Discrete graphics cards today often launch with memory capacities of at least 8 GB. In fact, you'll only see 8 GB with the cheapest Nvidia GeForce RTX 5050 and 5060 series. That said, although we're seeing the absolute minimum at 8 GB, I'm here to argue that we should have 24 GB of video memory allocated as standard, at least for mid-tier cards and above. Consider this: the RTX 5060 is marketed as a mid-range GPU, but you can purchase one with just 8 GB of RAM. That's not a lot, especially when running ray tracing and/or super-sampling.

We've all been there, running out of VRAM with your favorite game, video timelines that crawl to a halt in editor software, and AI workloads that seem to take an eternity to complete. The home lab is expanding, and everyone wants a piece of the action. PC gaming is becoming increasingly advanced with new effects and technologies that leverage all parts of the GPU. While it's great to see new-generation ray tracing and shader cores, VRAM is often overlooked, and we've seen the drawbacks of not having enough local memory for the GPU.

The aforementioned RTX 5060 can be bought with just 8GB, while the RTX 5060 Ti can be found with 16 GB, 4 GB more than the more expensive (and powerful) RTX 5070.

VRAM is vital to use the GPU

You always need VRAM

Gone are the arguments that "only enthusiasts require a GPU with 24 GB," since many games now pump lots of data into the component for outputting to the connected screen. We've seen the outcry from the community regarding the RTX 50 series and its seemingly conservative VRAM specs, with many believing the 5080 is underpowered with its meager 16 GB. 16 GB is still plenty enough for running games at 1080p and 1440p with settings cranked high and even some super sampling and ray tracing for good measure, but you're almost playing with fire.

VRAM matters because while it may not be as flashy as clock speeds and the number of cores for crunching numbers, the memory installed on the GPU essentially dictates what the card can (and more importantly, what it can't) do. Think of it like RAM for your PC. If you run out of system memory, the SSD or HDD will have to suffice. The same goes for the GPU. Out of VRAM? Looks like you'll either encounter some issues or force the system to rely on RAM. System memory on the motherboard is rapid compared to even NVMe SSDs, but it's still nowhere close to GPU VRAM.

Relying on RAM to help out with its slower bus will raise the ugly heads of asset pop-in when trees, characters, and other objects magically appear in front of you, if not game stuttering and complete system crashes. For low-range GPUs, 16 GB is a good figure to work with. 8 GB could be viewed as too little too late, with modern games demanding much more and developers sometimes relying on DLSS, FSR, and XeSS to help combat the issue of optimization. It's no secret that games are much more bloated than they once were. Have you tried firing up Star Citizen?

And it's not just lighting effects that will be taking up all your graphics card's memory. Modern games often have high-resolution textures, larger buffers for geometry, and other content that needs to be stored somewhere. Upscalers and even multi-frame generation only look to increase VRAM usage. 16 GB can prove too low to store everything, especially when moving up to 1440p and high-refresh-rate monitors. 4K and ray tracing would tip you well and truly overboard, depending on the game, of course. But you'd be forgiven for trying these settings on a GeForce RTX 5070 or 5080.

We're getting more creative with AI

More and more of us are starting to self-host services and content from home. Running Proxmox, Home Assistant, Jellyfin, among other containers and software, has never been easier and more accessible, thanks to the explosion in popularity of network-attached storage (NAS). Now, just about anyone can store their purchased movies, shows, and music and start streaming them to any device, anywhere in the world. The same goes for large language models (LLMs), similar to ChatGPT. You can run your own models using Llama and OpenWeb UI, but AI inferencing requires VRAM ... and lots of it.

It's why we're seeing such a supply issue with memory and other related parts since data centers are effectively scooping everything up, and Nvidia (and other parties) doesn't have an issue where its profits are coming from. AI is being used more and more, and I don't see this changing as companies look at some rather creative ways to work LLMs and other instances into our daily lives. I used to be a strong proponent against using LLMs and other AI-related features, but I'm now running my own models, have integrated them into my smart home, and rely on them for basic tasks.

Powering this is my server's GPU. Luckily, my RTX 4070 Ti has 16 GB of VRAM, which is enough to run a competent model. The only problem is that it's not great for running this model and handling Frigate object detection. That requires a separate GPU, in the form of an 8 GB RTX 4060 ― noticed how the RTX 5060 has the same VRAM capacity as its predecessor?

24 GB is great for modern GPUs

The more the merrier

Competitive gaming is often touted as the perfect reason for sticking to 1080p, setting the refresh rate to some impressively high figures. But this wouldn't be possible on a GeForce RTX 5050, and you won't see professional streamers or gamers work with these entry-level GPUs because not only do the cores and other parts of the GPU matter, but also the VRAM. Even if it's a largely basic game in that it doesn't take up a lot of VRAM or general resources, having that overhead ensures you'll always have enough at hand to avoid any of those GPU-related issues I covered above.

I ran some games at 1440p on an RTX 4070 Ti with 16 GB of RAM. It ran them well, but once cranking up settings, enabling ray tracing, and some other advanced settings, the GPU started to bog down with demand. That's when switching to an AMD Radeon RX 7900 XTX with 24 GB of VRAM completely transformed everything. Gaming was noticeably smoother at the same resolution. Sure, it's a better card overall, but I also didn't encounter any VRAM bottlenecking. The same goes for AI inferencing with OpenWeb UI.

Sure, you won't require 24 GB of VRAM to enjoy your favorite game at 1080p, but having the headroom available simply makes it possible to fire up all settings without a care in the world. The same goes for game optimization. Yes, we'd love for developers not to rely on DLSS and FSR to help out with performance, and largely they don't. But game engines are becoming increasingly more complex and require more system resources to run properly. And no, cloud gaming is not an answer for those of us who much prefer to enjoy a locally run game on the desktop or couch.

Let me be clear about my original statement, too. "24GB VRAM should be the new minimum for most GPUs." I'm not suggesting for one minute that a GeForce RTX 6050 should come rocking 32 GB of RAM, but Nvidia, AMD, and Intel need to ensure enough is present on the GPU to allow gamers to enjoy their favorite titles without needing to turn settings down due to memory allocation (or lack thereof). Yes, low-end GPUs for 1080p and eSports can come with 12 GB or maybe even 16 GB. Mid-range cards need 24 GB, and high-end cards can go for 32 GB and beyond.