Hacker News new | past | comments | ask | show | jobs | submit login

Both 3080TI and 4080 are roughly 600mm2+ Die Size. 2080 has a 700mm2+ on a mature node.

1080TI, the only outliner in your example has less than 500mm2 of die size.

And all of that are before taking account of GDDR memory costing, the board, higher TDP requirement and heatsink differences.

While I dont think 4080 pricing is good value or cheap by any standard. I also dont think it is ridiculously expensive. From a transistor stand point, you are getting 2.7x the transistor between 3080Ti and 4090. With 4090 You are buying 76 billion transistor, with 24GB of GDDRX for only $1600.




$/transistor could be the biggest variable here. My understanding is that we haven't been seeing this lower even with node shrinks. Analysts have speculated that this is why Xbox released a trimmed down console (series s)- because a cheaper "slim" wasn't realistic. Similarly Sony raised prices because inflation is outpacing their ability to reduce costs.


>$/transistor could be the biggest variable here. My understanding is that we haven't been seeing this lower even with node shrinks.

Purely from a Wafer, Transistor per Die Space hence Cost per transistor, we are still getting it cheaper per node generation. The ratio, cost per transistor may not be dropping "as much" as what we were used to. But it is still dropping.

However the Initial cost for product development on new node is increasing. This fixed cost, ( think CAPEX ) amortised over "fixed volume" would mean the cost per "die" would also be increasing.

So once you add those together you will see we are at the end of the power curve.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: