ModEl4In agree that 12GB is a problem, I haven't seen usage in Indiana with path tracing at FHD base resolution to check it and also 5070 will support neural texture compression enabling similar quality textures with smaller memory footprint (or higher quality textures in a given memory budget) but what you said even if we suppose that it isn't true today (FHD base...) certainly will be in the near future.

10tothemin9voltsAnd tomorrow is here, see my Indiana Jones example, and I don't think the neural texture compression -- it was like 0.4 - 0.5 GB less VRAM demand, according to a vid -- is going to help with that bc the VRAM utilization in that IJ scenario rises to ~15GB. Ofc not many exmples, but in the future more game gonna demand similar VRAM requiremetns (16GB VRAM for 1440p with even the lowest pathtracing setting), but better than nothing and it's free for all gens(?). Ofc, need to wait for independent reviews whether there's any catch with that.
The transformer based upscaling/DLSS is also free for all gens, that's also very nice (waiting for independent reviews whether theres any catch with that too).

For all AI LLM selfhosters enthusiasts, the GeForce 5090' 512-bit bus width theoretically means NV could release a clamshell RTX 6000 Blackwell 64GB VRAM workstation card (in a few months), like they usually do (384-bit: GeForce 4090 24 GB VRAM and "RTX 6000 Ada" 48GB VRAM).

• RTX Neural Texture Compression uses AI to compress thousands of textures in less than a minute. Their neural representations are stored or accessed in real time or loaded directly into memory without further modification. The neurally compressed textures save up to 7x more VRAM or system memory than traditional block compressed textures at the same visual quality.

10tothemin9voltsFor all AI LLM selfhosters enthusiasts, the GeForce 5090' 512-bit bus width theoretically means NV could release a clamshell RTX 6000 Blackwell 64GB VRAM workstation card (in a few months), like they usually do (384-bit: GeForce 4090 24 GB VRAM and "RTX 6000 Ada" 48GB VRAM).

ModEl4It's not only 0,4-0,5GB, but we will see in the near future (I hope) with actual implementations the memory requirement difference

• RTX Neural Texture Compression uses AI to compress thousands of textures in less than a minute. Their neural representations are stored or accessed in real time or loaded directly into memory without further modification. The neurally compressed textures save up to 7x more VRAM or system memory than traditional block compressed textures at the same visual quality.

igormpDon't forget the 3GB modules that are coming really soon. That would mean a 48GB model without clamshell, and a 96GB with.

notoperableEvery GPU ad I've seen from CES so far sounds like a marketing manager on Ritalin overdose

dragontamer5788GPU Makers know that AI isn't important, but it also doesn't "cost" very much to add FP16 or FP4 (or whatever) matrix-multiplication units to your GPUs.

Seriously, 4-bit numbers? Okay NVidia. I see what you're doing.

• The first product in the AMD Instinct MI350 Series, the AMD Instinct MI350X accelerator, is based on the AMD CDNA 4 architecture and is expected to be available in 2025. It will use the same industry standard Universal Baseboard server design as other MI300 Series accelerators and will be built using advanced 3nm process technology, support the FP4 and FP6 AI datatypes and have up to 288 GB of HBM3E memory.

Visible NoiseMaybe a little googling before posting will be helpful.

Visible NoiseIt’s FP4. Floating point. A wee bit more complicated than you make it out to be. Perhaps you should look it up. Here’s a starting point papers.nips.cc/paper/2020/file/13b919438259814cd5be8cb45877d577-Paper.pdf

dragontamer5788Oh right. Good point. Its 1-bit sign, 3-bit exponent, 0-bit mantissa. So "multipliers" are actually 3-bit adders.

Thanks for making me look it up. Its n worse than I expected. ("Multiplication" of exponents simply becomes addition. 2^4 * 2^3 == 2^7).

Visible Noiseif that’s what you got out of the IBM research paper then you’re too blinded by hate to understand. Upset that your gaming GPU are expensive because they found other users for then I presume.

dragontamer5788Dude, its elementary to design a 1-bit sign / 3-bit exponent / 0-bit mantissa multiplier. Seriously, look at the damn format and think about how the circuit would be made. This is the stuff a 2nd year computer engineer can do.

If you aren't into computer engineering, then I promise you, its not very complex. The hard stuff are the larger multiplier circuits (aka: Wallace Trees and whatnot), which are closer to 4th year or even masters-degree level in my experience.

Think like an NVidia or AMD engineer. Think about how to wire adders together so that it'd make a multiplication circuit. Some of these things are absurdly easy to do.

Visible Noiseif that’s what you got out of the IBM research paper then you’re too blinded by hate to understand. Upset that your gaming GPU are expensive because they found other users for then I presume.

igormpIt is awful for sure. But I the same time, I don't feel it's like Nvidia has been bribing companies to not use their competitors.