@HPC Podcast Archives - OrionX.net - HPC News Bytes – 20251215

Episode Date: December 15, 2025

- Nvidia H200 exports to China - H20, H200, Chinese chips: how do they stack up? - Few fast GPUs vs many slow GPUs - China’s electricity production - Datacenter electricity use in the US - Cell-phone sized AI supercomputer - HPC at the edge - Regulating AI [audio mp3="https://orionx.net/wp-content/uploads/2025/12/HPCNB_20251215.mp3"][/audio] The post HPC News Bytes – 20251215 appeared first on OrionX.net.

Welcome to HPC Newsbytes, a weekly show about important news in the world of supercomputing, AI, and other advanced technologies. Hi, everyone. Welcome to HPC Newsbytes. I'm Doug Black of Inside HPC, and with me is Shaheen Khan of OrionX.net. It was a big week for techno-geopolitics last week. The Trump administration bowed to the case presented by NVIDIA, C. CEO, Jensen Wong, and other technology executives, to allow exports of NVIDIA H-200 GPU,

something China's President Xi said is a positive development. The administration also bowed to its desire for 25% of NVIDIA's revenue from sales of H-200s to China. As for China, a CNBC report said the PRC will allow only limited numbers of H-200s into the country. The policy change is actually a double reversal. The U.S. had said no NVIDIA chips could be exported to China and the PRC said it would disallow NVIDIA imports as it strives to build up its domestic AI chip capabilities. We should note that the H-200 is a generation behind NVIDIA's current flagship Blackwell GPU and soon it will be two generations behind the upcoming Rubin. Meanwhile, the information reports that China's A.I. company Deepseek is using banned NVIDIA chips to build their next model. What this means in part is that between the desire of chip companies and the U.S. government to reap revenue from China chip exports and China's demand for advanced AI chips, it really means export restrictions face a number of challenges.

The big impact on Chinese AI development and on NVIDIA2, since it led to a $5 billion dollar write-off for them was when the H-20 chips were banned a few months ago. The H-20 was a downgraded version of the H-200, and it was designed to meet expert control parameters for the Chinese market, but was generally the same ballpark as locally made Chinese chips with better memory bandwidth and very fast GPU-to-GPU connectivity. So a very decent chip and better enough to be compelling at the time. The H-200, as you mentioned, is a generation behind the Blackwell family B-200 and B-300 chips. But it is about six times faster than the H-20,

about 50% more memory, and 20% higher memory bandwidth. It is also heads and shoulders ahead of locally made Chinese chips, which, as far as we know, include Huawei Ascend 910B and Cambercon, Saiwan 590, both of which are in the same ballpark or faster than H20, but way behind H-200. So the question for China is whether they can use more of their own chips to achieve the same performance and do it without too much loss of efficiency.

That's the same as asking whether many weaker and cheaper chips or systems can deliver the same performance as fewer, more powerful ones, and do so at acceptable costs. It is the same question that led the HPC industry towards Beowulf clusters versus big vector supercomputers at the time. Most NVIDIA competitors have to deal with that dynamics. How you get value from volume, so to speak, is a fundamental question of HBC. And if a nation state is pursuing that path, it's probably prepared to put up with much higher costs or inefficiencies than a business would, which could explain why China is

trying to look like it's not rushing to snap up these H-200 chips. I encourage our listeners to check out the in-depth podcast we did with Dan Neistead a few weeks ago, where we discussed this topic. Sticking with AI and China, the Wall Street Journal came out with a big story last week on an area closely related to AI, and that is electrical power. Quote from the story, China now has the biggest power grid the world has ever seen. Between 2010 and 24, its power production increased by more than the rest of the world, combined. Last year, China generated more than twice as much electricity as the U.S. Some Chinese data centers are now paying less than half what American ones pay for electricity. Among the many areas where China and the West differ is how they bring power plants online,

be they coal, natural gas, nuclear, or solar, and wind. While the West and the U.S. have regulatory and legal hurdles to be overcome, it's fair to say the government in China doesn't worry about these considerations overly much. Right, NVIDIA's unit sales just in Q2 were something like 11 million chips for gaming systems and 6 to 7 million GPUs for the data center, most of them Blackwells. If you assume 1 to 1.5 kilowatts average power draw per data center GPU, that adds up to gigawatts. The amount of energy used by data centers in the US is expected to grow from around 4%

of total a couple of years ago to 7 to 12% by 2030. That would be 325 terawatt hours to 580 terawatt hours according to estimates, almost all driven by AI. So as you mentioned, renewables supported by battery storage, natural gas, nuclear energy, plus extending transmission lines and the grid work that is required are really the ways to meet increasing demand from data centers and electric vehicles and everything else that is finally changing to use electricity. We ran a story last week on Inside HBC about a, quote, personal AI supercomputer from a company called Tiny T-I-N-Y, Tiny AI that they say runs 120 billion parameter large language models on a device about the size of a cell phone and that does not have GPUs or connect to the cloud or servers.

I thought this was impressive performance, but the article sparked criticism from a reader who said the NVIDIA Jetson Thor runs bigger LLMs faster and in the same form factor, and a maxed-out Apple Mac Mini also is capable of running similar 120 billion-sized models. I would say those are fair points, but we feel the tiny AI device is worthy of coverage, and inside HPC stands by its story. Yeah, that's right. Well, you know, GPUs started outside the data center in PCs and expanded into data centers in servers, and let's call them big PCs just to make the point, but are expanding back outside. Small boxes like the Mac Mini or the DGX Spark are really very cool, but they're

still a box meant to be fixed to a table. But a new wave of AI-enabled devices are now coming to market, and in that space, it's not just about the configuration. but about the specific use case, form factor, environmental requirements, transport, power, compliance, and many other requirements. As I like to say, if you need a soccer ball with a GPU in it, it needs to be a soccer ball first,

and you need to be able to kick it. And that sort of stuff determines the engineering and the components you need and how to build it. That is what HPC at the edge will mean. And to me, that's what the news points to, to the fact that digitization of human life requires computational power everywhere and in everything, not just in box-shaped devices in controlled environments. HPC everywhere really is the message.

President Trump signed an executive order, allowing the federal government to override AI laws at the state level with the intent of creating a unified national AI standard. Under the EO, states deemed to have passed restrictive AI laws, be punished by the Department of Justice. Tech companies have lobbied the White House to issue such a directive, arguing that a variety of state-based laws could hamstring the country's ability to compete on AI with China. What to do about AI is a global question, and we see different strategies at play in different parts of the world. AI is widely viewed as a global historic

phenomenon with human-scale opportunities and fears. So do you rush in and win the prize, or you rush in and get burned? Will you enable a new age of humanity and access new possibilities, or will you accelerate AI-enabled threats? What is certain, however, is that AI was not and is not stoppable. It is a global race. The executive order specifically mentions the Colorado AI Act as an example of a law that the administration may challenge, but other states like California, Texas, and Utah also have laws to regulate AI, whether it's transparent, and disclosure requirements or guarding against deep fakes or other aspects of AI right now. Some states, including South Dakota, have banned the use of deep fakes to influence elections or create imagery without permission.

Others like Idaho, Montana, and New Hampshire, have laws which go the other way to ensure AI development can continue. The administration's view is that some of these laws are too burdensome to comply, especially for startups. and some risk ideological bias, and therefore, a uniform regime across the country is the best way forward, and it believes it needs to preempt such state regulation. So state laws remain in place, but federal agencies are directed to pursue legal challenges and funding restrictions when state laws deviate from administration policies. As usual, such policies would ideally be passed as laws, which would provide more certainty for businesses, and also make them immune to change.

changes in administrations and probably counter-challenges. But the pressure to preserve U.S. AI competitiveness and especially against rivals like China, which do not typically have such compliance burdens, is very real. All right, that's it for this episode. Thank you all for being with us. HPC Newsbytes is a production of Orion X in association with InsideHPC. Shaheen Khan and Doug Black host the show. Every episode is featured on InsidehPC.com and podcast. posted on Orionx.net. Thank you for listening.