Hey everyone!
So let's say I'm thinking about making a clone of the Antminer S19j Pro with 104TH/s powering at 3068Watts. I believe it's built on 5nm tech.
I want to chat about a few things:
1) What’s the minimum investment to design a 5nm ASIC?
2) How long does it usually take from start to finish?
3) What are the technical specs we’re looking at (like transistors per 1TH/s and all that)?
4) What’s the price for a 300mm wafer?
5) Any other key questions we should consider?
If someone has links to similar discussions here, drop them in!
To kick things off, here are some of my thoughts.
The Antminer S19j consumes 3068Watts. Now, let's compare that with a 5nm AMD CPU. Take the AMD Ryzen 7 5800X for example, it has about 4.15 billion transistors and uses 105Watts.
Doing a rough calculation: 3068 divided by 105 times 4.15 gives us around 121 billion transistors for the Antminer S19j. Not super accurate, but I’d say the error margin is about 10% or so.
The Ryzen 7 5800X sells for around $449.
The cost for 1 billion 5nm transistors is about $2.25.
So, if we do the math, the prime cost for Ryzen 7 5800X would be 4.15 times 2.25, which equals about $9.34. For the Antminer S19j, it's 121 times 2.25, which gives us approximately $272.25.
When I say "prime cost," I mean that’s just for the wafer, not the final product like cut dies and packaged chips.
Got several hundred million dollars to risk? ref this article from 2019More to the point, once you have a chip designed you then have to test the design and there is NO guarantee that it will perform as expected. The Canaan A921 miner using 10nm chips from Samsung is a poster child for that -- actual chip performs around 50% of what modeling said it should run.
Fact is, if you have little to no experience in designing high performance mixed-signal chips you can dream but should forget about actually doing it...
NotFuzzyWarm,
I have no experience in designing chips, but for me it is very very interesting theoretical price of bitcoin ASIC.
I think for many bitcointalkers too.
Numbers from your article are saying nothing.
how many transistors are in this 28nm planar device or in 7nm soc ?
I can provide you info about russian quasi-CPU, each develop cost was +- 25mln usd.
baikal-m 28nm
die size 350mm2
transistors 3.5bln
elbrus-8c 28nm
die size 240mm2
transistors 2bln
so we have develop cost
71-104thousand usd per 1mm2 or per +-10mln transistors
hundreds thousands dollars, not hundreds millions.
it's 28nm technology, 5nm is more expensive.
***
Of course chip should be tested.
Fabs has special service called "shuttle". You can provide them design as low as 1mm2 and they will create chip for testing purposes.
https://europractice-ic.com/schedules-prices-2021/
GF 12nm 26000eur/1mm2
TSMC 28nm 9000eur/1mm2
Once again, we speak about thousands, but not millions dollars.
I can't find prices for 5nm shuttle. But I feel it will be in 100000eur/1mm2 region.
***
sha-256 core is very simple in design.
one-round need about 20k NAND gates or 80k transistors.
bitcoin uses two-round sha-256 = 160k transistors.
5nm has about 160mln transistors per 1mm2.
So for 1mm2 5nm design we just need to copy-paste 2000 sha-256 cores and design special control unit to synchronize theirs work.
It's really not hundreds millions dollars for such design.
That's all ! We don't need gigantic design such as 15mm*20mm. Such design of course can cost hundreds millions of dollars.
I like this thinking.
Why are we relying on the Chinese to make these and then make a killing off them?
We should already have USA made miners! These machines don't cost that much to make, and they are very simple machines.
One thing to be said though is that we would need to create our own chip manufacturing as well as we don't want to again: rely on China for chips.
-Graham
My friends, there's a reason for everything, including the fact that chips are made in Taiwan. And the cost to launch a new product is huge. It is very naive to think otherwise.
It only makes sense if you are backed by a firm that's willing to support your R&D and prototyping expenses.
I don't think if there is USA made miner I have never seen any miner made in the USA there are some unique miners that I heard from offordscott telegram and I think some of those miners are US made but outside the telegram I never heard any of them.
Most of the manufacturers are Chinese and I don't know if there is another miner created outside China. Maybe there are but not widely promoted here or outside the forum.
For making a clone of those miners I don't know you can make a clone of these units like Bitmain gears might have their own serial number built-in with a unique chip but if you can also develop your own program maybe you can make a clone that running the same as the original one. I'm sure you can save many people on buying very expensive miner.
There were US manufacturers of bitcoin miners, Butterfly Labs, Cointerra, Hashfast. They all ended up as bankrupt scams. (There are posters here still around who got burned bad who might chip in)
Then we had KNC in Europe, started out well, ended up a disaster, then Spondoolies in Israel, who again started out strong, then diminished to nothing in the onslaught of bitmain.
Although all these "manufacturers" pretty much depended on TSMC and Global Foundries for chip production.
There are others, if you're interested take a look here: https://en.bitcoin.it/wiki/List_of_Bitcoin_mining_ASICs not sure how up-to-date it is with the latest ASIC's but its interesting from a historical standpoint.
It could be cheaper to buy intellectual rights to some extinct manufacturer's design (for example: KNCMiner, Spondoolies, BW.com/FusionSilicon, LightningAsic...) than to start from scratch. But even that requires a lot of capital and further development to create a competetive asic chip. The miners power efficiency should be between 20 - 40 J /TH to be competitive vs. other manufacturers like Bitmain, Canaan, MicroBT.