terafo

10 ms·

61.
by terafo 3y ago
>already reduces sizes pretty effectively 8k texture with mipmaps takes about 100mb with BC7. Which is huge. Should be around 30-40mb with kraken and oodle texture. > I do wonder how much of it is that a Kraken license costs money
62.
by terafo 3y ago
Hitman would be very notable example. Fortnite almost got to 100 gigs when they enabled some kind of asset compression.
63.
by terafo 3y ago
It could've been handled as DLC that downloads with a game by default and can be uninstalled/disabled, if we are talking low-effort solutions.
64.
by terafo 3y ago
>immediately have to caveat Yeah, I could've indicated better that I'm talking about on-disk compression of files. >Sony included dedicated compression support in the PS5, licensed from a leading middleware company, speci
65.
by terafo 3y ago
By texture compression I meant something like LZ4 or Kraken on top of block texture compression, sorry for ambiguity. Of course they use block level compression. But it seems that it's the only thing they use, as I said in my top level
66.
by terafo 3y ago
> I doubt you can do it in 8K at that speed All inference of Stable Diffusion is done in FP16 at best, with multi-step models, and models themselves are highly inefficient. I would be surprised if in May 2024 4090 isn't capable o
67.
by terafo 3y ago
Because PNG isn't a good format for GPU textures. Oodle texture+Kraken+something like BC7 is an optimal solution and is used in many games.
68.
by terafo 3y ago
Jedi Survivor having uncompressed textures is a speculation, but I'm extremely confident that it is this way. The fact that games nowadays ship without LOD systems, without texture compression, without adequate streaming is not just sp
69.
by terafo 3y ago
8k texture with mipmaps takes around 100 megs uncompressed. Downloading it would be 8-10 seconds on 100 mbit connection. It is definitely possible to generate that image in 8 to 10 seconds on consumer GPU if you are using some kind of consi
70.
by terafo 3y ago
You can add lossless compression, like Kraken, on top of that. PS5 has hardware accelerated Kraken decompression. And you can encode textures for GPU in a way that is better compressable with lossless algorithms.
71.
by terafo 3y ago
And there was a hilarious situation with audio in Titanfall, which included 35 gigs of uncompressed audio(whole game took 48 gigs).
72.
by terafo 3y ago
> and if it weren't for the RTX graphics cards setting the precedent, there would probably be more research into making things prettier without having to "brute force" things. Not really. A lot of this research included
73.
by terafo 3y ago
And, besides, his solutions aren't really good. Radical solution would be shipping image generating model with the game and generating textures on the fly. That's how all texture-related data goes from being measured in hundreds o
74.
by terafo 3y ago
He completely misses the point. Games don't take hundreds of gigs because they have sub-optimal compression. They take hundreds of gigs because they don't use ANY compression. Jedi Survivor takes 139 gigs on Series X and mere 44 g
75.
by terafo 4y ago
I'd assume their rendering enigne is written for single API, such as OpenGL or Vulkan. Porting it to metal would be quite large project.
76.
by terafo 4y ago
I confused LLaMa vocabulary size, which is indeed 32k, with GPT-4 vocab size. Still, my point stands. You can add those characters there with miniscule cost.
77.
by terafo 4y ago
Nope, it's not about dataset. It's just bad tokenizer. Korean has couple of dozen of symbols in it's alphabet. Cyrillic languages have less than 50 symbols in total. Hiragana is 46 symbols. GPT-4 has 32k tokens IIRC. Includin
78.
by terafo 4y ago
Renting isn't purchasing.
79.
by terafo 4y ago
I was sleepy, I meant to say that it WASN'T trained on data of Meta users.
80.
by terafo 4y ago
It is a good candidate. Tech is good 6-18 months away, though.
81.
by terafo 4y ago
We are talking groups, not individuals. I think it is quite possible for couple of hundreds of people to cooperate and train something at least as big as LLaMa 7B in a week or two.
82.
by terafo 4y ago
LLaMa was trained on data of Meta users, though.
83.
by terafo 4y ago
I'd assume that 33B model should fit with this(only repo that I know of that implements SparseGPT and GPTQ for LLaMa), I, personally, haven't tried though. But you can try your luck https://github.com/lachlansnef
84.
by terafo 4y ago
With all my dislike to Electron, I struggle to remember even one Electron app that managed to use 6 gigs.
85.
by terafo 4y ago
It is as good as GPT-3 at most sizes. Instruct layer needs to be put on top in order for it to compete with GPT 3.5(which powers ChatGPT). It can be done with comparatively little amount of compute(couple hundred bucks worth of compute for
86.
by terafo 4y ago
AI training has very high minimum requirements to get in the door. If your GPU has 12GB of VRAM and your model and gradients require 13GB, you can't train the model. CPUs don't have this limitation but they are ridiculously ineff
87.
by terafo 4y ago
I wonder where such difference between llama.cpp and [1] repo comes from. F16 difference in perplexity is .3 on 7B model, which is not insignificant. ggml quirks are definitely need to be fixed. [1] https://github.com/qwopqw
88.
by terafo 4y ago
Read readme in repo.
89.
by terafo 4y ago
Wish you all luck in the world. We need much more clarity in legal status of these models.
90.
by terafo 4y ago
For this specific implementation here's info from llama.cpp repo: Perplexity - model options 5.5985 - 13B, q4_0 5.9565 - 7B, f16 6.3001 - 7B, q4_1 6.5949 - 7B, q4_0 6.5995 - 7B, q4_0, --memory_f16 According to this repo[1] difference

More ›