)
Yoooo
I used to be a yield engineer for a semiconductor fab. It was not uncommon to destroy a "boat" of 25 wafers from a drifting process or human error. That would be a loss of $1m in a flash. As a yield engineer the stress was insane. You'd tweak a parameter expecting it to improve yield and instead it would loose 5% (where breakeven point is 5%) and you wouldn't know until 1000s of wafers. The most underpaid job in the world. Nobody pats you on the back when you make the yield go up, but if yield drops due to an equipment issue you had nothing to do with your ass is grass. And everyone hates you because all you do is beat up people for hurting yield. I did it for 8 years. 8 years too many.
That sounds nuts, how can you go 1000s of failed wafers before noticing something was wrong, don't they test each wafer?
@@davidrave563 yes they do, there are test structures in between each die that they regularly test, but they don't catch everything ... And we don't know what those limits should be anyway until the product matures ... It's a lot of educated guesswork. Many test structures you can't test until a series of 5 steps (out of 100 total) have completed, and if the error is on the first step, you could process 20 boats of wafers until the deviation is detected.
It's not just manufacturing cost, there's research and development costs also. Through the process, there's developing, making and testing a new technology, like the 7 nm size transistors, then there's cpu design(circuit, logic), prototypes and testing. You may go through several iterations.
Also, verification of the design... .i.e Design verification..
You never get true 100% yields. This is why Intel has bins. Same chips post process locked to lower clock speeds or features disabled due to defects. So maybe a small handful of chips from the wafer are full featured at full clock rate. Many full featured at lower ranges of clock rates. Some at low clock rates with se features disabled And a handful of duds. There will also be some area of the wafer that is unusable due to the chips being square and the wafer being round.
he also forgot about mask costs, which are insane with these deeply scaled tech nodes.
And day to day operation costs.
Taken everything into the price they still markup too much,,,ppl are ready to buy and prices will be higher when this crisis is over.
The raw material costs are much less than half the base cost of production. When I worked in auto manufacturing supplier sector, we made a plastic part for $8 and less than $3 was raw materials, sold it to customer for $13 which they put on cars but sold the OEM replacement for over $80, this is common in all manufacturing. I guess I should also add the customer paid for 2 sensors, the R&D and the quarter million dollar mold for manufacturing which is why they had such a mark up.
The best thing is when the stuff comes from the same company as the cheaper stuff and they are exactly the same aside from labeling. My father worked in the 90s for a Company that sold everything that had to do with Carbrakes (or something like that) and sometimes they got the stuff that was supposed to be shipped to Mercedes but was cause of an error send to them,
@@mrn234 in my city,we have a warehouse from the top ball bearings manufacturer.
almost in front,there is a Caterpillar warehouse dealer.
you can buy the caterpillar branded bag original part ball bearing for 120, or you can go to the ball bearing actual manufacturer in front and buy the same for 30.
same ball bearing, same stamped code,different package, huge markup "for USA made quality"
@@xiro6 You can also buy most of the stuff on Amazon directly from the Chinese manufacturers via AliExpress at 50% off... Amazon often do not even bother importing and stocking the goods themselves - they just pass your order on to the Chinese manufacturer and you get the product directly from China, in the usual packaging as when you order from AliExpress. And you can't get it anywhere else, because no one outside of China is still manufacturing certain parts.
@@captainobvious-CH Can you tell me which suppliers in China would have some very low-priced gamer laptops? i need one so i can study, to run some heavy applications.. and as my national currency is tremendously devalued, it is extremely expensive and difficult to buy the full price.
R&D: I'm bout to end this profit margins whole career!
Yeah. The cerebras chip as an example has probably very low number of units sold, so you have to spread the cost of not only R&D but also tooling (like making the mask) over small number of sold units.
Guess what has an even higher profit margin, Apps or movies or games or music sold digitally.
@@noergelstein I'm sure that's true however with Apps and games you have a huge upfront cost and very high risk. Chances are with a CPU or GPU you can count on a certain amount of sales however if a game flops you have invested millions in creating that game and the losses can be staggering.
Yes thats the point.
@@joer8854 Still doesn't follow though; IF R&D comes up with a good chip then yeah it'll move, but that part can bomb just the same as a software project. They're also gambling that the market they were planning to sell to when they started developing the chip still exists and hasn't been upended by a competitor or rendered obsolete by a different technology.
@@FireStormOOO_ So what gpu sold less than 100 pieces. There are tons of games that sold less than 100 copies that the developer invested massive amounts of money. You can sell 0 copies or millions. It's gambling with having only a very vague idea of the odds. Entire studios live and die on a single game. If Intel never sells a single discrete GPU they will still continue to exist.
I've wanted more info like this from the youtube tech community for years. Thanks for putting us on to Sophie Wilson.
The construction of a single fab can cost billions depending on the wafer size and process.
Bosch - 300mm wafer - 130nm process - 2019 ≈ 1bn
Intel Fab 28 - 300mm wafer - 10nm process - 2018 ≈ 5bn
TSMC Fab 18 - 300mm wafer - 7nm process - 2018 ≈ 17bn
However the fabs wouldn't be built if there was no profit in it!! With the kinds of money being involved you now know why governments sponsor and protect these endeavours. These are economies in their own rights!!
Something is wrong here.
Intel's 10nm is BETTER/SMALLER than the TSMC 7nm.
So the fab price is wrong.
@@Wirmish The prices are taken from official press releases. Who cares what's better ... 🤦♂️
@@Wirmish no tsmc 7nm is smaller then 10nm Intel, the difference between them is not extremely high that's all
@@bertjedekat Well the TSMC 7nm node actually works.
Ian is on a roll. The last few videos have been A+++ content. Thanks Ian, love this channel, keep up the fantastic work!
yea, thats the shit.
I just hope he moves on the S-tier content so we don't get stuck with A++++++++++++++++ content for years. Sorry, I know that horse is dead enough already but couldn't help it
I just found this channel and how the hell does it not have more subscribers. This is top notch silicon level reporting compared to most other "techtubers"
It's a relatively new channel, all good growth comes slowly.
Yes, they have big „margins“, if you don’t calculate in research and planning. It’s not only staff costs but also experimenting and that sums up drastically making not too extreme margins especially when you also realize not all projects make it.
Also the fact that yields are
not to mention constant investing to new products and fabs.... without it, company and business dies in this industry
As well as a fab costing 15bn USD ish... Not exactly a cheap industry to get into :P
This is one of the most organically growing channel
Not really its Organosilicon biochemistry channel. 🙃
And? Strange comment
48.6K subscribers and still growing .
Depends on the cost of potatoes, oil and spices I would imagine 🤔
😂😂😂
Baked or fried?
@@TechTechPotato Lightly fried tastes best to me 🙂
@@TechTechPotato baked are crispiest. They're also ground down dehydrated potato mash, so they dry you out and they take a lot of potato to make. Potatoes are mostly water, after all, and they'll be trying to return to being mostly water.
@@TechTechPotato ua-cam.com/video/mLruhZjq-QY/v-deo.html
It's always amazing to hear from a true expert. Thanks for the video Dr Potato Man!
Its like determining Van Goghs paintings by the material cost. The R&D, logistics and marketing expenses will drive down the profit per processor way more than the material and production.
without even watching the entire video , I subscribed , and this is my first time on this channel
Your channel will definitely grow!!! Very interesting insight into the cost aspect of CPU/GPU manufacturing.
Omg thanks for this. I can't believe this is Anandtech. I've been an avid reader of your reviews in your wesbite since 2000. Way back to the Pentium 3/AMD Duron/Thunderbird era. Thanks.
What do you mean “this is anandtech”? Anandtech was started/ran by Anand Shimpi
@@codycast i know who STARTED it. I was meaning related to Anandtech.
Silicon substrates are just one raw material, there are lots of additional high purity chemicals needed.
These and all the other additional costs add up to a lot, and take that 20x ratio down to 0.6.
If TSMC had $50B revenue and made 12M wafers in a year, average price is $4k.
But isn't that already included in the wafer cost @ 12:50 ?
@@benjaminfacouchere2395 I mixed what I was talking about a little there. One was that the raw material cost for a processed wafer isn't just the silicon. That was TSMC's cost. The other idea was how does a $6k produced wafer that can be turned into $100k of retail parts lead to only 50% profit overall for AMD or Intel.
@@davidgunther8428 I see. My understanding was that the $6k/wafer are not the purchase price for AMD/Intel but the production cost for TSMC.
Would like to clear up Intel doesn't really do business much with TSMC. They have their own facilities. So their cost may be completely different.
Intel does billions of dollars of business with TSMC. 5% of TSMC's total revenue comes from intel. Used to be 8%. Intel has been a tsmc partner for over 2 decades.
As Dr Sophie Wilson mentions here, the problem is partly dark silicon that needs to be left blank so the chips don't get too toasty, this ties back to your later video about shapes.
If the chips could be a different shape some of these concerns could be alleviated. One of the apparent solutions would be stretching to the reticle limit and seeing how wide they can go before it melts itself on the highest available density.
Why i discovered this channel only today ? You are so underrated omg
In the 1990’s I was a R&D technologist and worked on a 5 year creative research and development agreement called a “CRADA” between Intel and the Department of Energy (DOE) for the development of EUV technology to expose integrated circuit wafers shrinking the light wavelength and increasing wafer yields. A multifaceted approach included the EUV source, lasers and focusing mirrors, ultra clean vacuum systems for coating (sputtering) to reduce wafer defects and the wafer handling between process steps. A new interferometer technology was developed by Dr. Gary Sommargren to measure multilayered mirror shapes prior to coating that ultimately made the EUV project successful, not diminishing the excellent work of the entire team of Physicists, mechanical and electronic engineers and their support by talented technologists. We helped change the world just a little. The was done at the Lawrence Livermore National Laboratory. It’s now the Lawrence Livermore National Security.
Kudos for featuring Sophie Wilson who, I believe, wrote the BBC BASIC interpreter and then, for an encore, designed the original ARM instruction set. And wasn't it Clive Sinclair who once said that wafer scale integrated circuits were the future?
At some point I'd love to get her on for an interview
I worked with her, very smart!
An interesting deep dive would be ASML, the most important tech company you never heared of.
They actually make the machines that make the chips. The high tech stuff they do is out of this world.
This is fantastic information. The high markups is what's been driving the entire industry. Us consumers, we don't really care all that much - either as individual or professional clients, we just want better, faster, more economic processors and if paying a bit more will accelerate the development, that's fine.
Probably the most in-depth video I've seen so far on wafer manufacturing . Great video!
The most important part of this:
There are plenty of budgets that could be cut in a crisis (marketing etc), cause the actual production itself is only 5-10% of the cost.
Hence silicon is here to stay, it's more then profitable enough.
Awesome Commodore T-shirt….memories from my first Amiga in 1989 :)
Impressed that you picked up on her presentation. I had seen it a while back too and thought the same
@9:47 Price did increase due to material cost went up. Hwei basically buy all that it can before the sanction.
But I think when the price went up by 10-20% (shipping cost etc), they didn't increase the price for their stable/long-term consumer to maintain the relationship.
New orders from new consumer however weren't so lucky to need to wait in line and still pay the full cost
Fascinating stuff, and yeah I would love to see a video on the whole silicon ingot stuff etc, the fact that chips are ultimately custom made crystals is almost alien-like in tech as much as a kind of beautiful.
Would certainly like to see more videos on the whole end-to-end semiconductor development and manufacturing process. One thing that has intrigued me is how far ahead are Intel and AMD working? When they release a Zen 3 to the public do they already have Zen 4's up and running in the labs? It feels that when you are interviewing senior industry figures, such as Forrest Norrod, they probably have a lot more information in their heads as it relates to future products that they have to be careful to not divulge.
Absolutely, AMD has multiple teams working on new processors so that each team is (roughly) finished a year or two after one another
Remember it takes YEARS before a CPU is ready for production, Zen 3 processor design probably started around Zen 1 launch
impressive how much meaningful informations you manage to bring up in your videos. Great job.
Great insight. Always learn a thing or two every time you put out content.
Excellent, informative video. Subscribed
You have answered a lot of questions I had now I have a lot more
Learning in a nutshell! The more you know, the more you realise you don't know
So according to Sophie Wilson we haven't been on Moore's original law (cost per transistor halving every 2 years) since before 2003.
I wonder how long it actually remained true.
Of course, the more modern version of Moore's Law is about how many transistors it makes sense to put on a chip - and that follows Moore's Law somewhat accurately.
I wonder how large Cererbras TAM is. There can't be a large number of potential users, but I as an outsider probably underestimates how many organisations that could use their systems.
Cerebras says they have dozens of customers, and they need to grow the company to help support the number of customers after the hardware.
I wonder if we will get back on that law when we go to non-silicon like Gallium Nitride?
@@PWingert1966 I doubt that. The key thing with Gallium Nitride seems to be increased energy efficiency to enable higher performance. I haven't heard anyone pointing at it as a way to reduce the costs of the chip.
Cool video, your intro always makes me happy! btw we need more CHIPS for our minimum specifications!
Please don't ever change that awesome intro graphic! :)
Man, I wish I could hire the artist to do an intro for me. :)
Wow! Another stellar video about a subject I did not know I wanted to know more about! Thanks a lot!
There are a lot more costs than the raw wafer and package! Fab depreciation is huge. Yield can’t be assumed to be so high (esp with larger die). Other consumables include gas, slurry, sputter targets, power, water, etc. and don’t forget to amortize the mask set cost too.
Yes of course. This is a simplified video based on a single set of numbers that aren't easily searchable.
Be aware it is common to quote wafer prices in 200mm wafer equivalents. Ie the actual 300mm wafer price is about 2.25x the 200mmequivalent price (add some edge effect correction to that)
Amazing content to be honest, I'm upset I haven't discovered him earlier. Just subbed.
The size of the die makes yield trouble go up superlinearly. If there are three "bad spots" on the wafer, at 600 chips per wafer, that's 0.5%, but at 60 chips per wafer, that's 5%.
Then again, tooling, design, and testing, and the risk taken on by starting that process, is a significant cost. As is the depreciation and interest on those 100 million dollar steppers lithography machines...
15:40 A factor in cost not listed here that's very important is testing. That includes quality, functionality and performance testing to make sure customers receive a part that works properly and as advertised. As well as binning to sort assembled packages according to their performance to begin with. This testing comprises a few major steps, one of which is what I do for a living.
The NRE for each CEREBRAS chip is extremely high, an NRE is definitely part of cost.
Retired Engineer as in "@RetiredEngineer" He's an awesome dude.... Great content as always, keep it up your growth is phenomenal!
Oh yeah #SiliconGang wut wut!
I don't even know how I ended up here, but I enjoyed it.
ASH and FOUPS would make the ultimate LP record player selection system. XD
Imagine a Vinyl museum with a warehouse of records and you can listen to any record on demand
@@hansdietrich83 You mean a cellphone?
Jk, not hating on you vinyl lovers, I understand the reasons for it. Just commenting about how amazing it is that we can store that amount of music in our pocket nowadays.
Subbed bro, very good knowledge thanks :)
gonna start looking into sophie wilson's lectures now!
For small runs you can't neglect the tooling costs for masks. More complex designs also need more masks per wafer and more layers and steps in the fabrication.
This video gives me a much better understanding of Fabs and wafer production, but I'm still blown away by the price. I thought I would be much more expensive!!
Also i think the most important factor for the price is the supply/demand mechanism
Thanks
Once again, LOVE the T Shirt. You have great taste in Tshirts Ian!
Yeah, as you said the cost for a fab is huge. The lifetime of a node is limited, so the fabs need constant upgrading. The chip designs are also extremely expensive, and it takes a large amount of experts around 5 years to design a gpu or cpu
Thanks great job in breaking it down!
Nice, looking forward to reading up on Cerebras. Thinking of buying stocks when they finally do an IPO.
I remember last year being hyped on AMD and TSMC and thinking why were their stock so low. Now they boomed but so did a lot of other tech plays.
AI and semiconductor companies are the future.
I've a video on Cerebras' latest chip, hope you get a chance to watch!
My job is literally to manage and reduce wafer cost for one of the big semi companies. Interesting video to see what info you have and what you don't ;)
you also need to factor in building & equipment depreciation impact on the wafer cost. You could have one company with different Fabs with different depreciation profiles (time delays in launch between fabs) on the same node with completely different wafer costs at the same volumes. Also geo impact of freight on raw materials, labour costs, layout with different fabs can be significant deltas
Thank you, have subscribed. Your channel is great
Thank you, now I understand where prices going - to the Moon) just because Intel and AMD can say "Deficit" and rise margin to x25 per wafer)
And special thanks to Steve from Gamers Nexus for posting a link to your video. Definitely cost to subscribe)
He posted a link to my video?
The breakdown of costs of materials is all well and good but there is that one huge elephant in the room... R&D, cost of maintaining equipment and so on... These costs account for most of the actual spending for a certain company and not the materials themselves
Very interesting. It would have been great to mention an estimate of some of the fixed costs, particularly the cost of mask sets. My understanding is that these have become really expensive, and amortization of their costs would likely make up a significant part of the cost of lower volume chips, like the Cerbras offering.
They have already masks when they work in that ultra sensible ambient
This is cool and all, but how does 7nm taste compared to the over processes?
Depends whether you use the standard N7 or N7+. You can really taste the EUV layers.
@@nathangamble125 The initial samples had a slight tang to them that remained on the tongue briefly after swallowing!
I have been waiting for this type of content since you started your channel
Dude! Nice videos! Good explanations, no fluff. Subbed.
Graet video
I always thought one silicon cost from a wafer will sell like the whole wafer
For example if company sell x or y cpu that cost x or y it means the whole wafer cost x or y
9900k sold for 500$ bucks that means whole wafer of 9900k worth 500$
OH MY GOD!!! FINALLY! THANK YOU
My Minimum Specification is *always* 42.
Choose arbitrary proprietary unit and scaling / gamma etc to suit...
#Universal
you need profits to pay for r&d, pay for salaries, pay for business strategies, trial and error, some cash on hand to survive bad times, money for legal litigations, insurance, etc etc. usually the min. is 30-40% but the ideal is 80% for a growth company as the risk is extremely high.
"What's your minimum specification?"
Not entirely sure how to answer this, but I'm currently on an AMD 5800XT with 32GB (4x8GB) DDR 3600 CL16, and an AMD 6900XT.
Don't forget to factor in ASP (Average Selling Price) of the items. Not every processor die is sold as the top SKU, so not every die makes the same money.
Great presentation. One thing you failed to mention is the NRE costs of going to wafer production. Reticules are damn expensive. This is a massive expense when you need 50+ reticules. Please discuss this too. Maybe do a video on the whole reticule design and development process ?
If you keep in mind, that you can buy an Espressif ESP32 dual core 240 MHz RISC machine for under $1.70 including shipping on AliExpress.
Ian, would you be able to elaborate on the image at 1:32?
What is the reason behind the fact that they can't flip some of those yellow chips that are being printed vertically (specifically the ones on the 3 and 9 o'clock) which would make them viable healthy chips?
There's a small ring around the processor that's called the 'edge loss'. You can't create processors right at the edge, and the edge loss is different for different factories/nodes and such. The yellow die would be green if the edge loss was lower, that's really what it means. The calculator already moves the mask/die location to maximize the number of green dies. There's a link to the site in the description if you want to play around with the tool.
@@TechTechPotato Thank you for the explanation, but I've heard you explain that in a previous video. That wasn't my question. I was specifically asking why those 3 and 9 oclock chips can't be flipped from horizontal to vertical to fit them inside the edge dead zone. Something to do with manufacturing?
ua-cam.com/video/skUCP2f4HIM/v-deo.html
Presumably that would require rotating the mask or wafer just to add a few more dies? I'd assume that'd add non-trivial complexity to the lithography machine, and it would only be something that could be leveraged by a subset of products using very rectangular dies.
@@Hostilenemy As well as the lithography problems hughJ mentioned, it would also introduce more complexity when cutting the wafer into individual dies, where it might not be possible (or impractical) to cut and hold an irregular arrangement like that.
Do you really think that they would happily throw away money?!
That layout gets all the attention from hundreds of eye, before starting anything...
It's worth mentioning that the Cerebras chip isn't the most expensive component in the box, by at least an order of magnitude (I don't think the amounts have been publicly released).
thanks for doing all the hard work all I needed to do was just watch really thank you
Does the cost of wafers go down over time? I think i read its $10000 a wafer for 7nm tsmc, google says $9,346, do you know if prices over time go down as the tech gets older.
I wish you would compare Amd apu 8core single die VS amd 8 core chiplet cost wise.
its important to understand that price is not related to cost, but to perceived value to the customer. People are happy to pay a lot for chips because of the improvement they offer to their lives and businesses. It is also important but also very difficult to recognise that a major part of the cost of the manufacturing process is the research, development, manufacture and depreciation of the machines that make the machines.
Thank you for this video!!
great vid, thanks Dr. Wafer Eater
Really interesting video! I need a bed/kitchen/bathroom foup
They need to make a huge amount of profit in order to continue R & D and to make sure they don't go broke if and when the orders stop for a period of time. There's also the cost of running everything else. Sure, they take advantage of their monopoly, but would you really prefer the possibility of them going out of business? Or for one of the other companies like Apple, Samsung, Microsoft, Intel, etc, to own them?
Please do make a video on Silicon Ingots! I would love to know more about those!
This was a great video thanks for doing this dive on this topic.
A friend of mine, a production manager at Intel, told me that the first generation Core CPUs were not coming along well. They were scrapping 97%.
"You're clearly buying it for more than it costs to make"
Externalities: "Am I a joke to you?" :)
All well and good talking about manufacturing costs, but how much is spent on r&d, management of it, publicity, marketing, and supporting roles, etc..that costs an ARM AMD an inteLEG too.
How many 8086s could fit on a 7nm 12 inch platter? That said I'm a fan of the 186, 386 486 686 and Pentium 4. 386 was the last processor without intentional hardware backdoors.
@18:50 It is like comparing a single component vs a singleboard computer. :)
This is unique and valuable content and is much appreciated. I waited for the cats before hitting like, though.
Very interesting video about chip manufacturing! 04:44 - Millions of wafers from a single ingot? That's one big ingot that stretches from here to the ISS :)
glad I found this channel. good stuff :)
I wonder if the rise of cost per gate is because they had to go from planar to 3d gates and also wonder when the switch to GAAFET happens the price will receive an ulterior unusual spike.
Anyway great video, thanks to it now i can watch and learn more stuff from Dr. Sophie Wilson's presentations.
I can tell that collab works well on YT. Just subscribed
Great content as always, one question: Dr Su always points out that margin on console-SoCs increase over the cycle, does this factor in a decrease of wafer costs and if so can AMD expect the cost of TSMC 7nm wafers to decrease any time soon?
TSMC have stopped discounting at the moment, I can't see that changing any time soon with demand so high
At the beginning of the cycle all the CapEx is done, which doesn't factor into future production. But usually over a processor cycle, between AMD and TSMC, there are additional efficiencies that can be made. A slightly better process step that gives better yield, or perhaps a tweak that saves 50 millivolts. Or in the past, a redesign for a smaller process node that ends up being cheaper.
The cost of TSMC 7nm wafers obviously isn't going to decrease any time soon when they can't get close to meeting demand at current prices and production is hobbled by ongoing water shortages.
As a non-native English-speaker, I had no idea, what an "Ingot" was. But today I learned, that "I" "stretch" "silicon seed" "into" an ingot, and that "the size" of "my" "ingot" depends on "the speed with which "i pull it out"..? This one(!) sentence will now haunt me for the rest of my life. Engineering can come in huge packages and is hard most of the time, I guess.I had no idea how much I did wrong in this regard. I did, however, know what a "Fap" was.
make a video on how cerebras gets around defects, that would be interesting, are they just including in chip redundancy or is this something more sophisicated?
Yes.
The best chips is always going to be Fish and 'chips'. You know what I'm saying, Potato. 😂
Wholesome and yummy too.
They might taste better, but suck a bit in calculation speed.
Nothing beats Haddock and chips, deep fried in beef dripping. Now given my age, its chips deep fried in rapeseed oil and cod given where i live Haddock is just not available.
Now let’s start talking about the vinegar / salt with which they’re seasoned. Malt and Himalayan pink are my personal favorites.
4:20 - Quartz IS a type of sand so yes, ingots are refined from sand. Check Intel learning video about the process if you want to know more.
Please make a video on Quartz to Silicon Wafers production
Could you do one on how assembling chiplets works? Such as the different methods and the future methods they are looking at? What I guess without any knowledge is they use a masking technique and they just mask two chips together, but it is a guess.