>It also helps me in latency because the best thing you can do for latency is be outside of the silicon. The wires in the substrate run 10 to 50 times faster than in silicon wires.
Can someone explain what he's talking about here, in more detail? I though the whole point of staying monolithic is because on-chip communication is faster and generally better than chip-to-chip communication, no? If what he's saying is true, why aren't we just still building processors on PCBs from discrete transistors? What metric is it that is 10-50 times faster? Surely not the signal propagation speed?
It has to do with the size of the wires. Larger cross section reduces resistance and increases propagation speed. The metal layers used on chip will have a significantly smaller cross section than on the substrate. When you use advanced packaging you have interposers that also use much larger cross sections on their wires. This is a huge simplification I’m making but that’s the gist.
Still though, if that is so, why has the industry generally moved to greater and greater integration, bringing caches, northbridges, memory controllers, GPUs, &c&c onto the silicon instead of disaggregating? Surely there are contexts and limits to this argument?
Yes, but there’s no way I could distill all my VLSI classes into a few sentences here.
A few are:
a) shore space on the chip is limited, you can’t go off chip for everything
b) taping out multiple chips is expensive
c) increasing the number of dies you package together reduces yield (Intel GPU Max)
d) if the distance is very short it will be much faster to stay on chip than to go over a phy and then the substrate. So it depends on what you’re doing which solution is better. Also as process nodes have shrunk so has the cross section of wires. So it’s gotten worse over time. (This probably the answer you are looking for)
>It also helps me in latency because the best thing you can do for latency is be outside of the silicon. The wires in the substrate run 10 to 50 times faster than in silicon wires.
Can someone explain what he's talking about here, in more detail? I though the whole point of staying monolithic is because on-chip communication is faster and generally better than chip-to-chip communication, no? If what he's saying is true, why aren't we just still building processors on PCBs from discrete transistors? What metric is it that is 10-50 times faster? Surely not the signal propagation speed?
It has to do with the size of the wires. Larger cross section reduces resistance and increases propagation speed. The metal layers used on chip will have a significantly smaller cross section than on the substrate. When you use advanced packaging you have interposers that also use much larger cross sections on their wires. This is a huge simplification I’m making but that’s the gist.
Still though, if that is so, why has the industry generally moved to greater and greater integration, bringing caches, northbridges, memory controllers, GPUs, &c&c onto the silicon instead of disaggregating? Surely there are contexts and limits to this argument?
Yes, but there’s no way I could distill all my VLSI classes into a few sentences here.
A few are:
a) shore space on the chip is limited, you can’t go off chip for everything
b) taping out multiple chips is expensive
c) increasing the number of dies you package together reduces yield (Intel GPU Max)
d) if the distance is very short it will be much faster to stay on chip than to go over a phy and then the substrate. So it depends on what you’re doing which solution is better. Also as process nodes have shrunk so has the cross section of wires. So it’s gotten worse over time. (This probably the answer you are looking for)
I can't even read the transcript, YOU KNOW. Comes up 35 times.