How Much Do CPU's And GPU's Cost To Manufacture?

[Dash Lambda]

This is something I've always wondered. It seems like we can only judge the value of a computer component based on the value of other computer components, making an environment in which manufacturers can charge whatever they damn well please.

 

I'd specify two costs:

-The cost, per chip, of manufacturing over the production lifetime of the model or series factoring in the cost of materials, facilities, workers, and unique (or consumable) tools.

-The same cost, except factoring in R&D as well.

 

-When I say cost of materials, I don't mean the cost of that amount of silicon or gold or whatnot. I mean the cost of the processed materials, like the wafers they're printed on (I think those are outsourced, right?).

-When I say facilities and unique tools, I only refer to things specific to the manufacture of that chip. That means dies, changes to accommodate a new process node or material, along with any other specialized equipment that may be required.

-Advertising and any other PR expense is irrelevant.

 

I guess it would also be prudent to draw a distinction between the chip and the package. It's less important for CPU's, but I'd think there's a stark difference between the cost of a GPU and the cost of a graphics card.

 

I don't expect there to be a definitive answer since nobody from Intel, AMD, nor Nvidia probably want to share, but I hope someone here has a bit of useful input.

[Syntaxvgm]

I don't know, but I know for making chips you have to consider the cost of each run, which will leave you with only so many higher end chips depending on how they are binned, so you have to sell so many of the lower end chips or basically discard them. If you priced the higher end ones to close to the lower end ones, you would risk not selling those chips you basically end up with. Not every chip made can be a flagship, so I do have a feeling the higher end chips, even not the top end ones, are much higher margin in theory. 

[Enderman]

well, this has been discussed about 72 200 000 times previously, including several dozen the LTT forum, and nobody has any kind of exact number

 

so I would suggest looking at those already existing topics instead of making a new one

[BuckGup]

The cheap part is making the chips. The expensive part it getting yields high and R&D. 

[GoodBytes]

Let's look at R&D....

An Nvidia new architecture (like a real new architecture, not refresh), requires about ~2000 engineer. They are paid on average 100,000k, and takes 5 years to develop. That is 1 billion of dollars. A refresh GPU is about ~1000 engineer.

 

You have employee benefits not counted for, marketing, HR, administration fees, etc.

[Syntaxvgm]

1 minute ago, SamStrecker said:

The cheap part is making the chips. The expensive part it getting yields high and R&D. 

Also, imagine you're intel, making your new enthusiast 10 core 1700$ because you realise "oh fuck, this competes with xeons AND it's unlocked"

[Prysin]

not much. a 4790k probably costs around 50-75 bucks to manufacture. But then you need to take the binning process into the equation.

[manikyath]

the "silicon" itself is actually surprisingly cheap.

 

but.. the R&D of the chips is quite astronomical, there's actually a youtuber that made a very interesting video from nvidia's "failure analysis" lab, which should give you an idea of the amount of money that is in this buisiness:

oh, and +1 to that engineer giving the tour

[Sauron]

Obviously it depends greatly on the specific part.

 

What you need to consider in these cases are yields. i7s come from the same silicon wafer as pentiums, but they are priced a lot higher; why? Think about the fact that out of a wafer you might be able to get about 80 cpus. Out of those, only 10% or so will be good enough to become core i7s. You're paying the rarity of the chip.

[JefferyD90]

19 minutes ago, Dash Lambda said:

This is something I've always wondered. It seems like we can only judge the value of a computer component based on the value of other computer components, making an environment in which manufacturers can charge whatever they damn well please.

 

I'd specify two costs:

-The cost, per chip, of manufacturing over the production lifetime of the model or series factoring in the cost of materials, facilities, workers, and unique (or consumable) tools.

-The same cost, except factoring in R&D as well.

 

-When I say cost of materials, I don't mean the cost of that amount of silicon or gold or whatnot. I mean the cost of the processed materials, like the wafers they're printed on (I think those are outsourced, right?).

-When I say facilities and unique tools, I only refer to things specific to the manufacture of that chip. That means dies, changes to accommodate a new process node or material, along with any other specialized equipment that may be required.

-Advertising and any other PR expense is irrelevant.

 

I guess it would also be prudent to draw a distinction between the chip and the package. It's less important for CPU's, but I'd think there's a stark difference between the cost of a GPU and the cost of a graphics card.

 

I don't expect there to be a definitive answer since nobody from Intel, AMD, or Nvidia probably want to share, but I hope someone here has a bit of useful input.

The BOM cost of a SandyBridge CPU was somewhere in the neighborhood of $20.  So the i7 that you paid $350 for, they made a $330 "profit" on.

 

Now I say profit cause that didn't count labor, manufacturing machines, cost of building, or even paying the engineering team.

[BrianS]

'bout tree fiddy

[BrightCandle]

As far as I can tell a wafer start is about $5000 including TSMC's profit margin with all its labour and such applied. So the yield combined with the harvesting of dies dependent on the size determines most of the cost of a GPU or CPU. A GPU does have additional costs in the RAM chips and the board itself and besides the RAM its not substantial, most of the cost remains in the die itself and somewhat in the VRAM chips.

[cleb]

As other people will say, cost of making a single cheap, probably not much, cost of development and research of that chip. Ridiculously high. 

[Bhav]

11 hours ago, manikyath said:

the "silicon" itself is actually surprisingly cheap.

 

but.. the R&D of the chips is quite astronomical, there's actually a youtuber that made a very interesting video from nvidia's "failure analysis" lab, which should give you an idea of the amount of money that is in this buisiness:

oh, and +1 to that engineer giving the tour

'This is our 2.6 million dollar backside probe system'.

 

Where can I get one of those?

 

[manikyath]

44 minutes ago, Bhav said:

'This is our 2.6 million dollar backside probe system'.

 

Where can I get one of those?

 

ask your "bean counter" for the money

[BrightCandle]

Some might find this article interesting around TSMC's process and associated costs: http://semiengineering.com/finfet-rollout-slower-than-expected/

 

Quote

At 16nm/14nm, there are 66 mask steps. With an 80% fab utilization rate at 16nm/14nm, the loaded cost is about $4,800 per 300mm wafer, according to Gartner. “It takes three months from wafer start to chip delivery,” he added.

And to give you an idea of what a wafer can net - a GTX 1080 is 314mm^2 (17.7mm x 17.7mm estimate) which comes out to 180 dies per wafer. A "mature yield" would be about 95% of those being viable although potentially cut down. Gives an idea of the sort of costs involved and the markup involved. Most silicon companies run at 50% profitability.

[2FA]

23 hours ago, Sauron said:

Obviously it depends greatly on the specific part.

 

What you need to consider in these cases are yields. i7s come from the same silicon wafer as pentiums, but they are priced a lot higher; why? Think about the fact that out of a wafer you might be able to get about 80 cpus. Out of those, only 10% or so will be good enough to become core i7s. You're paying the rarity of the chip.

Actually Intel manufacturers their dual cores separately from their quad cores, thus wasting less silicon on non-functioning parts. At least that is what I have been told.

[Sauron]

44 minutes ago, DeadEyePsycho said:

Actually Intel manufacturers their dual cores separately from their quad cores, thus wasting less silicon on non-functioning parts. At least that is what I have been told.

It's possible, I suppose it would depend on the generation etc. Still, the overall point still stands

[Garfieldus]

Lets try to do some math:

 

300mm silicon wafer has surface area of 70.650mm² (Pi * r²  or 3.14 * 150²).

 

14nm skylake quad core K series has a die size of 122.4mm² meaning we can have 577 CPU's on a single wafer. However we should remember we cannot effectively use all of the surface area of the wafer around the edges + space needed for cutting out individual CPU's. So for the sake of math I'll assume about 500 CPU's per wafer.

 

I found that complete wafer cost about 5 000$ making individual CPU to be about 10$ each.

 

However as many pointed out, there are other costs that are way more staggering. For example FAB it self can costs as much as 10 billion dollars or more + few more billion dollars in designing and refining new architecture.

[BrightCandle]

8 hours ago, Garfieldus said:

Lets try to do some math:

 

300mm silicon wafer has surface area of 70.650mm² (Pi * r²  or 3.14 * 150²).

 

14nm skylake quad core K series has a die size of 122.4mm² meaning we can have 577 CPU's on a single wafer. However we should remember we cannot effectively use all of the surface area of the wafer around the edges + space needed for cutting out individual CPU's. So for the sake of math I'll assume about 500 CPU's per wafer.

 

I found that complete wafer cost about 5 000$ making individual CPU to be about 10$ each.

 

However as many pointed out, there are other costs that are way more staggering. For example FAB it self can costs as much as 10 billion dollars or more + few more billion dollars in designing and refining new architecture.

 

Intel sells 300 Million or so processors a year. Even though a modern fab does cost 10 billion they get to spread that cost over 2-3 years, so around 600-900 million processors. The end result of which is that the fab cost is only $11 a CPU. That isn't separate to the price of wafer cost, the price TSMC charges for a wafer covers their costs, that is why those numbers seem so similar.

[DXMember]

for Intel it pays only 10~15% of the cost you purchase the CPU from the store to manufacture the chip

a big part of the cost is crazy high R&D expenses, equipping the manufacturing sites and modernizing laboratories, for Intel as they have very cost inefficient way to do it but a great one to rapidly advance the technology

and there's of course the profit margin