5 ms·

Do we not have 10ghz processors? My impression is that they exist, or you can at least overclock to that level, it's just neither reliable nor economical, and r

Do we not have 10ghz processors? My impression is that they exist, or you can at least overclock to that level, it's just neither reliable nor economical, and real workloads that require that much horsepower can by and large be parallelized, so the difference between quad-core 3.5 ghz and 10ghz is mostly academic.

ajross 14y ago
We have 10GHz digital logic for sure (Thunderbolt drives its differential lines that fast, I believe). You can do that for a small region of a chip with a handful of transistors. But you can't run a whole 50+ mm^2 die at that speed for the simple reason that it can't be cooled by non-exotic methods. You just can't get the heat out fast enough, so you can't ship the product. So for now, with ambient temperature cooling, the best modern chips top out at about 5GHz. That will likely improve slowly with each process generation due to efficiency improvements, but it's not going to change much.
- Daniel_Newby 14y ago
  If you can run a quad-core chip at 3 GHz, then there it is almost certainly feasible to run one core at 10 GHz. I suspect the real limit is in the size of the L1 and predecode caches. I expect they are limited by propagation delay, not transistor speed. The only way to get more cache is to use more cores.
  - ajross 14y ago
    Heat transfer doesn't scale quite like that. Once the surface area gets beyond "tiny subcircuit" you're limited by the essentially 1D transfer "up" through the packaging and can't rely on the ability of the cooler silicon in the surrounding regions. One core is still a really big circuit. And yes, caches are big regions and hard to make faster synchronously. But you can treat this with pipelining -- that's one reason even the 32k L1 data cache in SNB/IVB has a 4 cycle (!) latency. Note that the even the gargantuan L3 cache on these chips (which is literally 1/3 of the die area) is still running at full speed, just with a ~30 cycle latency.
    - Daniel_Newby 14y ago
      I was assuming silicon-on-sapphire soldered to a copper heat spreader. No numbers, but it feels plausible it could scale up. Wow. I did not know the L3 cache ran so fast.
    - wtvanhest 14y ago
      Lower energy consumption should also help the heat problem. Less waste in to heat.
sliverstorm 14y ago
Clearly they were talking about what would be widely commercially available. We probably had 10GHz test units in 2000, thanks to GaAs.
sbochins 14y ago
Guess the world record thus far is, 8.429 ghz. http://www.anandtech.com/show/4770/amd-sets-world-overclocking-record-with-8429ghz-bulldozer-processor http://www.anandtech.com/show/4770/amd-sets-world-overclocki... Doesn't really seem practical to run a datacenter using the methods in the video however. But, it's still possible.

ajross 14y ago

We have 10GHz digital logic for sure (Thunderbolt drives its differential lines that fast, I believe). You can do that for a small region of a chip with a handful of transistors. But you can't run a whole 50+ mm^2 die at that speed for the simple reason that it can't be cooled by non-exotic methods. You just can't get the heat out fast enough, so you can't ship the product. So for now, with ambient temperature cooling, the best modern chips top out at about 5GHz. That will likely improve slowly with each process generation due to efficiency improvements, but it's not going to change much.

Daniel_Newby 14y ago
If you can run a quad-core chip at 3 GHz, then there it is almost certainly feasible to run one core at 10 GHz. I suspect the real limit is in the size of the L1 and predecode caches. I expect they are limited by propagation delay, not transistor speed. The only way to get more cache is to use more cores.
- ajross 14y ago
  Heat transfer doesn't scale quite like that. Once the surface area gets beyond "tiny subcircuit" you're limited by the essentially 1D transfer "up" through the packaging and can't rely on the ability of the cooler silicon in the surrounding regions. One core is still a really big circuit. And yes, caches are big regions and hard to make faster synchronously. But you can treat this with pipelining -- that's one reason even the 32k L1 data cache in SNB/IVB has a 4 cycle (!) latency. Note that the even the gargantuan L3 cache on these chips (which is literally 1/3 of the die area) is still running at full speed, just with a ~30 cycle latency.
  - Daniel_Newby 14y ago
    I was assuming silicon-on-sapphire soldered to a copper heat spreader. No numbers, but it feels plausible it could scale up. Wow. I did not know the L3 cache ran so fast.
  - wtvanhest 14y ago
    Lower energy consumption should also help the heat problem. Less waste in to heat.

sliverstorm 14y ago

Clearly they were talking about what would be widely commercially available. We probably had 10GHz test units in 2000, thanks to GaAs.

sbochins 14y ago

Guess the world record thus far is, 8.429 ghz. http://www.anandtech.com/show/4770/amd-sets-world-overclocking-record-with-8429ghz-bulldozer-processor http://www.anandtech.com/show/4770/amd-sets-world-overclocki... Doesn't really seem practical to run a datacenter using the methods in the video however. But, it's still possible.