NVIDIA Blog

Other than TPUs they're also planning for 960,000 Rubin GPUs [1] which can do 33 teraflops fp64 each, so over 30 classical exaflops, and with emulation it could be more than 100 exaflops. [1] https://blogs.nvidia.com/blog/google-cloud-agentic-physical-ai-factori…

I recommend reading these two: https://developer.nvidia.com/blog/nvidia-800-v-hvdc-architecture-will-… https://blogs.nvidia.com/blog/gigawatt-ai-factories-ocp-vera-rubin/ almost everybody in the industry is embracing 800V DC mostly because of Vera Rubin and the increased electricity requirements.

This was announced a while ago, and it's just now going into effect: > The company is implementing its long-lasting promise revealed in 2024, with the option for users to purchase additional play time as needed. There's also some rollover between months (up to 15 hours)[1]. Previously: https://news.ycombinator.com/item?id=42078726 https://news.ycombinator.com/item?id=42085089 https://news.ycombinator.com/item?id=42090845 1: https://blogs.nvidia.com/blog/geforce-now-thursday-performance-members…

My understanding is that it's in vogue to use deep learning for complex control problems, and the results are fairly impressive. The idea is to train robotic motion end to end with RL. Not an expert so I don't know the strength and weaknesses versus classical approaches. https://blogs.nvidia.com/blog/eureka-robotics-research/ https://arxiv.org/abs/2108.10470

Wow neat! There's a great collection of Māori-made technology for te reo Māori. I'm thinking also of Te Hiku Media's work building a Māori speech recognition system: https://blogs.nvidia.com/blog/te-hiku-media-maori-speech-ai/

Honestly, as much as I hate them, NVidia does an amazing job with this sort of thing. Camera motion through a static scene: https://blogs.nvidia.com/blog/2022/03/25/instant-nerf-research-3d-ai/ I've got no idea where the most recent deep fake progress has come from - my google foo wasn't up to the task, but with a lot of more recent videos it seems that they're actually dealing with deformable 3D surfaces, and not just deformations of 2D projections.

Great idea! In fact look at what nvidia is doing with videocalls (spoiler exactly what you’re saying, but video) https://blogs.nvidia.com/blog/2020/10/05/gan-video-conferencing-maxine…

A recent alumni gift from NVIDIA's cofounder will "Build Fastest AI Supercomputer in Academia" for UFL.[0] When I was researching colleges to attend, Florida seemed amazing for in-state tuition. Florida is still ranked one of the best states for higher education.[1] [0] https://blogs.nvidia.com/blog/2020/07/21/university-of-florida-nvidia-… [1] https://www.usnews.com/news/best-states/rankings/education/higher-educ…

There have been great strides made in ai compression in recent years. I wonder if they are going to leverage that. https://blogs.nvidia.com/blog/2020/10/05/gan-video-conferencing-maxine… https://ai.googleblog.com/2021/02/lyra-new-very-low-bitrate-codec-for.… https://aibots.my/officialBlog/deepminds-ai-agent-muzero-could-turboch…

Nvidia has been producing their own ARM cores since at least 2014: https://en.wikipedia.org/wiki/Project_Denver The ARM core performance hasn't been spectacular vs the latest and greatest from Samsung and Apple but combined with CUDA and tensor cores it can provide pretty incredible results considering the cost, form factor, and power consumption. Carmel is also at least two years old at this point... Their project Denver announcement from 2011 discusses plans for server and HPC applications: https://blogs.nvidia.com/blog/2011/01/05/project-denver-processor-to-u…

NVidia did a nice project a few years ago to model the moon landing in great detail.[1] It's a demo for their graphics cards. They did ray traced renders using actual reflectance data for the objects involved. The lunar surface is reflecting enough light to provide the effect of diffuse lighting from below. It's like being on sunlit sand under a dark sky. NVidia comments that either the moon landing is real or NASA built a time machine to travel 50 years into the future to get a good GPU. (For reference, the most powerful computer in the Apollo era, the CDC 6600, was about 10 MIPS. Modern desktop CPU, about 50,000 MIPS. Current top NVidia gamer GPU, about 10,000,000 MIPS. No way could computers of that era render photorealistic pictures.) [1] https://blogs.nvidia.com/blog/2014/09/18/debunked/ [2] https://blogs.nvidia.com/blog/2018/10/11/turing-recreates-lunar-landin…

Or follow NVidia's security pick - SPARK, a subset of Ada [1] for their autonomous driving and automotive firmware software. [1] https://blogs.nvidia.com/blog/2019/02/05/adacore-secure-autonomous-dri…

I wonder if this exploit would be workable on older Tegra systems, like for example a Tegra 3 int he digital cockpit on the Audi S3/R8/TT [1] or the K1 they are selling now [2] - it would be really great to be able to modify and customize those systems. [1] https://blogs.nvidia.com/blog/2016/04/25/virtual-cockpit/ [2] http://www.nvidia.com/object/visual-computing-module.html

Reading this article along with the following one, is striking: https://blogs.nvidia.com/blog/2017/11/01/gtc-dc-project-maven-jack-sha…

For comparison, at 180 teraflops per TPU (assumption: at fp32) this is equivalent to offering a cluster with 15,000 Titax Xp GPUs (1200$ per GPU = 18 Million $ worth of compute power). If TPU teraflops are reported at fp16, then the number would be half. Titan X offers 12 TFLOPs per GPU https://blogs.nvidia.com/blog/2017/04/06/titan-xp/

While the Switch supports Vulkan, it isn't the main graphics API, that role belongs to NVN. https://blogs.nvidia.com/blog/2016/10/20/nintendo-switch/ So it remains to be seen which one gets the love of game developers.

> The only way I can see this thing taking off is if it can fall back into an Android That's actually a great idea. Someone else pointed out this press release [0] from Nvidia saying they're making the chip. Tegra's can run Android (Nexus 7, 9) which has the latest OpenGL ES and Vulkan support. [0] https://blogs.nvidia.com/blog/2016/10/20/nintendo-switch/

I love that this story basically contradicts another trend that is in Hacker News at the same time: https://news.ycombinator.com/item?id=12414764 https://blogs.nvidia.com/blog/2016/09/02/gradescope-brings-ai-to-gradi…

Nvidia's Maxwell GPU architecture seems to add some very nice VR features [1]. With Maxwell coming to Tegra, and Nvidia finally being able to utilize heterogeneous computing, it seems to me Samsung would be really smart to try and get the next Tegra chip for their Note 5 or whatever phone they use for Gear VR 2, next year. Oculus/Carmack also have some great relationships with Nvidia, and it seems Carmack has already been able to convince Samsung to do a lot of technical changes for them. I think if they make this happen, Gear VR 2 would be a lot better next year. [1] - http://blogs.nvidia.com/blog/2014/09/18/maxwell-virtual-reality/

In 1982 you still had "supercomputers" like https://en.wikipedia.org/wiki/Cray_X-MP because you could still make bipolar electronics that beat out mass-produced consumer electronics. By the mid 1990s even IBM abandoned bipolar mainframes and had to introduce parallelism so a cluster of (still slower) CMOS mainframes could replace a bipolar mainframe. This great book was written by someone who worked on this project https://campi.cab.cnea.gov.ar/tocs/17291.pdf and of course for large scale scientific computing it was clear that "clusters of rather ordinary nodes" like the https://www.cscamm.umd.edu/facilities/computing/sp2/index.htm we had at Cornell were going to win (ours was way bigger) because they were scalable. (e.g. the way Cray himself saw it, a conventional supercomputer had to live within a small enough space that the cycle time was not unduly limited by the speed of light so that kind of supercomputer had to become physically smaller, not larger, to get faster) Now for very specialized tasks like codebreaking, ASICs are a good answer and you'd probably stuff a large number of them into expansion cards into rather ordinary computers and clusters today possibly also have some ASICs for glue and communications such as https://blogs.nvidia.com/blog/whats-a-dpu-data-processing-unit/ ---- The problem I see with people who attempt parallelism for the first time is that the task size has to be smaller than the overhead to transfer tasks between cores or…

RL can be massively disappointing, indeed. And I agree with you (and with the amazing post I already referenced [1]) that it is hard to get it to work at all. Sorry to hear you have been disappointed so much! Nonetheless, I would personally recommend even just learning the basics and fundamentals of RL. Beyond supervised, unsupervised, and the most-recent and well-deservedly hyped semi-supervised learning (generative AI, LLMs, and so on), reinforcement learning indeed models the learning problem in a very elegant way: an agent interacting with an environment and getting feedback. Which is, arguably, a very intuitive and natural way of modeling it. You could consider backward error correction / propagation as an implicit reward signal, but that would be a very limited view. On a positive note, RL has very practical sucessful applications today - even if in niche fields. For example, LLM fine-tuning techniques like RLHF successfully apply RL to modern AI systems, companies like Covariant are working on large robotics models which definitely use RL, and generally as a research field I believe (but I may be proven wrong!) there is so much more to explore. For example, check Nvidia Eureka that combines LLM to RL [2]: pretty cool stuff IMHO! Far from attempting to convince you on the strength and capabilities of DRL, just recommending folks to not discard it right away and at least give it a chance to learn the basics, even just for an intellectual exercise :) Thanks…

Nvidia has a small lead on the industry in a few places, adding up to super attractive backend hardware options. They aren't invincible, but they profit off the hostility between their competitors. Until those companies gang up to fund an open alternative, it's open season for Nvidia and HPC customers. The recent Stable Diffusion results are great news, but also don't include comparisons to an Nvidia card using the same optimizations. Nvidia claims that Microsoft Olive doubles performance on their cards too, so it might be a bit of a wash: https://blogs.nvidia.com/blog/2023/05/23/microsoft-build-nvidia-ai-win… Plus, none of those optimizations were any more open than CUDA (since it used DirectML). > crunching 4/8/16 bit width numbers is no rocket science. Of course not. That's why everyone did it: https://onnxruntime.ai/docs/execution-providers The problem with that "15 competing standards" XKCD is that normally one big proprietary standard wins. Nvidia has the history, the stability, the multi-OS and multi-arch support. The industry can definitely overturn it, but they have to work together to obsolete it.

I am sure most at OpenAI realize that they don't have a moat[0] that will keep them ahead for an extended period of time. Being a first mover does have its advantages of course, funding and resources from MSFT also help, but if they continue to (temporarily) take away features from their paying customers, are not forthcoming with whether there have been changes to in-use models and generally make it hard to rely on and trust them to deliver a consistent service, that will further people looking for alternatives, be that other competitors[1] or local models[2]. Unless they can capture regulators in ways that would make local models impossible, they have a very narrow time window to retain a large section of the market in the medium term. Despite their best efforts at presenting themselves as the "ethical" ones in this industry and attempts to convince governments that open/local models must be regulated[3], the only way I can see them accomplishing longer term limits on local LLMs considering the many purposes GPUs and inference accelerators can serve to the public will be hard. I have said this before, but I view this similarly to the Crypto Wars[4] or more recently Nvidia LHR[5]. I can see a Clipper Chip[6] for GPUs, that tries to limit inference to features approved for the common public, (mainly gaming focused applications like DLSS) in the near future, but I am certain that like before, where there is a 10ft wall, 12ft ladders will emerge. Basically, OpenAI must…

>> I'm willing to bet that 99% of the performance difference will be solved by value types. Anyway, Cliff Click, who wrote H2O[1], a large machine learning platform, reports achieving Fortran speeds (i.e. maximum throughput) with pure Java[2]. This means that a C application won't even be 1% faster. I've had a personal conversation with Cliff himself. Java no matter what you do can't deal with hardware acceleration and gpus. We agreed on that. Numerical software is a different beast. I also kept mentioning "simd instructions" as well as things like openmp. You are talking about systems software. Unfortunately that matters a lot for machine learning. The axis along which you can get equivalent speeds should be specified here. "values types" != "runs on faster chips" You aren't likely to beat intel or nvidia's compilers at their own game here . Java will always be playing catch up to last gen's tech there. Disclosure: I'm more than aware of what's going on in the space. We compete with them for customers and have a very clear understanding of their offerings. H20 has a great k/v store based on the exact mechanics you're talking about. That's about it though. Also of note: Cliff doesn't work on h20 anymore: https://twitter.com/cliff_click/status/700817408110399492 >> Oh, sure, for some specialized use cases of course that's true, but you could use OpenCL in Java, too.... OpenCL isn't exactly the industry standard for this stuff. You always end up using cuda, and you…

Analytical Flavor Systems[0]. We built a AI for beer flavor profile consistency and quality optimization. NVIDIA wrote an awesome article about us here[1]. More Info Analytical Flavor Systems uses machine learning and artificial intelligence to build tools for the food & beverage industry. Our Quality, Process, and Market Intelligence services create real-time predictive decisions metrics at each stage of a products life-cycle. We leverage our predictive models across products & industries for flavor profile optimization, production process optimization, demographic targeting & cognitive marketing - helping companies create and sell the best product to their highest value consumers with every batch. Our Services __Quality Intelligence__: Real-time predictive quality control, assurance, and improvement from human sensory data. __Process Intelligence__: Real-time predictive process control and optimization from human sensory data + manufacturing & LIMS data. __Market Intelligence__: Linking flavor-profile, demographics, and sales data to find the highest value consumer demographics for a product's flavor-profile. [0]www.Gastrograph.com [1] http://blogs.nvidia.com/blog/2015/09/02/beer/

>Deep learning, modeled on the human brain, is infinitely more complex [than machine learning]. Unlike machine learning, deep learning can teach machines to ignore all but the important characteristics of a sound or image-a hierarchical view of the world that accounts for infinite variety. -Yann LeCun I strongly disagree with a lot about this quote, even though it comes from a brilliant man I highly respect (my thesis research was inspired from some of his older work, and in my job we work with techniques he developed around deep learning). What I dislike is it utilizes hype to make deep learning seem mystical; in actuality it's a natural extension of old techniques that clearly fall under "machine learning". Deep learning is a neural network with 3 or more layers instead of the 2 layer networks that were developed in the 1980s. People tried 3 layers back then, and they didn't work well. Yann LeCunn and other researchers found cool ways to get 3+ layers to work in the 90s and again in the mid 2000s. More recently, researchers have just thrown a ton of data and computational power at them to get them to work. But fundamentally this was an extension of established techniques. This article that recently hit the front page actually breaks it down really well: https://blogs.nvidia.com/blog/2016/07/29/whats-difference-artificial-i… I think my main point here is that deep learning is quite accessible to people who are learning machine…

Your points are excellent, but number 2 seems a weak: why think of it as a single "machine"? If that's the metric, dual- and quad-processing machines would counter any advantage. Plus or minus some overclocking, a GTX280 and a Core i7 960 both consume roughly the same number or watts (130 for the processor vs 200 for the card?) and cost roughly the same (at least they did when I bought comparables). So while you can stick (and probably should) stick multiple cards in the same machine, it seems a little silly measure this in terms of multiples of single CPU performance. NVidia's blog response is interesting, and the some of the comments are actually worth reading: http://blogs.nvidia.com/ntersect/2010/06/gpus-are-only-up-to-14-times-… To me, the interesting question here is the degree to which the code has been optimized in each case, and the difficulty of doing so. A while ago there was a competition to attempt to generate SHA1 hash collision: http://www.engineyard.com/blog/2009/programming-contest-and-the-winner… The GPU's trounced every other approach. I participated in the contest minimally (running slightly tweaked versions other people's code on both an Nvidia GPU and a Core i7) and I was amazed at how much more efficient the GPU solution was. I don't recall the exact ratios, but 'orders of magnitude' was definitely not an exaggeration. But that's sort of a contrived example. So what I'd love to see here is a framework…

Official blog post: https://blogs.nvidia.com/blog/geforce-now-thursday-linux-native-app/

Azure has a lot of GB300s: https://blogs.nvidia.com/blog/microsoft-azure-worlds-first-gb300-nvl72…

In a way does this study try to detract from the public backlash against datacenters? Look at the author of the blog. https://blogs.nvidia.com/blog/author/joshparker/ Is this NVIDIA lawyer trying to influence the public perception of datacenter for his company?