brochiller
Member
You knew this before and still thought Pascal was going to be in the retail units, correct?
Nintendo Switch Dev Kit Stats Leaked? Cortex A57, 4GB RAM, 32GB Storage, Multi-Touch.
- Thread starter Robert at Zeboyd Games
- Start date
- Status
You knew this before and still thought Pascal was going to be in the retail units, correct?
Polygonal_Sprite
Member
Don't forget that they were editing in real time for an hour after the article went up to remove whole paragraphs which people were pointing out to be factually inaccurate with regards to the tech they were talking about lol !
OK. I didn't think retail hardware would have even gone into production four months before launch.
That sounds solid. A shame if a change in plan brings you grief online. I've always enjoyed and appreciated your input.
Since I'm waiting to hear back from folks, I can't for certain but it would seem unlikely for them to switch architecture at this point.
Could the retail unit have a Pascal chip and the dev kit be using Maxwell? I guess that's possible if performance is the same - though I'm not entirely sure if that is possible.
I shared Pascal in July and reiterated my confidence in Pascal in September. I heard final dev kits in November with Maxwell.
Schnozberry
Member
That's not the only difference. The Pascal SM design used in the desktop cards (GTX 1070 and 1080) has poor FP16 performance in comparison to the design in the Tegra X1.
If they could achieve the X1 design on 16nm, they would be far better off.
I'm starting to see that mod's point with you now. You really aren't saying anything definitive, and you backed down way too quickly. Sorry, but I don't believe that you have or have had legitimate sources.
It's not possible.
LOL, get out with this shit
That's what I was thinking. It's a strange decision for them to make all things considered, but there must have been a reason for that. Maybe that was the cost advantage early rumors were talking about. Makes the Nvidia blog post about their relationship with Nintendo even funnier to me.
Sorry this is gonna impact how people take the information you present, man. I enjoy reading your stuff and I hope you continue despite this.
As someone who only spout nonsense during the NX era you really shouldn't be the one who calls out anyone, if I may say so.
Papacheeks
Banned
This makes a lot of sense to me.
It would indicate the recent rumor about battery life on newer kits being 5 Hours if it was 16nm.
For what it's worth, and I know people doubt me too, which is totally understandable because we should all be skeptical of everything until it's officially revealed ... my source says documentation with final dev kit states Maxwell.
But I don't doubt Nate at all either. Dudes info has been in line with what I've heard for like 98% of things.
It's a officially a custom chip though. The specs in the op are exactly a Jetson TX1 board, so we were wondering, where are these modifications that would give us a custom chip?
Well, if Nintendo put some SRAM (2-4MB) on die and asked for a die shrink to 16nm, that would be a custom chip. And at that point it wouldn't even matter if it's Pascal or Maxwell, because the slight architectural advantages the former has compared to the latter are largely irrelevant in a device like this.
This could be what happened or could have been a plan being considered.
Sorry, I've been doing this line of work for too long to fall for bait like this. Nintendo looked in Pascal. That is fact.
sir_bumble_bee
Member
Well, considering Nvidia themselves have said it's a custom chip, this possibility is still on the table.
Papacheeks
Banned
It makes a lot of sense and would answer the confusion we've had from different sources on pascal/maxwell stuff.
And would answer how we heard 3 hour battery life, then later on heard rumors of final devkits giving 5hours or so battery life.
brochiller
Member
Thanks. I appreciate hearing your information regardless of what the haters say
How would you put this info alongside the rumored Pascal?
A dev kit VS final hw gap or a running change?
The latter could be worrisome imho for possible third party expectations even if I think the only possibile difference could be for battery life in portable mode
Someone correct me if I'm wrong but isn't Pascal just a die shrink of Maxwell? Wouldn't a 16nm Maxwell chip just be a Pascal chip?
It can varying from a lot of variables.
For example, if that 1TF is, in true, what that machine can do when docked, then yes, it is very close to the Xbox one.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Now I understand why you avoided the question for so long.
Schnozberry
Member
FP16 performance on pascal based consumer graphics cards is much worse than that of the Tegra X1 design.
With Tegra X1, all 256 CUDA cores are capable of both one FP32 and 2xFP16 instructions per clock cycle. On the PC consumer based GTX cards based on Pascal (think GTX 1050-1080), there is only one 2xFP16 CUDA Core per module of 128.
Skittzo0413
Member
No there are apparently some other minor changes made in the architecture. This could all be a matter of semantics though if the final custom SoC is indeed made on a 16nm node.
so it is actually better if it is a maxell shrinked?
This makes it sound like a 16nm Maxwell chip would actually be the best of both worlds, offering tolerable battery life and superior FP16 performance at once.
So really best case scenario instead of worst case scenario.
Not fully. Edit: beaten with more details above.
Schnozberry
Member
Yes, for a low power system. For desktop cards, FP16 performance isn't much of a factor. If Nvidia designed the API for the Switch around maximizing the performance of a portable deseign, it will use mixed precision and encourage use of FP16 wherever possible.
Custom Tegra X1 Maxwell SM's die shrunk to 16nm would be the best possible combination.
Ha. Talk about a turnabout.
Still unclear if we are getting 20nm or 16nm but this info certainly makes either scenario make more sense.
hoping for the best then!
Skittzo0413
Member
Doesn't Parker also benefit from 2x performance for FP16? Which would indicate any Tegra Pascal chip would similarly have that advantage?
Again, seeing as how Parker is the only Pascal Tegra chip we know of yet, this could all be largely a matter of semantics.
My man, if what you're saying is true, then we really needed you in that other thread. We could have prevented 20+ pages of people blowing an aneurysm over something that could turn out better in the end lol.
I get it now (finally !). Thanks !
im pretty sure he posted it there too, but you know, Nintendoomed
Skittzo0413
Member
His post is a tiny bit misleading-
Pascal is only more power efficient due to being made on a 16nm process node, which draws less power and has less material. Maxwell, as an architecture, can also be made on a 16nm process node and get the exact same power efficiency gain.
People are conflating Pascal with 16nm and Maxwell with 20nm/28nm, which isn't really the case. A custom Maxwell chip can definitely be made on a 16nm process. And hopefully this is the case.
brochiller
Member
I mean 600 man years of development or whatever and you'd think they could pull it off.
Schnozberry
Member
I don't believe Nvidia released a detailed SM design for Parker. It does have native FP16 capability, but that may just mean it's just using die shrunk SM's from the X1, or even the compute heavy GP100 design used in pascal cards targeted at high performance computing.
Schnozberry
Member
You can't prevent Neogaf aneurysms. You just have to kind of ride them out.
Switch is now Maxwell and Nate is a fucking liar that never had legitimate sources, or something..
That wouldn't have mattered. People were reading the title alone and reacting. People were posting, "See Nintendo fans! You were dumb for thinking this was on par with PS4/XB1!" without even knowing the nature of the discussion that goes on in the switch threads. Thraktor makes a well-reasoned post that gives insight on the differences between Maxwell/Pascal that gets reposted page after page, yet you still have people only reacting to the thread title, saying "im out". Like lemmings jumping off a cliff.
I still believe that Maxwell is worst for battery (size and consumption ) but of course that thread was still embarassing
Schnozberry
Member
Certainly, at 20nm it would be. If it was die shrunk, then the difference would be minimal if any at all.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
You're assuming Tegra Pascal (AKA Parker GPU) derives from GP104 and lower desktop models. It doesn't. It's a direct derivative of the TX1 Maxwell, with all fp16 advantages from that. The statement it's a die-shrink of TX1 Maxwell is much closer to the truth than some think.
Bookmark this whitepaper*, then we can discuss how Parker GPU differs from that once the respective whitepaper is out.
* Notice the 3rd gen Delta Color Compression and end-to-end compression features? Yes, those got into desktop Pascal (but the Delta Color Compression is dubbed 4th gen, which is precisely 3rd gen with some fixes; edit oh, hey, I was wrong - they did introduce a new 4:1 compression block - that's something completely new).
It's not Pascal by itself, but 16 nm node advantages, architecture wise they are almost the same.
Active cooling on main unit might be indicative of 20nm though, but higher clockspeed than I expected.
I had forgotten about consumer Pascal being 16fp nerfed to avoid competing against their server gpus, so that's a nice perk from Maxwell (not that I'm a fan of people talking about half-precision sort of being the next "secret sauce", but there are gains to be had there)
Just stated my opinion, I don't care about the responses.
Either way, I think this is about the high-end of what to expect from switch in well-optimized games: https://www.youtube.com/watch?v=aEHhOmlyhJQ
Either way, I think this is about the high-end of what to expect from switch in well-optimized games: https://www.youtube.com/watch?v=aEHhOmlyhJQ
Schnozberry
Member
Yeah, I couldn't find anything related to the SM Design on Parker out there. It does claim native FP16 support, but it wasn't really detailed. I was assuming it was a shrink of the Maxwell design, which would very well be the same thing that happens with the Switch.
The missing piece of the puzzle is lithography. 20nm would be a bit of a letdown, and would limit the clock speeds quite a bit in comparison to what's been quoted for Parker.
I'm talking about Tegra Maxwell to Tegra Pascal, nothing to do with desktop parts. Tegra Maxwell and Tegra Pascal are the same as far as SM design goes, in fact the same in almost every way. The node is different and they've improved colour compression, thats really the only significant differences. But the node could also be made the same with Maxwell, they just had no reason to because they already had a nearly identical chip called Pascal on that node.
- Status