4 ms·

Could please someone clear up the following for me: So his processor draws up to 335W. The core voltage is to my knowledge around 1.3V, so this processor draws

by deng 4y ago

Could please someone clear up the following for me: So his processor draws up to 335W. The core voltage is to my knowledge around 1.3V, so this processor draws over 255A. How is this possible? Shouldn't these tiny structures inside the processor immediately evaporate with currents like these? I mean, they are not superconductors, they must have some resistance? Why do I have to spend so much money on AWG4 cables for my battery storage, and somehow these rules do not apply to microprocessors?

spijdar 4y ago
Not my expertise, but I assume this is (one of) the reasons why CPU packages have so many dedicated power delivery pins, like in https://www.techpowerup.com/img/RJC7aj7xUhLmvxVD.jpg https://www.techpowerup.com/img/RJC7aj7xUhLmvxVD.jpg
formerly_proven 4y ago
Like half the pins in a CPU socket are ground and power. The copper planes carrying power from the VRM to the socket are massive. Wide and thick. The CPU die itself has large networks of metal interconnect that get finer and finer to supply smaller and smaller areas, much like our circulatory system. You can see these on die shots of older parts as well, e.g. http://www.righto.com/2016/12/die-photos-and-analysis-of_24.html http://www.righto.com/2016/12/die-photos-and-analysis-of_24.... notice the thick traces which start to fan out quickly. Also consider that there's 10+ billion transistors for those 200 A to go around.
nigerianbrince 4y ago
I read somewhere that the energy density in modern cpus is 10% of the surface of the sun. Don't quote me on that.
- adgjlsfhk1 4y ago
  I'm pretty sure it's much higher. The sun isn't very energy dense.
- katmannthree 4y ago
  I'd be curious to see how one measures peak energy density in a modern CPU. As far as power density goes, the i9-13900K seems to be 335W / (257 mm^2 die area * ~1mm water thickness) ≈ 1.3e9 W/m^3. The power density of the sun in comparison is around 2.8e2 W/m^3, roughly what you'd find at the center of a large pile of compost (although the temperature at the sun's core may be just a wee bit higher).
- sliken 4y ago
  A cubic meter of sun produces ... 0.27 watts! The sun's output is 3.8 x 10^26 watts and the volume is 1.4 x 10^27 cubic meters Surface of the sun is 6.09×10^12 km^2, which comes out to 62.4 watts per square mm. The AM5 socket is 40x40mm, so that would be 100,000 watts of the sun's surface.
LeifCarrotson 4y ago
Your #4 cables do not dissipate heat nearly as effectively as the liquid cooling solution they're using on this test bed [1]. For this processor to continue dissipating 335W without overheating, it needs to dissipate that 335W of heat energy into hundreds of cubic feet of air every minute. Six fans push and pull the air through hundreds of tiny fins, which are conducting heat out of liquid coolant continuously pumped through tubes soldered to those fins and through a giant copper waterblock millimeters from the point where this heat is produced. And even with all that, it probably can't actually run at 335W continuously. You don't have a $200 cooling solution on every foot of your battery cables, they're wrapped in PVC or silicone, buried in conduit, in plastic wire wrapping, or through insulation in a wall cavity. As an EE, those rules are painfully applied to microprocessors, such that a ton of PCB design energy, mechanical constraints on packaging, and other costs are spent removing that heat. You have the convenience of just buying the recommended AWG4 cables and it just works, regardless of the severely sub-optimal thermal design you might employ when running that cable. In theory your batteries could pass 255A through a single strand of bare copper from an Ethernet cable for a few tens of milliseconds before it went into thermal runaway and liquefied/evaporated. Remove that heat fast enough (liquid nitrogen?) and you could run tiny cables to your battery storage. [1] https://www.ekwb.com/shop/ek-aio-elite-360-d-rgb https://www.ekwb.com/shop/ek-aio-elite-360-d-rgb
- deng 4y ago
  Thank you. While that makes sense, it still blows my mind... These things are absolute marvels of engineering and really push the laws of physics. Next to 255A of current, these things can almost reach 6GHz frequency. Just incredible.
  - sliken 4y ago
    Heh, yes, sort of. Sure it's over 5 GHz. But their somewhere around 19 pipeline stages. So 19 parts of the chip, do 1/19th of the work of a single instruction, under ideal conditions (data is in registers, functional unit is available, etc). So something as simple as a=b+c gets split into 19 pieces. Sadly waiting on cache (l1, l2, and l3), main memory (4 32 bit channels approximately 100ns away), missed branch predictions, etc is fairly common. So delivered performance is much less than you'd think form a 6 way issue @ 6 GHz * 8 to 16 cores. Sadly we are still stuck at 128 bit wide memory, as a result even the most parallel of codes like cinebench go up by 50% when you double the number of cores.
adgjlsfhk1 4y ago
The biggest reason is that the cpu is only carrying those 250A for about 1 inch. Resistance is directly proportional to distance.
tails4e 4y ago
1.3v blows me away, at 7nm transistors were typically using 0.8v due to oxide breakdown at high voltages. I wonder how they can run at 1.3v and not kill those tiny transistors gates. Dynamic power being proportional to voltagr squared will hurt perf/watt significantly also.

by deng 4y ago

spijdar 4y ago

Not my expertise, but I assume this is (one of) the reasons why CPU packages have so many dedicated power delivery pins, like in https://www.techpowerup.com/img/RJC7aj7xUhLmvxVD.jpg https://www.techpowerup.com/img/RJC7aj7xUhLmvxVD.jpg

formerly_proven 4y ago

Like half the pins in a CPU socket are ground and power. The copper planes carrying power from the VRM to the socket are massive. Wide and thick. The CPU die itself has large networks of metal interconnect that get finer and finer to supply smaller and smaller areas, much like our circulatory system. You can see these on die shots of older parts as well, e.g. http://www.righto.com/2016/12/die-photos-and-analysis-of_24.html http://www.righto.com/2016/12/die-photos-and-analysis-of_24.... notice the thick traces which start to fan out quickly. Also consider that there's 10+ billion transistors for those 200 A to go around.

nigerianbrince 4y ago

I read somewhere that the energy density in modern cpus is 10% of the surface of the sun. Don't quote me on that.

adgjlsfhk1 4y ago
I'm pretty sure it's much higher. The sun isn't very energy dense.
katmannthree 4y ago
I'd be curious to see how one measures peak energy density in a modern CPU. As far as power density goes, the i9-13900K seems to be 335W / (257 mm^2 die area * ~1mm water thickness) ≈ 1.3e9 W/m^3. The power density of the sun in comparison is around 2.8e2 W/m^3, roughly what you'd find at the center of a large pile of compost (although the temperature at the sun's core may be just a wee bit higher).
sliken 4y ago
A cubic meter of sun produces ... 0.27 watts! The sun's output is 3.8 x 10^26 watts and the volume is 1.4 x 10^27 cubic meters Surface of the sun is 6.09×10^12 km^2, which comes out to 62.4 watts per square mm. The AM5 socket is 40x40mm, so that would be 100,000 watts of the sun's surface.

LeifCarrotson 4y ago

Your #4 cables do not dissipate heat nearly as effectively as the liquid cooling solution they're using on this test bed [1]. For this processor to continue dissipating 335W without overheating, it needs to dissipate that 335W of heat energy into hundreds of cubic feet of air every minute. Six fans push and pull the air through hundreds of tiny fins, which are conducting heat out of liquid coolant continuously pumped through tubes soldered to those fins and through a giant copper waterblock millimeters from the point where this heat is produced. And even with all that, it probably can't actually run at 335W continuously. You don't have a $200 cooling solution on every foot of your battery cables, they're wrapped in PVC or silicone, buried in conduit, in plastic wire wrapping, or through insulation in a wall cavity. As an EE, those rules are painfully applied to microprocessors, such that a ton of PCB design energy, mechanical constraints on packaging, and other costs are spent removing that heat. You have the convenience of just buying the recommended AWG4 cables and it just works, regardless of the severely sub-optimal thermal design you might employ when running that cable. In theory your batteries could pass 255A through a single strand of bare copper from an Ethernet cable for a few tens of milliseconds before it went into thermal runaway and liquefied/evaporated. Remove that heat fast enough (liquid nitrogen?) and you could run tiny cables to your battery storage. [1] https://www.ekwb.com/shop/ek-aio-elite-360-d-rgb https://www.ekwb.com/shop/ek-aio-elite-360-d-rgb

deng 4y ago
Thank you. While that makes sense, it still blows my mind... These things are absolute marvels of engineering and really push the laws of physics. Next to 255A of current, these things can almost reach 6GHz frequency. Just incredible.
- sliken 4y ago
  Heh, yes, sort of. Sure it's over 5 GHz. But their somewhere around 19 pipeline stages. So 19 parts of the chip, do 1/19th of the work of a single instruction, under ideal conditions (data is in registers, functional unit is available, etc). So something as simple as a=b+c gets split into 19 pieces. Sadly waiting on cache (l1, l2, and l3), main memory (4 32 bit channels approximately 100ns away), missed branch predictions, etc is fairly common. So delivered performance is much less than you'd think form a 6 way issue @ 6 GHz * 8 to 16 cores. Sadly we are still stuck at 128 bit wide memory, as a result even the most parallel of codes like cinebench go up by 50% when you double the number of cores.

adgjlsfhk1 4y ago

The biggest reason is that the cpu is only carrying those 250A for about 1 inch. Resistance is directly proportional to distance.

tails4e 4y ago

1.3v blows me away, at 7nm transistors were typically using 0.8v due to oxide breakdown at high voltages. I wonder how they can run at 1.3v and not kill those tiny transistors gates. Dynamic power being proportional to voltagr squared will hurt perf/watt significantly also.