I'm definitely making some assumptions that the Switch firmware is probably in the > 100MB range (would be good to get some hard numbers on that) and all the units are the same. Any reprints that include patches will be unambiguously over the threshold, though, as the day one patch pushes it to 16.2GB. It's really hard to say where the exact threshold actually is, but this game is close enough that it's probably above it.
Did Smash launch on a 32GB card? Because if not, it crossed the threshold and must have transitioned to 32GB after the first run.
Smash's initial cart image was 13.6 GB, probably below the threshold. The full size including updates climbed above 16GB at some point, but I couldn't tell you what version they're printing on the carts now.
Just checked, the only documented cart revision I can find was only at 2.0.1, but this database is probably not exhaustive. It's possible they haven't spun up a new cart revision since it crossed the threshold.
www.doesitplay.org
It would be near disaster level for it not to have a stronger launch than BOTW. I don't remember the exact number but I wanna say it shipped like 3M in the first week?
The attach Ratio for BOTW was greater than the consoles sold at launch iirc.
I don't know if it was attach rate or just straight-up copies sold, but yeah. I was under the impression BotW sales numbers exceeded Switch numbers because basically everyone who bought a Switch bought BotW, and then also many people who owned WiiUs bought it.
so is a 32GB cartdridge?
most likely
It's a little hard to game the numbers out, as sales charts only report physical sales, but Nintendo will sometimes drop press releases about total sales for big games. But Scarlet/Violet did 10 million in its first 3 days, according to Nintendo. If I had to make a guesstimate, that represents a ceiling on the speed of sales for Switch games, which are somewhat throttled by Nintendo's ability to physically create and ship carts.
Perhaps a die shrink of T239 in the absolute best case scenario.
The tweet is referring to Blackwell/RTX 50XX, not T239. Could be a potential target for a die shrink later down the line if Nvidia's using it for something else, though.
Complete guess on my part: target 540p handheld, 720p docked.
Yeah it’s too quiet. I don’t see it coming this year.
Could be a compromise due to rising costs and inflation, and wanting to prioritize soc and ram.
videocardz.com
No doubt it will start much bigger, with like 8x as many systems out there as there were by the end of 2017. Matching/exceeding the eventual total of a previous evergreen, though, always hard to say.
The compression happens at production time, but the benefits extend to runtime since it allows higher resolution textures using less RAM. The earlier link said this new method took more processing time, but I have no idea what kind of differences there would be considered notable or negligible when it comes to textures.
I know, I was the first one to post the paper. But that compression in real time doesn't seem beneficial when you you have to store that 256MB source. That's the biggest benefit here, cutting down the sizes so you can save on space. That in turn saves on ram usage, which you can compress further if needed. It takes more time to compress from the high quality, 4K master, but from an already compressed, 2K source, it'll be much faster (quality is unknown though)
Does this work well on ampere?
Doesn't seem beneficial, agreed, but why would that be a consideration anyway? Textures aren't being stored uncompressed today, so it's not a change.
To me being able to get higher resolution texture output without needing more RAM/bandwidth seems like a bigger deal. Even for a game installing to 200GB on an arbitrary medium, what the GPU could actually display is a bottleneck. And with RAM increasing much slower than it used to, finding other ways to work around that bottleneck seems like a great thing.
they only tested it on a 4090, so it'll probably be much slower on ampere
this is just to get more quality out of a given file size. so this whole thing isn't really novel, but a better way to do what's already being done
)
If it's doing it in a different way that produces much better results, that seems pretty novel to me. And it's not just getting more quality out of a given file size, but getting more quality out of a given VRAM size.
In theory it should work as well on Ampere as on Ada, but there's going to be a significant performance cost either way, as you're moving from work that's done basically for free on texture units (BC decompression) to the tensor cores.
So rendering a 4K image with trilinear BC7 textures takes 0.49ms on an RTX 4090, and NTC bumps that to between 1.15ms and 1.92ms. I'd be very surprised if the BC7 case is actually decompression limited at all (the RTX 4090's texture units have a throughput of 1.3 trillion texels a second), so let's assume the actual rendering takes 0.49ms, which means the neural texture decompression takes between 0.66ms and 1.43ms. The RTX 4090 tensor cores are capable of around 330 Tflops FP16, whereas on T239 the absolute best case scenario is around 15 Tflops of FP16 tensor core performance. So assuming that it's 100% limited by tensor core throughput, on T239 you'd be looking at around 14.52ms to 31.42ms for the same task. Even at 1080p, that would imply 3.63ms to 7.86ms just for texture decompression. Of course this is a very crude measurement, but that's quite a high cost when BC/ASTC decompression is basically free.
If you look in table 5 in the paper, they claim a PSNR of 28.96 and a bits per pixel of 0.55 if the method is limited to four channels, which is the maximum supported by any current block format. Up in table 3, when comparing different compression algorithms with the same dataset, they show that ASTC 12×12 (which, FYI, Switch supports) achieves a PSNR of 28.90 with a bits per pixel of 0.43. So limited to the same channel count of existing algorithms, ASTC can beat them in terms of compression for a similar PSNR. That's not to say that their method wouldn't win out over a many-channel block format, but I strongly suspect that applying more traditional approaches to higher channel count textures could at least get close in performance to NTC while having the very significant benefit of being decompressible in hardware within texture units.
This ls very funny and describes me as well whenever I read tech heavy posts, ahahahhahahahame reading every other Thraktor post
So, just for funsies, how about a little explainer? This loops back to the "physically based rendering" conversation from the Metroid Prime remaster.
Can engine channels be compared to Photoshop Layers? Like, we see Tears of the Kingdom ‘s box art, but there’s a lot going on that’s deeper than a single image: it’s layers of placement of Link, edited lighting here, contrast there, etc
While I don't think it's likely, I'd love to see it. I think they'll just drop Home out of the blue with a date for more info on the Teal Mask. Then, new console reveal, then a week or two later, Teal Mask trailer, updated release window for Indigo Disk, "with enhancements for [REDACTED]".
You mean the person that leaked the Present’s content and also commented on a next gen patch for S/V?
Hope that was from an earlier build. Then again, the game was five months out when that trailer was shown, so who knows.Re: textures.
Pikmin 4's bulborbs are less detailed than Pikmin 3, with no normal map to give their spots depth.
That's the one, though they also leaked the existence of Tera forms (backed up by the most reliable Pokemon leaker) and that was nowhere in the presentation
Ooh, interesting theory. I think this could be possible.
nope. outside of games that are poised to get dlc, like Zelda, Pokemon, Splatoon, I don't foresee too many patches. from third parties, maybe if the game is a GaaS, like Apex, Fortnite, and Overwatch
8nm would make the Drake design very difficult to use. it's big, consumes a good deal of power, and reigning that in is pointless when it was cheaper to just make a smaller chip.
Same reason as it was back when the Nvidia leak happened. Power consumption.
And Nvidia's capacity for 4N>> their capacity for 5LPP/LPE, the only other suitable node for the size of the processor at the same power consumption as Nintendo Switch (HAC-001)(V1)
