NVIDIA has been exploring AI and datacenters lately; see what they’re up to in these recordings of their recent NVIDIA AI presentation and the rest of their Computex showcase.
Recently, at Computex Taipei 2023 which was held May 30th – June 2nd of this year, NVIDIA presented some AI related predictive generation tool based explorations they have been working on, as well as some developments in datacenter machines that could reduce the cost and size needed to produce a high workload throughput. These are some interesting developments for folks who use high-powered machines and GPU datacenters for work and industrial tasks.
Check out these videos of some of what they presented during the event: a text-to-song generative tool, as well as improvements in GPU datacenter machines that they say will greatly improve efficiency and result in higher throughput under a given power limit.
NVIDIA AI text-to-song generation
As an example of NVIDIA AI exploration, NVIDIA used an AI tool from Voicemod to put some lyrics to a tune and created a short song, which they invited the audience to do a karaoke style sing-along to during their presentation.
NVIDIA GPU datacenter machines
NVIDIA is hoping to meet the increasing demand for the services that GPU servers and datacenters can provide, and get around the problem of power limitation, as well as cost, by developing more dense, faster, and cost-effective datacenter machines. Almost all datacenters face the issue of power limitation, so NVIDIA has decided to dedicate their teams to developing a way to achieve more work under the same given power limitations.
Increasing the density of the computing power for GPU datacenter machines lowers the number of and size of computers necessary to complete a task, and thus the necessary power draw, and as a result also the cost required to sustain the same work output. This also raises the potential performance throughput ceiling for a datacenter as it will be able to achieve more work under the same power limitation. These improvements also result in a smaller physical machine, which is more conveneint when it comes to storage for a datacenter.
Watch NVIDIA GPU datacenter machines presentation clip
NVIDIA also brought a physical example of one of these improved datacenter machines to the presentation, so you can see just how much less mass can produce the same amount of performance as massive datacenters with many large machines have in the past. The machine costs $200,000 and packs a lot of power, despite its small size.
It can replace an entire room of less efficient server machines, which can also help save on costs to store this equipment, as a smaller room can house much more power in it using denser datacenter machines. It offers advantages from being much smaller in footprint, but it is also interesting to note that due to its density, even the individual compute trays of this datacenter machine are actually very heavy, as well as powerful.