Currently Online 
This post explains why 8nm doesn't make sense from a "Nintendo goes cheap and low power" perspective.
Either our assumptions about power consumption are wrong or the process node (8nm) is wrong. Which one seems more likely?
"However, while it's reasonable to design a chip with intent to clock it at the peak efficiency clock, or to clock it above the peak efficiency clock, what you're not going to see is a chip that's intentionally designed to run at a fixed clock speed that's below the peak efficiency clock. The reason for this is pretty straight-forward; if you have a design with a large number of SMs that's intended to run at a clock below the peak efficiency clock, you could just remove some SMs and increase the clock speed and you would get both better performance within your power budget and it would cost less."
"The above section wasn't theoretical. Nvidia and Nintendo did sit in a room (or have a series of calls) to design a chip for a new Nintendo console, and what they came out with is T239. We know that the result of those discussions was to use a 12 SM Ampere GPU. We also know the power curve, and peak efficiency clock for a very similar Ampere GPU on 8nm.
The GPU in the TX1 used in the original Switch units consumed around 3W in portable mode, as far as I can tell. In later models with the die-shrunk Mariko chip, it would have been lower still. Therefore, I would expect 3W to be a reasonable upper limit to the power budget Nintendo would allocate for the GPU in portable mode when designing the T239.
With a 3W power budget and a peak efficiency clock of 470MHz, then the (again, not theoretical) numbers above tell us the best possible performance would be achieved by a 6 SM GPU operating at 470MHz, and that you'd be able to get 90% of that performance with a 4 SM GPU operating at 640MHz. Note that neither of these say 12 SMs. A 12 SM GPU on Samsung 8nm would be an awful design for a 3W power budget. It would be twice the size and cost of a 6 SM GPU while offering much less performance, if it's even possible to run within 3W at any clock.
There's no world where Nintendo and Nvidia went into that room with an 8nm SoC in mind and a 3W power budget for the GPU in handheld mode, and came out with a 12 SM GPU. That means either the manufacturing process, or the power consumption must be wrong (or both). I'm basing my power consumption estimates on the assumption that this is a device around the same size as the Switch and with battery life that falls somewhere between TX1 and Mariko units. This seems to be the same assumption almost everyone here is making, and while it could be wrong, I think them sticking with the Switch form-factor and battery life is a pretty safe bet, which leaves the manufacturing process."
"So, what manufacturing process can give a 2.5x improvement in efficiency over Samsung 8nm? The only reasonable answer I can think of is TSMC's 5nm/4nm processes, including 4N, which just happens to be the process Nvidia is using for every other product (outside of acquired Mellanox products) from this point onwards. In Nvidia's Ada white paper (an architecture very similar to Ampere), they claim a 2x improvement in performance per Watt, which appears to come almost exclusively from the move to TSMC's 4N process, plus some memory changes."
With 4N just about stretching to the 2.5x improvement in efficiency required for a 12 SM GPU to make sense, I don't think the chances for any other process are good. We don't have direct examples for other processes like we have for Ada, but from everything we know, TSMC's 5nm class processes are significantly more efficient than either their 6nm or Samsung's 5nm/4nm processes. If it's a squeeze for 12 SMs to work on 4N, then I can't see how it would make sense on anything less efficient than 4N.
Last edited by sc94597 - on 15 September 2023
