How do you calculate the cost of a chip?

[JediMindTricks]

Instead of listening to all these Nvidia apologists, I want to calculate the cost of the chip myself, and how much it ultimately costs to sell us the damn thing.

I have this calculation for starters https://siliconanalysts.com/chip-price-calculator-by-tech-node/

Trying to calculate the cost of the 5060ti chip, at 181mm square, I get chip cost $260. I inputted the x axis at 10mm and y at 18.1mm but not sure if that's right, and I picked TSMC 4nm as the node.

And pls, help me in determining the cost of 8gb of vram, the other components, shipping, their profit margin and whatnot and let's see if we're getting shafted at $379 USD.

 

[boppingXQWB]

Ian goes into detail here. Check the timestamped chapter and the succeeding chapter.

The Tech Poutine #23: AMD's Moving to 2nm

Lots of news this week ahead of the foundry events coming up. AMD's 2nm Venice, both AMD and NVIDIA building in Arizona, there's a new ChatGPT, and Lip-bu at...

www.youtube.com

The tool he uses: https://semianalysis.com/die-yield-calculator/

Looking at TPU's die shot of GB206-300-A1, it has a height to width ratio of 1.116. Solving for 181mm², that's 14.21mm by 12.74mm.
In the video he states N3 defect rate is 0.075 defects/cm² (28:29), 5nm-tier (N4/4N/N5) should be lower but we will use 0.075.
I don't remember if he said wafer price for 5nm-tier, but we will go with N2's $30,000 per wafer price: in the video he states that it will likely be closer to 20k or something for Nvidia as they are a TSMC partner, and N4 or w/e the heck 50-series actually uses should be even less.

Even with the generous overestimation of the defect rate and wafer price, we get...

$122 per die.
Again, that's with overestimation of both defect rate and wafer cost, and without even accounting for how some of the defective dies that NV can use some of the defective dies for 5060.

edit: editing the original post's editing mishap
edit 2: at 35:81, Ian uses $20500 for N4 wafer cost and 0.07 for defect rate. Using those numbers, it would be $82.33 per die.

 

[evernessince]

Ian goes into detail here. Check the timestamped chapter and the succeeding chapter.

The Tech Poutine #23: AMD's Moving to 2nm

Lots of news this week ahead of the foundry events coming up. AMD's 2nm Venice, both AMD and NVIDIA building in Arizona, there's a new ChatGPT, and Lip-bu at...

www.youtube.com

The tool he uses: https://semianalysis.com/die-yield-calculator/

Looking at TPU's die shot of GB206-300-A1, it has a height to width ratio of 1.116. Solving for 181mm², that's 14.21mm by 12.74mm.
In the video he states N3 defect rate is 0.075 defects/cm² (28:29), 5nm-tier (N4/4N/N5) should be lower but we will use 0.075.
I don't remember if he said wafer price for 5nm-tier, but we will go with N2's $30,000 per wafer price: in the video he states that it will likely be closer to 20k or something for Nvidia as they are a TSMC partner, and N4 or w/e the heck 50-series actually uses should be even less.

Even with the generous overestimation of the defect rate and wafer price, we get...
View attachment 396573View attachment 396574
$122 per die.
Again, that's with overestimation of both defect rate and wafer cost, and without even accounting for how some of the defective dies that NV can use for 5060.

A quick note on this, the cost provided is based on the publically disclosed wafer cost. There will almost certainly be a different, lower price agreed upon in Nvidia's wafer supply contract. In addition, wafer cost decreases over time so that should be taken into consideration.

Mind you, one can look at Nvidia's gross margins and deduce that, being more than twice of Apple's, certainly they don't have an issue with increasing silicon cost. Silicon cost is always rising and that is offset by the increasing growth of the market. With this huge expansion into AI, it has massively increased the amount of revenue Nvidia brings in which should allow Nvidia to leverage economy of scale to lower chip prices and as development costs are spread over a great number of sales. Only they aren't doing that, in fact they are increasing prices instead. Their gross margin reflects this and has massively outpaced the cost growth of silicon. It's to Nvidia's benefit that wafer cost increases, it edges out potential competitors.

 

[JediMindTricks]

Okay, so let's say it's under $100 for a 5060ti chip, how much would the rest be? Is there a site or calculator that could estimate the cost of everything else once you have the cost of the chip?

 

[boppingXQWB]

There are websites like https://www.trendforce.com/price/dram/gddr_spot that can further help with estimating the BOM cost. 8Gb = 1GB. Using the average session price, 1GB of GDDR6 is $2.276. 16GB of GDDR6 would be $36.4. No idea how much GDDR7 costs, but it's probably not much more considering NV are willing to offer 8GB additional for $50 of additional MSRP. A large part of it will be NV, AIBs, and retailers' margins. You could assume around $50 (MSRP for 8GB of GDDR7 to end users) to $70 (double GDDR6 spot price) for 16GB GDDR7 and you shouldn't be too far off.

It gets far more difficult with other misc. parts that are AIBs' responsibilities like caps, inductors, VR modules, PCB, etc.

While trying to do a BOM breakdown is a excruciating fun thought experiment at best, I think the important takeaway is NV are charging more than ever per mm² of die with the 40 and now the 50-series. BTW, if the 40-Super-series GPUs were included, it would look a bit better than the 40-series.

The author has added toggles for inflation and cut-die calculation—two parameters whose absence I previously lamented in another post.

NVIDIA GPU Core Count Visualization

Interactive visualization of NVIDIA's CUDA core counts across GPU generations.

mr-september.github.io

 

[dark_globe]

at $429 msrp , you can imagine not only nvidia but also the 3rd party partners like MSI , gigabyte etc have to make profit as well
and also the local vendors selling the card directly to the customer
so we already have at least 3 parties that will have to make profit on this $429 chip .

i was working for a big company selling electronics here in europe and from what i know we could be selling $700 TV at 50% discount (some old model , black fridays etc.)
and still make $100 profit off of it as a vendor .
which would mean we had to buy this TV for under $300 when it was brand new and the company selling it to us still made profit ...
i think it is save to assume that the manufacturing cost of this small xx60 chip is around $150 or below so that everybody can get the cut .
(shipping cost is not really a factor if you imagine how much cards get shipped via each individual transport ,
it would icrease the cost of each individual chip by $5 maybe (not even that most likely) .
(the only question remaining is how much is each party making off of it
and whether or not it is ethicall to be selling a $150 chip at $450)
personally i think greedia is crossing the line and these cards should be branded as 5050/5050Ti and sell for $250-300
5060Ti should have a 5070s chip and 12gb of VRAM for $450 .

 

[freeagent]

If Hyte makes 5 bucks on a case after all is said and done, I have no idea lol..

Engineering, R&D.. who knows man. Probably not much.

 

[Wirko]

A quick note on this, the cost provided is based on the publically disclosed wafer cost.

No, it's based on leaks and informed guessing, and always lacking the necessary context. Quantities? Guaranteed yields or maximum number of defects? Time to delivery?

Design costs are huge and rising too, and that is only acceptable if they are amortised over an ever-growing number of chips produced. Ever-growing until when?

 

[bug]

Those formulae look spot on. If chips would be built on $0 research.

If you already calculated $260 for the bare chip, $380 for the finished product look far from a rip off. Which what AIBs have been complaining all along, that Nvidia's MSRP are eating into their margins.

*VRAM would add another $30-50 (not sure exactly, but GDDR7 is not exactly cheap atm), a cooler can add $20-50, depending on quality.

 

[TSiAhmat]

there were 2 videos (1 & 2) on the topic of GPU prices from buildzoid. (Actual Hardware Overclocking)

In short:

The GPU dies are quite huge & pricey to boot. Which is why we now see this development. (Not to mention the AI & Business demand of those chips)

We should criticize NVIDIA, AMD & Intel for some of the decision they make, but sadly price & availability aren't the ones.

(Which is why the 9060 (xt) 8gb won't be a cheap card, because cost-cutting vram right now doesn't save a lot of money/cost)