Aren't we pretty sure that Switch is not using a Vulkan derivative as its main API? I thought it's using a proprietary Nvidia API ("NVN"?).
There's no evidence of BW starvation any more than there's evidence of ALU starvation, any more than there's an evidence of command-buffer starvation, any more than there's evidence of the engine lead having had a bad case of a hangover. I hope I've made myself clear.
They do (and it's not simple by far).
Neither of those cases given 'clear examples of BW starvation'. Just because it's not strictly higher blending rates that they exhibit over other areas in the game, but also notably higher geometry rates. Also, while any GPU could be BW starved with the 'right' approach, alpha effects per se would not be a good candidate for that on the TX1, for reasons I already explained. You'd need a technique that requires the tile to get resolved to RAM before readback.
No, and I don't claim to have in my head, either.
BW/pixel comparison would be meaningful if it used correct figures. Your analysis above does not.
Not so much lazy in the sense that they didn't put work into it, but they rushed it and its not optimized.
I dunno. Whatever true successor we get to Pascal and uses like a 7nm node that can do 1 to 1.25 TFLOPS, should be able to trade blows with PS4 when you factor newer hardware efficiency and mixed precision mode. This I also assuming they use the same amount of bandwidth, amount of RAM and CPU power
Naw. The A57s in the Switch are a lot more similar to the Jaguars (and "Jaguar evolved" in Scorpio) than the chips from the previous generation. Those chips were clocked a lot higher than the ones in the current consoles, but they generally not nearly as good. At 1020MHz, the individual cores in Switch are roughly close to the Jaguars in the PS4 in performance. If developers are able to utilize all 6+ cores for gaming in the PS4 (which may may not be close to the case with some game engines like the previous versions of Unity), the Switch can do about half of that with its 3 cores. It would have been nice if Nintendo was able to clock the CPU higher, but it is still much better than the tiny CPU in the Wii U.
Not true even before the patch.
I didn't say they didn't put work in it. What you're saying us what I mean. They literally ported the Wii U version on to the switch, and brute forced it on switch. They made it working with better framerate and slightly better resolution, though considering the Wii u is 4x as powerful, they could have done more if given more than 10 months. . Not that I'm really complaining.. Zelda is great on switch, though has some hiccups.
If NVIDIA and Nintendo didn't make a deal, I wonder if we would have even seen anymore smaller-scale Tegras. Before now, Tegra's technology's success in the phone/tablet industry was very limited.
You have. Given that stance, we can conclusively say that it's premature for you to be dismissing people who think Switch is bottlenecked by bandwidth. And it's premature for someone to say that Switch definitely is bottlenecked.
I apologize for the errors, then. Can you please explain what's wrong with the figures?
Wait, I'm not dismissive of everybody who ever said switch could be BW-limited - as I said, any GPU could be made to hit its BW limits under the right circumstances. I'm dismissive of people who claim switch's GPU performance was BW-limited (and BotW was a 'clear indication' of that) because of that main RAM BW. We have a couple of wiiU up-ports where the switch demonstrates clearly superior BW, and the wiiU (one of the data points in your table) had way more BW than 25GB/s to its fb, and way less BW to its main RAM.
Okay. That must mean you think Switch's effective main RAM bandwidth is comparable to other current machines (since otherwise you couldn't dismiss the claim). So see below....
Definitely, there are clearly details that my rough calculations didn't capture! But if you can dismiss claims of bandwidth starvation because you know bandwidth is fine, then what is the final effective bandwidth/pixel of Switch? I assume you have to know that it's comparable to PS4 Pro and Scorpio at least, if not standard Xbox One and PS4.
Sorry, I didn't follow. Which claim?
We must be talking across each other at this point. I don't know what the Maxwell2 BW in the TX1 is - I've said as much already. What I'm saying is that the 'See, the alpha effects in BotW prove switch is BW-capped at 25.6GB/s' crowd got it wrong. How do I know this? Because the BW available to the GPU is not 25.6GB/s, particularly when it comes to the alpha effects.
The claim that Switch is definitely bandwidth-starved. To categorically dismiss those claims--which you have done--you must know that it isn't bandwidth-starved, right?
The problem is, I don't recall any people saying only 25.6GB/s could result in a bottleneck. The claims seemed broader: that Switch is bandwidth-starved, with no threshold at a specific number. So I'm not sure what exactly you were dismissing in the initial post I responded to. If I understand what you're saying here correctly, it's that just the 25.6 GB/s number is false. The Switch could still easily be bandwidth-starved, no matter what the true effective bandwidth is.
No, to categorically dismiss those claims one must know that those claims were not based on sound reasing. For instance, if I definitively claimed time was 12:34 from looking at my watch, you could refute my definitive claim if you knew my watch was inexact, even if you didn't exactly know the correct time. Not because the time could not hypothetically be 12:34, but because you knew I had no ground for making such a definitive statement.
Check the first BotW DF analysis thread. From post #447 on (apologies for not providing a link - I'm on mobile here)
But as you say, I could only refute your surety about the specific value; I couldn't refute an assertion that it's too late to see a movie (since I don't know what time it actually is). Given that analogy, it does seem you intend only to dismiss claims that 25.6 GB/s is demonstrably inadequate bandwidth, not all claims that bandwidth is inadequate. And this is where I don't know who you're really referring to.
No problem, here's the thread. There is only one post in the entire thread that specifies the Switch bandwidth, this one:
A single claim doesn't seem like a wide target for your argument, even if we read the tentative "maybe" as sarcastic.
I think Blu and I was looking into that awhile ago, and LCGeek also has some links with benchmarks. It did have data on A57s handedly beating Jaguars clock to clock. However, I do admit that it may not be enough to make up for to the ~640MHz difference between Switch's CPU speed and the Jaguars in the PS4. Switch's clock speed is 60% of those Jaguars, so the actual performance will be somewhere above that. I may check it out later; searching things like this on mobile can be troublesome.
This is the post from the thread I was referring to (again, apologies for not providing the link originally): http://www.neogaf.com/forum/showpost.php?p=231418033&postcount=447
Yea 25.6gb vs 58.4gb. 16nm would probably also allow them to clock the CPU so it would be a lot closer to current gen in performance. But no point in crying over spilled milk.
Yes, its clock normalisation is borked, thanks to Geekbench's habit of showing the clock of the LITTLE core in the test results and me overlooking that. Huawai Angler is a snapdragon 810 with big cores running at 1.9GHz, so the table should be:
Single-Core Score 1304 967 686
AES 945 916 497
LZMA 1221 1191 643
JPEG 1561 1177 822
Canny 1659 988 873
Lua 1173 799 617
Dijkstra 1416 1366 745
SQLite 1206 751 635
HTML5 Parse 1381 902 727
HTML5 DOM 1259 819 663
Histogram Equalization 1697 1022 893
PDF Rendering 1631 926 858
LLVM 1562 825 822
Camera 1700 1003 895
SGEMM 472 300 248
SFFT 977 472 514
N-Body Physics 778 650 409
Ray Tracing 892 744 469
Rigid Body Physics 1424 977 749
HDR 1743 1333 917
Gaussian Blur 1284 813 676
Speech Recognition 964 1012 507
Face Detection 1355 861 713
Memory Copy 1399 1056 736
Memory Latency 1821 3459 958
Memory Bandwidth 1312 647 691It may be worth it. You get a 60% clock speed power efficiency with Pascal. Bandwidth can be held a bit more back in handheld mode
Tx2 power consumption is not higher than tx1
With Splatoon 2 it seems like the devs are prioritizing dropping no frames over everything else and it runs the same engine probably.
Single-thread, yes.
Yep.It's cool. I'm glad that you told us so that I wouldn't lead people with false infomation.
Ok, so is this what we are looking at?:
first column is A57 @ 1.9GHz, second column - Jaguar @1.6GHz (PS4), third column - first column / 1.9 (Switch).
If that the case, the Switch's performance may be around 70% of PS4's in single core performance.