Xbone's GPU - 1.31 TFLOPS, Switch - 393 GFLOPS docked. That's exactly 30%.
Xbone's CPU - 8 (but only 6 for games) Jaguars clocked at 1.75 GHz vs. 4 A57 clocked at 1 GHz.
How do you end up from this to 45% ?
Yes, to 540p most likely.
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
Nintendo Switch Dev Kit Stats Leaked? Cortex A57, 4GB RAM, 32GB Storage, Multi-Touch.
- Thread starter Robert at Zeboyd Games
- Start date
- Status
Xbone's GPU - 1.31 TFLOPS, Switch - 393 GFLOPS docked. That's exactly 30%.
Xbone's CPU - 8 (but only 6 for games) Jaguars clocked at 1.75 GHz vs. 4 A57 clocked at 1 GHz.
How do you end up from this to 45% ?
Yes, to 540p most likely.
I agree with you, but we can split hairs over the %, it should be 45% unless you are talking about the slightly faster xb1s. It would only need to be 40% to do 720p of 1080p games though, much like the difference between switch portable and switch docked.
As for the CPU cores, we are speculating until we hear about a split, which we haven't. If and when we do, we can look at those numbers, currently what we know is Developers have 4 A57 cores, which is what I was saying.
480p would be the correct ratio, so similar to what you get on Wii U gamepad but without all the extra blurriness from streaming (think how reds looked vs every other color) It isn't idea, I'm not trying to say this is what I would have made, but it isn't a lost cause either.
ChrisMahoney
Banned
Fans would be required for the Switch to not get throttled while docked and played in a hot room
EatChildren
Currently polling second in Australia's federal election (first in the Gold Coast), this feral may one day be your Bogan King.
I'll drink to the day Nintendo releases any hardware where the CPU isn't by far the most absurdly weakest link. It's borderline historic at this point how routinely Nintendo fuck this up.
Shogmaster
Member
45% of XBO? Which stage of grief is this now? Delusion? Must be an extra stage for Nintendo diehards. lolz
I don't think you guys have any idea of how ridiculously awful sub res games can look on a portable screen.
There's this thing called lowering details, and with today's temporal filtering techniques, releasing a game that is straight up 540p would mean not having a fucking clue about what they're doing.
There's this thing called lowering details, and with today's temporal filtering techniques, releasing a game that is straight up 540p would mean not having a fucking clue about what they're doing.
The funniest thing here is thinking back to those wsj early articles that talked about Nintendo using industry leading chips in response to the criticism they got for the Wii U hardware not being able to keep up with the competition. Reading that quote after seeing the CPU clockspeed is both hilarious and infuriating.
Accounting for Nvidia architectural advantages over AMD. The ratio is about 4:3 in NVIDIA's favour, so Switch would have 393*4/3 = 524 GFLOPS AMD equivalent. 524/1300*100% = 40%, so we are far along already. Don't know why 45 instead of 40, though.
People have been complaining about "tablet CPUs" in the PS4 and Xbone, so it may not be that Nintendo went out of their way; they just had a really low target.
You'll have your day of drinking.
TLZ
Banned
Isn't that ratio valid only on PC, though?
Hmm, I couldn't say. I know it was used multiple times before in these discussion to compare with XB1 and PS4, but maybe I am wrong? Could use clarification from someone who is sure about whether this is true for consoles as well, please.
Shogmaster
Member
Right. And we can disregard the RAM bandwidth, RAM amount, CPU core count etc.?
A system's ability ain't just the floating point operations numbers, yeah?
Not really. The 1.39TF 750ti can still keep up with the 1.84TF PS4 just fine in most titles and that's with the PS4's supposed console optimisation advantage.
Accompanied by what CPU?
Shogmaster
Member
Switch might have RAM bandwidth advantage.
Accompanied by what OS overhead ;P
Yes, I pointed that out back in page 98.
http://m.neogaf.com/showpost.php?p=226673825
I was arguing about fab nodes to disprove that it can't be running on a 16nm node considering the low clock speeds while also having a fan.
I should have been more specific and said that it shouldn't need a fan while it is portable at those clock speeds considering its Wattage should be around <5W. (With reference to one of the recent patents that state that the fan is active even while portable.)
Speaking of which for anyone else reading.
http://www.anandtech.com/show/9289/the-nvidia-shield-android-tv-review/9
Here is a picture that shows the thermals of the Shield TV. We know it uses a TX1 and is at its stock clock speed, it uses around 20W while running games and it uses a fan.
They showed it at 34°C which is only around a 10°C increase compared to room temperature which means it's warm but not in anyway hot compared to say the PS4 which can get up to 56°C which is noticeably hot when you feel the air coming from the vents but not near overheating obviously.
http://wccftech.com/playstation-4-heat-regulation-test-thermal-camera/
So yes? I hope it is a bit more, this chip is realy not suited for gaming. I know you can not compare a pc to a dedicated machine but... I do not see it being able to run AAA games reasonably. At any settings or resolutions.
I think most people doesn't understand what BIG.little means, most devices and specifically Tegra use cluster migration, meaning that it either uses 4 a57s or 4 a53 in low power mode, but in a device that is going to work at high power mode most of the time, having those a53 means almost nothing aside from saving a bit of battery.
On the other hand, this matter is the missing piece in the Switch puzzle, in my eyes if there has been some kind of customization over the Tegra X1 is in the CPU setup, my Espresso theory maybe too wild but if it's not true I would expect a couple of extra a53 working in an HMP setup along with the a57s (Nvidia has already made an HMP setup for Tegra Parker, so they know how to do them)
But I would advise against expecting higher clocks or extra SMs, that seems like wishful thinking and setting yourselves for extra disappointments.
Yeah, using 1-2 CPU cores and 2-3 GB RAM is like no overhead at all.
The point is that it's not that easy to provide a ratio outside PC where you can use similar specs and just switch the GPU.
If you add on top the memory and memory bandwidth and compensate with tile base rendering it becomes even blurrier.
Still don't see how all these could account for an increase to 45% of Xbone. Not at these clocks. Maybe if you use a bit of FP16, but then it becomes even blurrier.
Read on.
XB360 is 240gflops with the ancient AMD architecture from 2005. Wii U is 176gflops with the ancient AMD architecture from 2008. Yet []quote[]"Some of the developers we spoke to indicated to us that the console will have 50% more processing power compared to the PlayStation 3 or Xbox 360. This is yet to be confirmed by Nintendo."[/quote]
XB1 is 1330gflops with the ancient GCN architecture from 2011. Wii U's 176gflops is only worth about 128gflops of GCN. putting XB1 at about 10x the Wii U's performance. enjoy Now for the next part, Maxwell and Pascal have identical flop performance, this is also true of GCN and Polaris (GCN 4.0) which is lucky for us because we can then match up the GTX 1060 and the RX 480 to find that flop for flop, the 4.3tflop GTX 1060 is 40% faster than the 5.8tflop RX 480 on average.
393+40%= 550+ this is just to match the architecture differences. There is always going to be outliners, but this is true among the largest samples. XB1 at launch was 1228gflops, -55% is 552gflops, putting Switch's performance over 45%. XB1 currently is 1330gflops, -58.5% is 552gflops, putting Switch's performance over 41.5% XB1. XB1s is 1400gflops -61% = 546gflops, putting Switch's performance over 39% XB1s
Again without any new features or mixed precision.
We've been talking about jaguar cores vs A57 cores, I'd compare that to PS4 rather than XB1, but it isn't far off the mark for XB1 either as cores are only clocked ~9% higher.
All about context, I'm only talking about the graphical capabilities.
Shogmaster
Member
There is so much stretching and reaching in that post, you really need to start a yoga studio.
It's a propeller in case you have nowhere to use the kickstand.
Shogmaster
Member
And graphics pipeline ain't just a bunch of shader units either.
To be fair all three went with a weak cpu this gen.
Actually I don't think the Wii U cpu is absurdly weak compared to the rest of the hardware. Its absurdly weak overall.
I'm comparing architectures. Do you think AMD makes worse drivers on PC than for XBox or PS4? they would be well suited. We've also seen maxwell on pc out perform PS4 with far less gpu performance. Do you just not believe in Architectures yielding different results? I've even shown you VLIW5 vs GCN PC to PC, there is literally 100+ benchmarks to compare GCN/polaris to Maxwell/pascal, they point to a 40% advantage on average. Going across to consoles, you have developers optimizing for consoles, so you see a smaller advantage, but it is still ~30%. I don't see why you have such a problem with the idea that Switch's performance can be compared to XB1 through benchmarks of these architectures on PC where they would have the same optimizations.
lwilliams3
Member
Do you believe there is an advantage of GCN vs NVIDIA Cuda cores, but not to the percentage he is calculating?
I'm doing what I can to add to this thread, we are comparing benchmarks of A57 to Jaguar, but somehow maxwell/pascal vs polaris/gcn isn't valid?
The whole previous architecture stuff was to prove that architectures do provide increased performance. It isn't some fairytale
Shogmaster
Member
I will repeat as many times as you guys need to understand it: A system ain't just it's FLOPS number, and the system's graphics pipeline isn't comprised of just a bunch of shader units.
We are discussing what we know, we can't discuss what we don't know, but I'm comparing the entire graphics pipeline of Pascal to Polaris and finding 40% better performance on average, that is with Pascal having less memory bandwidth too, so I think you are just missing the concept of what we are doing here. Discussing specs that we do have.
I believe in architecture yielding different results, but I believe that you can't just use the ratio from PC and apply it to Switch and Xbone. And going for the max ratio and then even round it up to 45%. Especially since the memory bandwidth plays a big role in all of this.
Shogmaster
Member
Pascal? Polaris? What the hell are you doing man? As you said, we don't have all the information needed, but you are happily cranking out precise numbers based on Pascal and Polaris which has nothing to do with anything. Lordy... You need to take a break, man. You are just masturbating mathematically at this point.
Memory bandwidth of the GTX 1060 vs RX 480 is much lower actually. The 45% was rough estimations done in my head, I'm at a computer right now so I did the extra work and came away with better than 41.5%, but it really only needs to be at ~40% to do 720p versions of 1080p XB1 games. Infact the entire clock speed difference, puts the handheld at 40% the docked speed and does exactly this.
I'm doing my best to give us references we can use to determine Switch's performance since it is more or less known what the base performance is. I've said over and over that these are estimations of baseline performance.
PdotMichael
Banned
If only there were Tegra X1 devices to compare superior nVidia flops to plebs AMD flops.
Didnt he get permajunniored during the WUST threads?
Shogmaster
Member
If only... *scratches chin stubble*
Eolz
Member
Anyone that played on Vita probably knows how bad it looks. It's pretty bad yeah.
Shogmaster
Member
I'm just at work killing time in a speculation thread, speculating. You are brow beating and not reading. Pascal and Polaris have no performance advantage in the metric I'm measuring (flops) over their "previous" architectures. Hell Pascal didn't even exist 2 years ago, it was added to the roadmap because volta was delayed and Nvidia decided to shrink maxwell again.
Here PARKER is maxwell based, it is just called pascal:
Nope, it was a thread I created at 4am, posted a fudzilla article that was anti apple in tone, the mod that juniored me is unknown and when I asked another mod about it, they didn't know anything but didn't want to step on toes. I never asked again because I don't really need to post threads. Also, I'm right here man, you can just ask me about it.
ok, kinda makes sense. Nvidia flops are always more efficient than AMD flops.
lwilliams3
Member
Lets change the topic: what do you think is up with the dev kit? They had barely edited TX1 docs, but we now getting these clock frequency numbers and it doesn't seem like it wasn't downgraded throughout its development.I didn't follow all the drama. Thanks for the heads up. I will move on.
- Status