What sort of chances are there for Drake to produce roughly x8 the raw GPU flops that Switch is currently capable of?
- Switch portable - 157 GFlops portable @ 307.2 Mhz
- Switch docked - 393 GFlops docked @ 768 Mhz
- Drake portable (if x8) - 1.258 TFlops @ 409.6 Mhz
- Drake docked (if x8) - 3.146 TFlops @ 1.024 Ghz
Unlikely, but not totally insane.
The biggest indicator of what a next-gen device will do is just number of transistors. Nintendo put 6 times the number of SMs in in their GPU over their last one, I think 6x is likely where they'll start, and then clocks will get adjusted up or down as power can tolerate. Going 33% further seems like a stretch
Would that even be feasible? Estimated power draw on the different nodes? And how would that compare to other devices?
We know that Drake has power saving tech that other Ampere cards don't have. So I think some of our power estimates are a little shit, honestly.
Also, we don't know if Nintendo will even go with flat clock speeds. It's entirely possible that Nintendo uses a totally different strategy - like dynamic clocks with a fixed power pool, with higher GPU games underclocking the CPU and vice versa.
Steam Deck uses RDNA2 and can produce 1.6 TFlops, but being part of an APU, I believe it lacks Infinity Cache, so it doesn't have that benefit.
Infinity Cache isn't a benefit. It's tradeoff. Cache is expensive, but so is bandwidth. AMD skimped on bandwidth and added cache. In some cases, that is a win, in others it is a loss, and overall it comes out in the wash. APUs don't need infinity cache if they just have a crapload of memory bandwidth. Steam Deck is fine.
RDNA 2 FLOPS and Ampere FLOPS are surprisingly comparable. Part of why Nvidia has invested so much in features in the last few generations, is because both teams have very similar architectures when it comes to raster performance.
Dunno what else it lacks compared to desktop GPUs. Would Drake using Ampere have anything missing from desktop versions (besides the RAM bandwidth that also affect other portables)?
Bandwidth matters in relationship to the amount of textures/geometry you're pushing, and how many frames - in other words, bandwidth needs scales up as your GPU's performance increases.
We don't know Drake's bandwidth, because we don't know at what clock they'll set their memory controllers, I would expect memory performance to drop in handheld mode. But Ampere cards tend to sit around 30 GB/s/TFLOP, (with some lower). Drake will need extra for the CPU, but it's likely that Drake will be fine bandwidth wise.
Other devices like the Aya Neo2 go even farther than Steam Deck, including "boosting" itself to 3.38 TFlops by setting its GPU clock to 2.2 Ghz, but boosted is just not sustainable.
Those machines are bogged down by Windows, and don't have "real" docked modes, running the same profile regardless of whether or not they're hooked up to an external display.
I would expect a Drake based device to underperform a Steam Deck in handheld mode, but absolutely slaughter in battery life. I would similarly expect it to outperform in docked mode.
If Drake is limited - and there are those who disagree - it's single core CPU performance, not the GPU at all.
