alt Hacker NewsSo I got curious about the progression of processing power, specifically how long ago did a GPU have equivalent to the latest iPhone chip? The iPhone 17 Pro has the A19 Pro, which has ~2.5 FP32 TFLOPS. The RTX 5090 has ~100 TFLOPS, so a factor of 40. Obviously there are higher end cards than the 5090 and FP32 performance is only one of many metrics so nothing about this is perfect but it is interesting.
The first consumer NVidia GPUs with similar FP32 FLOPS performance were in about 2011-2012 but were expensive. By 2016-2017, the 1060 was a very accessible consumer card with similar performance. So you're looking at about a 10 year lag from best consumer GPUs to a GPU with similar performance to a modern phone.
This is what people are spending trillions on. Put another way, their investment is going to be worthless in 10-15 yyears, absolute max. That's a very short time to recoup trillions in investment.
Obviously this depends on further shrinking and improving chips but I'm old enough to remember that same discussion and it being unknown if the future was XIL or EUV or if both of these would fail. Still, we are getting down to a handful of silicon atoms wide.
But the future here I think will be in interconnects so you don't need ever-bigger chips and you can scale horizontally much more effectively.
Oh and for comparison, the M5 has ~4.2 TFLOPS and the M5 Max has ~18 TFLOPS, for comparison.
As for it being a war, of course it is. That's what the US government does: it protects the interests of US companies and their owners. Look at the history of Bombardier-Boeing or all the atrocities committed in the name of the United Fruit Company, including multiple military coups and the ongoing embargo of Cuba.
US companies want an AI moat. China doesn't, ergo China is the enemy because no moat destroys US tech company value.
Replies
Ehhh, the question comes down to can you cool a chip with ~100 TFLOPs in the size of an Iphone package. Not really as much about the cost of the chip itself or if you can cram it in.
Packing in more transistors, sure probably possible, packing in more transistors while keeping it cool enough to touch? Totally different ballgame
> So you're looking at about a 10 year lag from best consumer GPUs to a GPU with similar performance to a modern phone.
Two competing viewpoints to this:
1) It is getting harder to make the same performance gains, so maybe that 10 year window grows to 15 or 20.
> Put another way, their investment is going to be worthless in 10-15 yyears, absolute max.
2) The value of a GPU is not its flops relative to to other GPUs. Its value is it's output minus it's cost. If the value of its output is stable, or grows, it doesn't really matter if its efficiency relative to the latest and greatest diminishes.