Hacker News Reader

Comment by elabajaba | Hacker News Reader

The 4080 12GB's performance is about the same as where a new 60 series card sits compared to the previous gen 80 series card (60 series tend to be within +-10% of the previous gen 80 series, 4080 12GB is at most a 60ti based on this)

The 4080 16GB is around where a 70 series card tends to sit.

[0] https://www.overclock3d.net/news/gpu_displays/nvidia_release... Ignore the dlss2 vs dlss3 numbers, just look at the base numbers since dlss3 numbers are interpolated framerates

View on HN

Yes, a youtuber who goes by ‘I’m a mac’ made a great (at least I thought) series of videos breaking down how this generations of die sizes compare to previous generations: (The first where he starts comparing generations with graphs) https://youtu.be/9_bqwEy5AQ4

Though I have to wonder, do we want bigger 40 gen dies? The 4090 already guzzles at 400w at full tilt, and gets ar least 200+ fps at pretty much all games except maybe cyberpunk, and can complete video editing workloads at twice the speed of the 3090.

Do we need an even bigger lovelace chip?

Considering the 4090 gets 95% of the performance in games if you set a 60% powertarget [0](~90% in synthetics), I feel like they just juiced the power draw like crazy to be able to say they had the fastest gpu this generation (same with AMD's Ryzen 7000 CPUs, where you can drop from ~240w to 140w for 95% of the multicore/100% of the single core performance).

Also I was playing the Darktide beta, and at 4k I want about double the framerate I currently get on a 6800xt (FSR 1.0 and DLSS (according to friends and comments on reddit+discord) both look bad in it). We're moving past crossgen games at this point too (and UE5 seems very heavy), so for eg. 100+ fps or 4k something in range of ~50% faster than a 3090 would be nice.

[0] https://youtu.be/60yFji_GKak?t=876

The power draw of the 4090 is less than the 3090Ti and AMD 6950XT…

My guess is that AI workloads are going to be a significant factor of demand for high-end cards. It used to be gaming and mining and now we have AI art generation that requires high-end cards to be productive.

Now mere mortals can play around with Stable Diffusion and other art-generation technologies, this is going to increase demand for these types of cards.

Folk are already using lower end cards with Stable Diffusion, even the 4GB models.

My 3080 is well capable of generating 512&768 images with a high step count in under a minute, and 500 odd steps (just for fun) in around 2 minutes (or even less with different samplers)

The interesting bit is when we get more VRAM (24-32GB) and can actually train the models with off the shelf hardware without breaking the bank.

And soon, it won't just be the people who are "playing around" with AI models on their home machine. It will be games and commercial products using ML at the client level that will have system requirements around GPUs with a lot of memory for the better results. A lot of this can be done at the cloud level, but if you're clocking heavy hours on cloud GPU, that subscription model is going to be very expensive and will introduce latency if a lot of data needs to be exchanged. Since you can't (by the license agreement on NVidia's drivers) just throw a bunch of consumer-grade hardware in a datacenter to do ML loads like you can for mining, the price shoots up if you want to do the heavy lifting on a central server.

Reply on news.ycombinator.com