3 ms·

Increasing clocks tends to have a greater-than-linear cost on power, as you need transistors to switch quicker so often need a higher voltage, which causes more

by kimixa 2y ago

Increasing clocks tends to have a greater-than-linear cost on power, as you need transistors to switch quicker so often need a higher voltage, which causes more leakage and other losses on top of the switching cost itself (that all turn into heat). Higher clock targets also have a cost for the design itself, often needing more transistors for things like extra redrivers to ensure you get fast switching speed, or even things like more pipeline stages. Plus not all area is "transistors" - it's often easier to place related units that need a lot of interconnectivity with shorter interconnects if an adjacent, less interconnected unit isn't also trying to be packed into much of the same space, routing on modern chips is really difficult (and a place where companies can really differentiate by investing more). For tasks that tend to scale well with increased die area, which is often the case for GPUs as they're already focused on massively parallel tasks so laying down more parallel units is a realistic option, running a larger die at lower clocks is often notably more efficient in terms of performance per unit power. For GPUs generally that's just part of the pricing and cost balance, a larger lower clocked die would be more efficient, but would that really sell for as much as the same die clocked even higher to get peak results?

netbioserror 2y ago
>For tasks that tend to scale well with increased die area, which is often the case for GPUs as they're already focused on massively parallel tasks so laying down more parallel units is a realistic option, running a larger die at lower clocks is often notably more efficient in terms of performance per unit power. I should've considered this, I have an RTX A5000. It's a gigantic GA102 die (3090, 3080) that's underclocked to 230W, putting it at roughly 3070 throughput. That's ~15% less performance than a 3090 for a ~35% power reduction. Absolutely nonlinear savings there. Though some of that may have to do with power savings using GDDR6 over GDDR6X. (I should mention that relative performance estimates are all over the place, by some metrics the A5000 is ~3070, by others it's ~3080.)
- bgnn 2y ago
  Yeah the power consumption scales, to first order, with Vdd^2 (square of power supply voltage) but performance scales with Vdd. Though you cannot simply reduce the Vdd and clock rate and do more pipelining etc to gain back the performance. If you are willing to back off on performance a bit you can gain hugely on power. Plus thermal management of it is more manageable.
cubefox 2y ago
> Increasing clocks tends to have a greater-than-linear cost on power Old source, but this says the power cost of increasing the clock frequency is cubic: https://physics.stackexchange.com/questions/34766/how-does-power-consumption-vary-with-the-processor-frequency-in-a-typical-comput/61937#61937 https://physics.stackexchange.com/questions/34766/how-does-p...

by kimixa 2y ago

netbioserror 2y ago

>For tasks that tend to scale well with increased die area, which is often the case for GPUs as they're already focused on massively parallel tasks so laying down more parallel units is a realistic option, running a larger die at lower clocks is often notably more efficient in terms of performance per unit power. I should've considered this, I have an RTX A5000. It's a gigantic GA102 die (3090, 3080) that's underclocked to 230W, putting it at roughly 3070 throughput. That's ~15% less performance than a 3090 for a ~35% power reduction. Absolutely nonlinear savings there. Though some of that may have to do with power savings using GDDR6 over GDDR6X. (I should mention that relative performance estimates are all over the place, by some metrics the A5000 is ~3070, by others it's ~3080.)

bgnn 2y ago
Yeah the power consumption scales, to first order, with Vdd^2 (square of power supply voltage) but performance scales with Vdd. Though you cannot simply reduce the Vdd and clock rate and do more pipelining etc to gain back the performance. If you are willing to back off on performance a bit you can gain hugely on power. Plus thermal management of it is more manageable.

cubefox 2y ago

> Increasing clocks tends to have a greater-than-linear cost on power Old source, but this says the power cost of increasing the clock frequency is cubic: https://physics.stackexchange.com/questions/34766/how-does-power-consumption-vary-with-the-processor-frequency-in-a-typical-comput/61937#61937 https://physics.stackexchange.com/questions/34766/how-does-p...