3 ms·

How long before we can have open source CPU cores made a la carte by pure play foundries? Beyond Moore's Law I wonder if we'll just see a massive drop in the c

by api 11y ago

How long before we can have open source CPU cores made a la carte by pure play foundries? Beyond Moore's Law I wonder if we'll just see a massive drop in the cost of computing power to the point where something near parity with a 48 core Xeon is like $10. Then it would become very cheap to put tons of them in a box and build little desktop supercomputers.

tcoppi 11y ago
I think silicon still has a bit longer to go before we will run into the end of Moore's Law on it(not even counting other substrates for transistors or logic gates), beyond simply scaling there are a bunch of other techniques to squeeze performance and lower power usage that will be economically attractive if you can't shrink anymore, we've already seen some of them like clock gating and voltage scaling, but they can go a lot further. Once we get to the terminal node(probably 5nm with our current understanding) I think we will also see a lot of architecture innovation as well. Not x86 vs ARM or MIPS etc, but fundamentally different ways of thinking about processing units. I think FPGAs, GPUs, and other currently specialized architectures will become even more important, it is possible that everything in a "PC" could be as configurable as an FPGA is today. A lot of things open up once we hit the scaling wall and are forced to make better use of the given resources. I also think die sizes will grow, including the possibility of "3D" wafers, but that will bring its own manufacturing and design challenges as well. Anyway I think the end of classic Moore's Law scaling is still far enough away on the horizon that nobody can really predict what will happen, but if it does indeed happen as soon as the 2020s we will be in for a treat no matter what happens I think.
ksec 11y ago
Never? Or not for the next 10-15 years. The cost of producing (in your example ) 48 Core Xeon exceed $600 for any pure play foundry. This does not include any profits for Foundry to operate and no return on investment from their initial billions of equipment. Even if moores law is working at the same pace as before, you still won't see it in 15 years, let alone how moores law has slowed, and dead in price / transistors terms. Price for 10nm are in 10 billions. And will go even more expensive when we go down further.

by api 11y ago

tcoppi 11y ago

I think silicon still has a bit longer to go before we will run into the end of Moore's Law on it(not even counting other substrates for transistors or logic gates), beyond simply scaling there are a bunch of other techniques to squeeze performance and lower power usage that will be economically attractive if you can't shrink anymore, we've already seen some of them like clock gating and voltage scaling, but they can go a lot further. Once we get to the terminal node(probably 5nm with our current understanding) I think we will also see a lot of architecture innovation as well. Not x86 vs ARM or MIPS etc, but fundamentally different ways of thinking about processing units. I think FPGAs, GPUs, and other currently specialized architectures will become even more important, it is possible that everything in a "PC" could be as configurable as an FPGA is today. A lot of things open up once we hit the scaling wall and are forced to make better use of the given resources. I also think die sizes will grow, including the possibility of "3D" wafers, but that will bring its own manufacturing and design challenges as well. Anyway I think the end of classic Moore's Law scaling is still far enough away on the horizon that nobody can really predict what will happen, but if it does indeed happen as soon as the 2020s we will be in for a treat no matter what happens I think.

ksec 11y ago

Never? Or not for the next 10-15 years. The cost of producing (in your example ) 48 Core Xeon exceed $600 for any pure play foundry. This does not include any profits for Foundry to operate and no return on investment from their initial billions of equipment. Even if moores law is working at the same pace as before, you still won't see it in 15 years, let alone how moores law has slowed, and dead in price / transistors terms. Price for 10nm are in 10 billions. And will go even more expensive when we go down further.