I don't even want to share what I have in mind. This forum has become horrible for people making guesses that happen to come out wrong. I still get thrown in the face stuff I predicted wrong from years ago, some people here are [censored] eager to use strawmans at the first opportunity.
(...)
(BTW what the hell do these people do in their lives that's so boring that they get off on searching/keeping/whatever years-old quotes from unimportant anonymous users in internet forums?!)
Theeere we go. Here it is.
Out of curiosity, how long did it take to find that specific 4 month-old post on page 116 of a closed thread?
Was it worth it, at least? Is there that much satisfaction from engaging in strawmans?
You were wrong before so you must be wrong again! Haha you're discredited!!
I don't mind having been wrong about the SoC supplier. I was the first user to bring the nvidia news to the forum, it's the very first post in this thread, lol.
Why not engage in my arguments instead of pulling strawmans? How does that benefit discussion? Bonus internet points for you, I guess?
I mean you already have your own thread where
you made your own rules of prohibiting people from speculating on speculation. What more do you want?
Same numbers (because they were silicon peak).
If there is such a thing (nominal clocks can be surpassed through multiplier configuration and higher voltage and/or current), I'm positive TX1's
silicon peak wouldn't be at 2GHz CPU and 1GHz GPU.
Get access to nvidia's custom kernel and a competitive overclocker working on it and I'm positive that by feeding up to 50W into TX1 while attaching a CPU cooler to it then both the GPU and CPU could be clocked significantly higher than what the Shield TV can do with 22W (pulled from the wall BTW).
Which, in the end, means that those clocks wouldn't be
silicon peak as you say. At best, they would be the maximum granted by the specific devkit that was handed to the developers.
But then why would the devkit even allow 1GHz GPU and 2GHz CPU if the maximum target clocks were so much lower?
So you think it's much bigger silicon with more shader cores?
The "much" part is your own deduction. It could even have substantially more shader cores and be smaller.
1 - Cortex A53 cluster stripped out
2 - Video encoder can be much simpler/smaller as no one is going to use the Switch to record 4K HEVC videos
3 - Same goes for image processor, it doesn't need to handle JPEGs at 600MPixel/s because the Switch won't be used to take e.g. 40MPixel pictures at 15 frames-per-second.
4 - PCI-Express interface and some other I/O are useless and can be stripped out too.
5 - if Cortex A57 cluster is running at low clocks then they can be printed for higher density / smaller footprint.
Take all of these into account and maybe it fits an extra SM with the left over space.
Plus, if this is a custom chip, there's no reason to assume the Streaming Multiprocessors as a fixed and known value.
It could be each SM with more ALUs like Kepler's SMX as a closed and hardware-specific development environment could allow for higher utilization from the get go.
Pascal consumer chips for example have no 2*FP16 throughput, but they do have dedicated FP16 ALUs. We could be seeing one or more smaller SM units with FP16 ALUs only.
These are just ramblings and each of these tweaks has a very low chance of being true. I'm aware of that.
The only message I want to convey is there's a lot of missing pieces besides the clocks. SM count and SM configuration are just an example of missing info.
The logic here clearly points to the UE profile being wrong.
You could assume the Unreal Engine 4 developers made the Switch performance profile wrong, or you could just look at it as yet another piece of the puzzle that doesn't fit.
If that spec PC can't run it at a decent framerate, there's no way a portable is!
It doesn't seem to be that far away from 30 FPS. Console implementations punch above PCs with similar theoretical specs. The 650 Ti is a 4 year-old GPU and the Maxwell already boosted performance-per-GFLOP by a lot.
According to nvidia, a Maxwell 1 SM has 90% of the performance of a Kepler SMX. So 4 Maxwell SMM at 1GHz (512 ALUs, 1TFLOP) would have 90% the
real performance of 4 Kepler SMX at 1GHz (768 ALUs, 1.5 TFLOPs).
I'd be backing DF here rather than the ToTTenTranz horse.
Lol I'm practically on the same horse as DF, which is
there's a lot of missing info.