Wednesday, February 17th 2021
NVIDIA GPUs Have Hotspot Temperature Sensors Like AMD
NVIDIA GeForce GPUs feature hotspot temperature measurement akin to AMD Radeon ones, according to an investigative report by Igor's Lab. A beta version of HWInfo already supports hotspot measurement. As its name suggests, the hotspot is the hottest spot on the GPU, measured from a network of thermal sensors across the GPU die, unlike conventional "GPU Temperature" sensors, which reads off a single physical location of the GPU die. AMD refers to this static sensor as "Edge temperature." In some cases, the reported temperature of this sensor could differ from the hotspot by as much as 20°C, which underscores the importance of hotspot. The sensor with the highest temperature measurement becomes the hotspot.
GPU manufacturers rarely disclose the physical locations of on-die thermal sensors, but during the AMD Radeon VII, we got a rare glimpse at this, in a company slide, with the sensors being located near components that can get the hottest, such as the compute units (pictured below). Igor's Lab put out measurements of the deviation between the hotspot and "GPU temperature" sensors on a GeForce RTX 3090 Founders Edition card. There's a much narrower deviation between the two (between 11-14°C), and than the one between hotspot and Edge temperature on an MSI Radeon RX 6800 XT Gaming X Trio (which posts a 12-20°C difference).
Source:
Igor's Lab
GPU manufacturers rarely disclose the physical locations of on-die thermal sensors, but during the AMD Radeon VII, we got a rare glimpse at this, in a company slide, with the sensors being located near components that can get the hottest, such as the compute units (pictured below). Igor's Lab put out measurements of the deviation between the hotspot and "GPU temperature" sensors on a GeForce RTX 3090 Founders Edition card. There's a much narrower deviation between the two (between 11-14°C), and than the one between hotspot and Edge temperature on an MSI Radeon RX 6800 XT Gaming X Trio (which posts a 12-20°C difference).



44 Comments on NVIDIA GPUs Have Hotspot Temperature Sensors Like AMD
I believe the higher differential on water is from an aggressive OC though, and when the card is pushed to its limits
My conclusion is that we can't conclude anything related to chip design based on the difference between edge temp and hotspot temp.
It is not that simple, GA102 being a much bigger die also aids with thermal transfer because of the larger surface area, and lower thermal density.
Also for IC designs, generally the smaller die that achieve similar performance is considered more efficient.
Smaller die-area = higher yields etc. Although Samsung 8nm and TSMC 7nm are very different processes, thus not directly comparable.
Good cooling does not prevent hot spots even if the other areas are well below spec, and on top of that, if you are air cooling Nvidia GPUs, you'll run into high temps anyway because of GPU Boost unless you handicap your card.Nah its clearly because the GPUs are not that edgy o_O:kookoo:
Its going well here. Jesus christ.
The better the cooler, the lower the temps will be overall, including any hotspots, whether you have a sensor there or not, so it matters very much.
Imagine the scenario where the overall temp is 50C, but a hotspot is 20C higher. That still only makes 70C, so the chip is within spec and won't overheat, because of that powerful cooler.
Now imagine that same chip with an inferior cooler. The overall temp is now 70C, but the hotspot has hit a whopping 90C, or maybe even more, so the chip could be overheating, or very close to its limit.
They only do if you undervolt them and AIBs develop coolers sized to what the board can put through and the components can take. You're dreaming if you think a stock air cooler will do anything positive for hot spot temps. It'll just boost higher, because you can market boost clocks. And if a cooler stays far below, its a waste of money.
My Asus Strix RX 5700 XT runs at 75 C edge temp and 95 C hotspot temp on stock settings. Decreasing the power limit by 25% lowers the GPU voltage and clocks, but also the fan speed, resulting in the exact same temperatures (with less noise). It's not wrong, it's designed this way.
A specific load could lead to a very large delta between hotspot/GPU temp on a specific architecture but if that give huge performance, that is not a bad chip design.
The end goal of chip design is performance/watt/cost. Having a small delta between hotspot and GPU temp isn't really something that make a chip good. I get your point but doing a very bad design like putting all the hot stuff together would lead in the end in a bad performance/watt/cost ratio. So we are back to that.
Your stretching. Majorly.