On the eve of Computex 2024 [url=]http://www.nvidia.com]NVIDIA[/url] CEO Jensen Huang took to the stage for the now traditional pre-expo Keynote address. He arrived on the back of an astounding year for the company driven by unprecedented growth in AI-related industries, making it one of a select few $1tn companies; as you might have expected, consumer GPU discussion wasn't exactly at the forefront of the presentation.
In a rather long and rambling address, Huang expounded on a topic now familiar to those who have been following NVIDIA's journey since 2016: the symbiotic development of their hardware and software portfolio into one which could be leveraged for ever more complicated neural networking tasks. Touching on innovations such as the transformer for Large Language Models, generative AI, digital twins and the now hundreds of pre-trained AI models available through CUDA, Huang sketched out why NVIDIA is in such a dominant position as we potentially reach the threshold of a computing paradigm. At least, that's the plan.
All good software needs good hardware and Jensen once again showed off Blackwell, their upcoming GPU architecture which will replace Hopper and Ada Lovelace in datacentre and consumer silicon respectively later this year. The next DGX supercomputer, based on Blackwell, will incorporate up to 9x more GPUs and have over 18x the interconnect capacity compared to the Hopper variant, offering 4.5x improved perf/watt in the same form factor. This will also be capable of training and inferencing over even larger datasets (or alternatively, doing either much faster), loosening a key bottleneck for many domains.
NVIDIA envisions that AI will be a factor in the world's entire industrial output, currently estimated at $100tn, in the near future. By providing low-cost Cloud services and deploying pre-trained NIMs to aid in the integration of AI capabilities (including such tools as AI assistants for customer service roles), NVIDIA hope to be a huge part of it.
Just prior to a brief finale demonstrating applications for NVIDIA Omniverse physically-based simulation for training real-world robotics, Jensen Huang also finally revealed the codename for the generation after Blackwell. Rubin GPUs will be paired with Vera CPUs - replacing 2025's Ultra variant of Blackwell and Hopper - and debuting with more HBM4 memory, NV LINK 6, faster switch technology and more.
--
Sadly the keynote proper barely made mention of consumer-level GPU developments, though in wider context that's probably not all that surprising. A flurry of post-Keynote press releases did however detail some of what consumers invested in the GeForce ecosystem can look forward to in the second half of 2024.
Project G-Assist
If you've ever been stumped to what to do next in a game, or just simply some in-game optimisations you could make to improve performance, Project G-Assist might be just what the doctor ordered. Utilising a database of knowledge provided by the game developer, this AI assistant will respond to prompts with important context clues including what's on screen to make recommendations such as where to go next, what weapons to try, or what skills to take. The aim to is decrease the information burden for new players and making learning curves more gradual while not over-simplifying games.
While this sounds great in theory, more complex titles than your average FPS will require pretty immense and well-curated wiki's. For instance, handling in-game queries for the next Call of Duty instalment may be quite straight-forward; Path of Exile however? Just a rudimentary crafting guide would be asking a lot from any AI assistant.
RTX-powered Digital Humans
Utilising AI to generate in-game NPCs has long been a dream for game developers, but with the advent of AI-power digital assistants their time might finally be here. NVIDIA are offering a range of tools to create NPCs that are more realistically visually, verbally and conversationally, better understanding both the user than the world in which they reside. Through NVIDIA NIMs, game will be able to create them in the Cloud and with the help of RTX AI PCs.
Tools such as these based on generative AI will prove to be controversial. They have the potential to effectively replace motion capture and voice actors, and the dataset on which they are trained will come in for plenty of scrutiny too. On the other hand they could significantly streamline game development by automating a very personnel-intensive task; it's a genuine catch-22 for the industry.
RTX Acceleration for Windows Copilot Runtime
Speaking of controversy, Microsoft is bringing small language model (SLM) capabilities to desktop PCs in upcoming Windows previews and NVIDIA are backing them with the tools to run them on local device RTX hardware. NVIDIA's role is currently oriented towards providing the API framework to prospective developers of these new models and the apps on which they will be based. Expect them to become much more widely discussed on the release of future preview builds.
Readers may recall the recent furore over Recall, a Windows Copilot tool that will provide a history of in-OS user actions by analysing sequentially captured screenshots. Well, this is just one of the first-party features NVIDIA are hoping to support through the development partnership with Microsoft.
RTX AI Laptops, SFF GPUs and G-SYNC Displays
Consumer hardware isn't really NVIDIA's focus at Computex this year but the green team did offer a few morsels. Perhaps the most significant long-term will be the reveal of a new range of RTX AI laptops equipped with an RTX 4070 GPU and an as-yet undisclosed CPU with hardware-accelerated AI capabilities as part of its SoC. These laptops launch as part of a drive by Windows Copilot developments, and the CPUs discussed have greater AI capabilities than those released by both Intel and AMD thus far.
To improve options available to mainstream consumers in the desktop space, NVIDIA are choosing to highlight Small Form-Factor Ready GPUs and compatible cases. DIY system builders will be able to confirm compatibility through a 2-step process, easing the decision-making process quite a bit. More info at their Build Small, Play Big article.
And finally, NVIDIA announced a new cache of G-SYNC Compatible displays. Eleven new models are expected to be revealed over the course of Computex, eight of which will be integrated into the June 4th GeForce driver update. The models include OLED panel designs from ACER and plenty of other high refresh rate SKUs from the likes of ASUS, Lenovo, LG and Philips.
--
The general tone of NVIDIA's Computex announcements this year was muted, with an expectation of big things to come that aren't quite ready yet. It's almost certain that we'll see the GeForce RTX 50-series later this year - in the autumn post-Gamescom if previous timings are an indication - but at a datacentre and enterprise level NVIDIA seemed like they were holding their breath for new AI model reveals in the second half of 2024. And Jensen, or his Digital AI Twin, will be there to collect plenty of plaudits for what has been made possible by their technology.
