@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20241230

Episode Date: December 30, 2024

- China, Advanced GPUs, Advanced AI - High Tech companies pursue government contracts - Neuromorphic chips, artificial fast neurons - Farewell 2024, thank you @HPCpodcast listeners [audio mp3="https://orionx.net/wp-content/uploads/2024/12/HPCNB_20241230.mp3"][/audio] The post HPC News Bytes – 20241230 appeared first on OrionX.net.

Welcome to HPC News Bites, a weekly show about important news in the world of supercomputing, AI, and other advanced technologies. Hi, everyone. Welcome to HPC News Bites. I'm Doug Black of Inside HPC, and with me is Shaheen Khan of IranX.net. There's a story from the Wall Street Journal that Chinese companies are doing impressive things in AI in spite of American trade sanctions that cuts off China's access to the most advanced AI chips from NVIDIA. The journal cited Chinese AI companies DeepSeek and Moonshot AI for developing

workarounds to their processing disadvantages by developing extremely good software and hardware training stack capabilities, according to Anthropix's Jack Clark. DeepSeek, for example, released a preview of its new LLM in November that it said compares favorably with OpenAI's O1 reasoning model announced in September. Moonshot AI, which is backed by Alibaba and Tencent, meanwhile, said it developed a math model with capabilities close to those of 01, while Alibaba said one of its own experimental models outperformed the preview version of the U.S. model on math. The journal noted that the companies haven't released detailed information

about their models and, quote, evaluating the claims is difficult because there isn't a single agreed-upon test of an AI model's capabilities, unquote. Still, some U.S. AI specialists said they are impressed. A former OpenAI fellow, Andrew Carr, said China is, quote, catching up faster, adding that deep-seek researchers trying to replicate OpenAI's reasoning model, quote, figured up faster, adding that deep-seek researchers trying to replicate OpenAI's reasoning model, quote, figured it out within a few months. And frankly, many of my colleagues are surprised by that. Leading in AI, or at least being up there among the top countries, is the ambition of many countries. This requires data, algorithms, hardware, and the expensive

talent and skills that they require. And there's a lot of room for optimization techniques everywhere. The idea that China's capabilities in AI are top-notch is not new. Some years ago, the prevailing view was that in some applications of AI, China is ahead of the West. Evidence in favor of that view included and continues to include a large share of AI research papers and patents, strong government support, which also means access to massive amounts of data due to less stringent data privacy regulations, several well-known and not so well-known companies doing advanced work, and social and industrial applications that make it real. And then there

is nimbleness, both in terms of catching up with methods and algorithms and finding ways to get around chip restrictions. Training even a trillion parameter model is a matter of months, not years. So slower chips can get you there a bit later. And everyone is looking for less demanding and faster algorithms too, some of them well-known. Examples of those include trial and error thinking or mixture of experts, otherwise known as MOE, an acronym that we see often. So really, not having access to the latest GPUs seems to be less of a wall and more of a slope. And sometimes a bit of slope is the point and the delays can add up.

We saw an interesting blog post last week from the publication The Information, this one by Corey Weinberg, stating that some of the most valuable tech startups, including Palantir, SpaceX, Anduril, Scale AI, OpenAI, and others, are forming, quote, a consortium that will jointly bid for U.S. government work in an effort to disrupt the country's oligopoly of prime contractors such as Raytheon and Northrop Grumman, according to the Financial Times. You know, Shaheen, this makes complete sense considering that AI is entering so many aspects of life and business, including goods and services required within the federal sector. According to Weinberg's post, Silicon Valley has gotten its first taste of what success in government contracting can look like,

thanks to the soaring valuations for SpaceX and Palantir in particular. That's right. Some of those companies already sell a lot to the government. One could say they probably would not exist without government contracts, and they seem to be well-connected to the incoming administration. As it relates to efficiency, well, anything can be done better, so hopefully they will help improve efficiency, but anything can be done worse too. The stereotypical Silicon Valley model is to move fast and break things, and fail fast and iterate, but this mindset doesn't always work for mission-critical or life-critical situations. This is part of what government contractors and systems integrators do. It's not just raw technology, they're not just middlemen, and it's not just lobbying. While

joining forces can pre-integrate some or all of a solution, what they really need to do in order to disrupt traditional prime contractors is to change what the government buys versus who it buys it from. That sounds consistent with Weinberg's article when it said, quote, a coalition could also put up a bigger fight against government agencies it thinks are writing unfair contract requests, end quote. In a quest for better and lower energy AI processing, we've discussed neuromorphic chips, which digitally mimic neurons. There have also been various efforts to build artificial neurons, which is very intriguing, but looks hard to make it go fast. Well, a team of seven scientists from Hong Kong and China have published a paper on, quote, photonic neuromorphic computing, which could be used to build an artificial neural net that relies on fast lasers to communicate among neurons.

The authors stated that, quote, the widely studied photonic spiking neurons suffer from limitations in signal processing speed is due to the refractory period, and most of them are optical injected, requiring additional laser sources and modulators. The researchers experimented with a bio-inspired integrated photonic-rated neuron based on an electrical injected quantum dot laser. This new type of signal injection method enables the laser to avoid the pulse response rate limitations caused by the refractory period, thereby achieving high processing speed. The laser neuron does not require external laser pumps and modulators, leading to a simple and low-energy consuming system. All right, that's it for this episode as we bring HPC News Bites to a close for 2024.

It's been another enjoyable year podcasting with you, Shane, and I'm looking forward to a great 2025. We'll be doing a 2024 year in review episode on the At HPC podcast soon after the start of the new year that may even include a few choice predictions for developments we think will become a reality a year from now. But you know, with the pace of progress accelerating so rapidly, perhaps our predictions could come true by July, maybe even April. April is right. We can certainly make them true on April 1st. And yes, we're now in the fourth year of this podcast, Doug, and thanks are due. Besides your insights, eloquence, sparkling humor, how am I doing there, Doug? Very well, very well.

But more seriously, we want to thank you, our listeners, for spending your precious time with us. We very much appreciate the steady growth of our global audience since we began in late 2021. And we want to encourage you to reach out to us with suggested topics and guests. Happy holidays to all of you. And here's to a great 2025. Thank you. HPC News Bites is a production of OrionX

in association with Inside HPC. Shaheen Khan and Doug Black host the show. Every episode is featured on InsideHPC.com and posted on OrionX.net. Thank you for listening.