VRAM

you can't really use it.

Even 96 gigabytes of VRAM is kind of hard, even if you go to 128 gigabytes or the highest models I'm operating are running at 256 gigabytes of VRAM and then 50 gigabytes more for like KV caches.

Like these models, for production grade systems, it is questionable.

I mean, I do a lot of work on my main profession and I'm doing a lot of AI stuff there and things.

Previously you had to enter power save for every core, now you can select some cores to enter power save and some to still run at full tilt.

Older AMD GPUs will also better recover from hangs or crashes And you also get that really nice fix to improve how AMD GPUs that run out of VRAM will behave, meaning you can play some games that require more VRAM than you have and still run them at decent performance.

For example, the latest Indiana Jones games has been tested and can run well when requiring 9GB of VRAM if you only have 8.

This stuff was already in SteamOS, but it is now in the kernel as a whole, so it will help every Linux gamer.

Well, at least if you use an AMD GPU.

But yeah, he was saying the latest and greatest combos, which... goes either way, but he didn't make an argument for either one of those.

But I get where he's coming from with that, but also it's like, no, you want to see what the GPU can actually do when not constrained by another component within the system and then the overhead testing is additional testing it's like saying how much vram a graphics card or a gpu needs which you know can depend on the variables at play But also testing VRAM usage and how that actually impacts visuals and performance is such an in-depth investigation that it's not practical to include that in a video.

It's something you do afterwards and then note it in following reviews of GPUs.

It's just so difficult to test VRAM usage accurately, which is why so many outlets have struggled with it in the past.

You've got to play the game.

And it's the exact same reason why I was hesitant to jump headfirst straight into this shader compilation testing, because it is such a difficult thing to test.

And then you've got to play through the game and work out what's shader compilation stutter, what's traversal stutter, and then record that and compare it to another.

It's a whole lot of work.

So yeah, the RTX 6000 Pro, I mean...

Ada to Blackwell, gin on gin, doubling the VRAM, adding 422 in-code decode.

And it's a great card because it rivals like an H100 kind of, I mean, not in terms of speed, right? But like in terms of almost like VRAM wise, but it just, It does everything, and it's really designed for – it's interesting you say that, that you brought that up, because that card is really professional.

It's not a GeForce, but it's a professional-grade card where that industry used to serve engineering, manufacturing, media and entertainment for a lot of things workstation-wise where they were doing – like, you know, simulations or design or CAD work or whatever.

And then now kind of with the advent of AI, there are some things that just straight up need tensor cores and they just go.

And there are some very important updates for gamers, hardware enthusiasts, or if you just want to make your system run faster.

If you are gaming on an AMD graphics card with lower VRAM, like an 8GB card, Linux 7.3 is honestly literally a game changer.

Are you

good enough saying lower VRAM now, though, man? Because they're releasing brand new cards right now that don't even have that much.

Well, just a few years ago, that was considered low.

So, but yes, yeah.

They're like, I wish I had 8 gigs.

Well, that's why Linux kernel is focusing on these eight gig cards because there's a lot of new ones coming out and people still have a lot of old ones because of the GPU prices and RAM apocalypse and all the things.