With Nvidia Ace taking up 1 GB of VRAM in Inzoi, Team Green will need to up its memory game if AI NPCs take off in PC gaming

At this year’s GDC event, Nvidia showcased all its latest RTX technologies, including Ace, its software suite of ‘digital human technologies.’ One of the first games to use it is Inzoi, a Sims-like game, and while chatting to Nvidia about it, I learned that the AI model takes up a surprising amount of VRAM, which raises an interesting question about how much memory future GeForce cards are going to have.

The implementation of Nvidia Ace in Inzoi (or inZOI, to use the correct title) is relatively low-key. The family you control, along with background NPCs, all display ‘thought bubbles’ which give you clues as to how they’re feeling, what their plans are, and what they’re considering doing in the future. Enabling Nvidia’s AI system gives you a bit more control over their thoughts, as well as making them adapt and respond to changes around them more realistically.

In terms of technical details, the AI system used is a Mistral NeMo Minitron small language model, with 500 million parameters. The developers had experimented with larger models but settled on this size, as it gave the best balance between responsiveness, accuracy, and most importantly of all, performance. Larger models use more GPU resources to process and in this specific case, Inzoi uses 1 GB of VRAM to store the model.

That may not seem like very much, but this is a small model with some clear limitations. For example, it doesn’t get applied to every NPC, just those within visible range and it won’t result in any major transformations to a character’s life. The smaller the language model, the less accurate it is, and it has the potential to hallucinate more (i.e. produces results that aren’t in training data).

While Inzoi’s AI system isn’t all that impressive, what I saw in action at the GDC made me think that Nvidia’s Ace has huge potential for other genres, particularly large, open-world RPGs. Alternatives already exist as mods for certain games, such as Mantella for Skyrim, and it transforms the dull, repetitive nature of NPC ‘conversations’ into something far more realistic and immersive.

To transform such games into ‘living, breathing worlds,’ much larger models will be required and traditionally, this involves a cloud-based system. However, a local model would be far preferable for most PC gamers worldwide, which brings us to the topic of VRAM.

Nvidia has been offering 8 GB of memory on its mainstream graphics cards for years, and other than the glitch in the matrix that is the RTX 3060, it doesn’t seem to want to change this any time soon. Intel and AMD have been doing the same, of course, but where 16 GB of VRAM is the preserve of Nvidia’s high-end GPUs, such as the RTX 5070 Ti and RTX 5080, one can get that amount of memory on far cheaper Arc and Radeon cards.

But if Nvidia Ace really takes off and developers start to complain that they’re being restricted about what they can achieve with the software suite, because of the size of the SLM (small language model) they’re having to use, then the jolly green giant will have to respond by upping the amount of VRAM it offers across the board.

After all, other aspects of PCs/computers have needed to increase the minimum amount of memory they sport because of AI, such as Apple with its base Mac spec and Lunar Lake laptops having 16 GB because of Copilot.

It’s not often that one can say AI is doing something really useful for gamers but in this case, I think Nvidia’s Ace and competing systems may well be what pushes graphics cards to consign 8 GB of VRAM to history. Not textures, ray tracing, or frame generation but realistic NPC responses. Progress never quite goes in the direction you always expect it to.