I mean you're really splitting hairs here. It's true the GPU was not solely made for cars, they use Orin in other things as well, but the point is that orin was designed to use a lot more power than a tablet GPU is, that's why it can have features like DLSS and RT cores which a tablet probably isn't gonna have.
I mean, maybe. I don’t really think that applies in this case though. A GPU is still a GPU regardless, how it does and how it is supported is what determines its use case. ORIN uses a modified version of the Ampere GPU microarchitecture, but this architecture is also similar to the one found in the Datacenter GPUs. So it’s not really a “car” GPU in this case.
With a tablet GPU? Well, there’s not really a thing as tablet GPU. To pass off as something that can function for a tablet, what is necessary is that it qualifies for the power targets. The industry has more or less homogenized where one size is trying to fit all. Apple for example uses the same GPU and CPU across a variety of processors. A newer generation gets a new name for what the efficiency cores and the performance cores are of course, but what works in the iPhone can work in the iPad and has worked before. Hell, if I’m not mistaken, Apple is putting their A13 Bionic in the new display!
In the case of Drake, which is the name of the chip that will be in the Switch Next, if you want to get more into the nitty gritty, it has more in common with the desktop GPUs than with ORIN. While ORIN has more in common with the Datacenter GPUs. All are ampere of course, but slight modifications that aid it or reduce an unnecessary aspect of it to make it work in its use case.
the reason ORIN uses more power, is really just down to being clocked so much higher and/or having so much more to it on the whole entire board that demands it. For example, the chip has DLAs which are meant for training neural networks, it’s like a mini supercomputer. A dedicated gaming console doesn’t really need this. It has 12 CPU cores, consoles at most have 8 for area and efficiency concerns, Switch doesn’t need 12. ORiN has a PVA which is a programmable vision accelerator, and these all have their own individual clock frequencies. it’s a whole different beast, and this is separate from the GPU itself!
But anyway, here’s why orin can be a confusing product, or rather I should say, why NVidia is so confusing:
Officially, ORIN had 17B transistors when it was revealed in 2019. Then later one it had 21B transistors. The gentleman in the video is using an ORIN AGX which has the full feature set if I’m not mistaken, just two configs of 32 or 64GB of RAM. Anyway, here we see that he mentions ORIN AGX has
13Billion Transistors.
Of course I should mention, these are pre production OEMs that NVidia provided, I don’t know if that means anything in this case.
And I’m unsure if it’s that ORIN is 21B transistors in total and the person from this video is just referring to the GPU+CPU+Caches only equates to 13Billion Transistors, excluding the automotive only feature that do take up their own allotment of silicon. Or if ORIN is actually 13Billion transistors, instead of 21B.
Needless to say, this is confusing and throws a whole wrench in the conversation.
Actually, if the 2048 GPU cores +Caches + 12 CPU cores =13B transistors, then Nintendo with 1536+Caches+8 CPU cores would be very much below 13B and obviously the 21B transistors.
You are correct that Nintendo was in talks with Nvidia long before the switch. In fact, it was well reported that Nintendo was considering a tegra for the 3ds. But the X1 was definitely chosen for its price, considering the pascal was an option and they elected for the older version.
Pascal wasn’t even out when they selected for the TX1.