600 watts of power consumption is now considered established for Nvidia’s top models of the Ada-based GeForce RTX 4000 generation. How much of this the GPU alone could absorb now stems from a bypass that already led to valid results with Ampere.
GPU consumption = graphics card consumption – rest consumption
igor’s lab once again the task has been set and from the reported 600 watts of “Total Graphics Power” (TGP), which includes the GPU, memory and all other components of the graphics card, to calculate the power loss of the Ada GPU AD102, presumably larger in the GeForce RTX 4090.
Nearly two years ago, the same calculation experiment for GA102 with a rumored and ultimately correct TGP of 350 watts on the GeForce RTX 3090 indicated 230 watts of power consumption for the GPU alone.
AD102 with up to 450 watts of power consumption
For AD102 calculations igor’s lab with 450 watts of power consumption for the GPU alone. That would be nearly double that for GA102 on the GeForce RTX 3090, though the TGP is increased by “only” 71 percent. This is due to the fact that the GeForce RTX 4090’s circuit board and memory system will probably only slightly increase the consumption from 120 to 150 watts, so the additional consumption is mainly due to the GPU.
The memory system could even consume less electrical power (40 watts) than in the GeForce RTX 3090 (60 watts), because Nvidia, as in the RTX 3090 Ti (test), only uses twelve 2GB chips instead of twenty-four 2GB chips. 1GB is set. In return, the power loss from the power supply and circuit board increases from an estimated nearly 40 watts to just over 100 watts.
According to the latest rumors, 18,432 Cuda cores are supplied with the 450 watt RTX 4090, although this value has not yet been confirmed for the GPU configuration of the next provisional top of the range, nor for the AD102 full expansion. GA102 on GeForce RTX 3090 offers 10,496 CUDA cores, the full configuration on GeForce RTX 3090 Ti even 10,752 CUDA cores. So the AD102 doesn’t offer twice as many Cuda cores as the GA102, but Nvidia is said to have overhauled the cores significantly, and there’s also talk of much larger caches in the rumours. Instead of using the 8mm process at Samsung, the Ada is manufactured at TSMC using the 5nm process.
most likely close
igor’s lab emphasizes that the assumptions from industry sources on which the sample calculation is based were not confirmed to the exact watt, but were generally evaluated as realistic. Since the GeForce RTX 3090 Ti already has the expected memory for the RTX 4090 and board partners are already running 600 watt tests with their board, there is already more information this time around than before the GeForce RTX 3090 launch at end of 2020.