What just happened? After AI's massive success in boosting framerates, it's now getting ready for the next big thing in gaming – powering conversations. Nvidia is preparing to roll out a new artificial intelligence system called ACE that will allow players to interact with NPCs using plain speech.
The first title set to feature ACE is Mecha Break, an upcoming online mech battle game releasing on the PC, Xbox Series X/S, and PlayStation 5 in 2025. The game will let players pick from different classes of mechs, such as assault, melee, and sniper, to duke it out in battles that take place on both land and in the air.
The real innovation is ACE, which will power an AI-controlled NPC that can engage in back-and-forth conversations by understanding your voice commands. Nvidia says ACE can "see" you and identify objects around you through your webcam using GPT-4o language models. It combines on-device AI processing with some cloud processing to generate the NPC's voices.
In the demo video, the player asked the NPC a few details about their mech and mission they'd otherwise be Googling for, picked a mech, changed its paint, and launched the mission – all by using nothing but their voice. However, the responses still have that robotic quality, which is a shame considering OpenAI's models have gotten ridiculously good at imitating humans.
Unfortunately, conversation quality and quick response times happen to be inversely related, and one can't have more of both at the same time. The latter is especially important in gaming, where you certainly wouldn't want to wait several seconds to squeeze basic conversation out of characters. From the demo above, it appears the AI NPC is focused on performance over naturalistic conversation.
That seems deliberate considering that a good chunk of the AI processing for ACE happens locally on your graphics card rather than relying solely on cloud compute. Nvidia has optimized a pared-down 4B version of its language models to run on RTX GPUs all the way down to the entry-level RTX 2060, taking up just 2GB of VRAM.
That miniaturized model descends from Nvidia's beefy, cloud-based Nemotron model, which requires four A100 GPUs and well over 100GB of VRAM to operate. It's not as robust as its cloud counterpart, but the 4B version can still provide snappy AI-generated responses in around 300 milliseconds.
Mecha Break is just a start, and ACE's implementation is fairly limited, restricted to a single NPC character. We'll have to wait and see if future games build this tech more comprehensively into core mechanics and storytelling. There are also unanswered questions around how ACE will work for gamers without Nvidia GPUs – whether it will default to cloud processing, use non-AI scripts, or even potentially be unavailable.