NVIDIA and audio/video AI GEN
NVIDIA and audio/video AI GEN
Mg. en Ing. Carlos Gerardo Said
Nvidia has benefitted the most from the AI wave so far. So much so, in fact, that the U.S. Department of Justice is investigating complaints that the chipmaker is engaging in anticompetitive behavior.
The company’s chips do apparently have a weakness when it comes to a growing field of AI research: video generation.
The main issue is that Nvidia’s graphics processing units, even its popular H100s, don’t have enough memory to handle the immense amount of video data needed to train these models.
"Training data [for video generation models] can scale into petabytes pretty easily, as opposed to a few hundred terabytes for language-only models,” said a cofounder at video generation startup Luma AI.
To give you a sense of the magnitude of that—a petabyte is equivalent to a whopping 1.000 terabytes, or around 500.000 hours of video.
Another bottleneck for AI video startups is when GPUs don’t transfer data between each other fast enough. That’s important because video data is significantly larger than language or image data, so it’s even more important that it can be transferred between chips quickly.
The difficulties of training and running video generation models with today’s GPUs have likely held back the progress of the AI video market. OpenAI’s Sora, was teased nearly six months ago, and we can´t see yet to see any sign of a public release.
When Luma released its video generating product the flood of users overwhelmed its systems, leading to the website going down for much of the day. That forced the Luma team to add thousands of GPUs in days from its three to four different providers—not an easy engineering task.
Nvidia GPUs are pretty much the best option available today to train video models with.
Companies are testing out Nvidia’s next flagship AI chip, Blackwell, whose significantly larger memory will likely address some of the issues that come along with training video generation models.
But there’s also an opportunity for rivals, namely Advanced Micro Devices. Its MI300X chips already have more than twice as much memory as Nvidia’s H100s.
Comments
Post a Comment