<div id="readability-page-1"><div id="body">For as long as I have been a reporter and analyst in the IT sector, November has always been supercomputing month. Way before there was a TOP500 ranking of supercomputers in June 1993 but just as I was leaving university, the first Supercomputing Conference was held in Orlando in 1988. And that November SC show set the cadence for high-performance computing for the decades that followed.With the original SC conference, just under 1,500 people and 36 companies showed up. This time around, more than 16,500 people attended SC25 in St. Louis, Missouri, and 559 exhibitors were hawking their wares. As our HPC pal Dan Olds quipped at dinner, of those 559 companies, 732 of them were showing off plumbing to keep GPU-accelerated systems from melting.This being the HPC crowd, high-precision FP64 floating point computing is still an important aspect of the systems that run all manner of simulations and models. While there are tricks to take the solvers written for 64-bit processing and adapt them to the lower precision vector and tensor units that have been embedded in GPUs to accelerate AI training and inference, this is still not common practice and in fact we do not know how practical this approach is at scale. There are certainly technical reasons for HPC centers to want to do this, given the high prices they pay for FP64 computing. All things being equal, HPC centers would rather not have to gut their applications to get a 10x improvement in applications. Then again, they seem to be on a Sisyphean task of doing just that over the more than six decades since supercomputing as we know it - arguably starting with the CDC 6000 designed by Seymor Cray for Control Data Corp - came into being. The reason is that absolute performance often requires jarring architecture changes, ones that the mainstream of enterprise computing cannot stomach but which HPC centers exist precisely to bear. Over at The Next Platform, we have occasionally compared the cost of GPUs with gold just to get a sense of how precious these intricate devices are. And for Supercomputing Month at El Reg, we thought it would be fun to actually compare the cost of GPU accelerators and systems to precious and exotic metals that are used in semiconductor manufacturing. So we gathered up some data since the GPU acceleration for HPC applications wave really got going in 2013 with the &#34;Kepler&#34; K40 GPUs.Nvidia has very rarely given out the weight of its SXM GPU cards, but there is some data out there when people resell them. We did our best to estimate the weights of these cards up to the current &#34;Blackwell&#34; B200 GPU generation. The weight of the DGX servers is known, so we also calculated the price per ounce for these systems too. And because we are focused here on double-precision 64-bit floating point with occasional divergence into single-precision floating point, we have added the performance specs for each card to the table. Take a gander:<div><img src="https://imageproxy.pbkrs.com/https://regmedia.co.uk/2025/11/28/nvidia-gpu-versus-precious-metals.jpg/query-aW5mZXJfeT0xJng9NjQ4Jnk9MjQw?x-oss-process=image/auto-orient,1/interlace,1/resize,w_1440,h_1440/quality,q_95/format,jpg" height="240" width="648" original-src="https://imageproxy.pbkrs.com/https://regmedia.co.uk/2025/11/28/nvidia-gpu-versus-precious-metals.jpg/query-aW5mZXJfeT0xJng9NjQ4Jnk9MjQw"/>Click to enlarge</div>And just for fun, and knowing how turkey has its price artificially reduced this time of year, we added in the price of live &#34;fed cattle per hundredweight&#34; and converted this wholesale price of a live cow to ounces to compare to GPUs and gold.As it turns out, it may feel like GPUs are made of gold because of their high cost, but gold is actually more expensive per ounce than an Nvidia datacenter GPU. As far as we can tell, the price of GPU accelerators peaked at just under $650 an ounce in 2024 with the &#34;Hopper&#34; H200 devices, and gold was 4x as expensive per ounce at $2,624 as 2024 came to an end (these are all year-end prices per ounce in the table above). The cost of GPU accelerators has actually come down as 2025 has come to a close with the Blackwell B200 SXM6 device trading at around $330 per ounce as gold has skyrocketed to more than $4,000 per ounce. The gap is now 12.4x between GPUs and gold.Platinum and palladium are still considerably more expensive per ounce than GPUs, but germanium is about half the price even after a huge spike this year. Silver and gallium are much cheaper, and copper is relatively free although a penny in the United States now takes about 4 cents to make, which is stupid. And while it may seem like beef is more expensive than gold, or GPUs, it is about 12 cents per ounce for a live cow and around 55 cents to 60 cents per ounce to buy a good steak from the supermarket. It can cost many times that again to get a great steak at a very good steakhouse, of course, but the entrée has side dishes and you are also helping to stimulate the economy when you buy it. ® </div></div>

INTC

The article discusses the high cost of GPUs in the context of supercomputing, comparing their price per ounce to precious metals like gold, platinum, and palladium. Despite their high cost, GPUs are cheaper than gold, which has risen to over $4,000 per ounce. The article highlights the importance of GPUs in high-performance computing and their evolution since 2013.

- The Supercomputing Conference, held annually in November, has grown significantly since its inception in 1988, with over 16,500 attendees and 559 exhibitors at SC25 in St. Louis.  
- The article compares the costs of GPU accelerators to precious metals, revealing that while GPUs are expensive, gold remains significantly pricier, with a price gap of 12.4 times as of late 2025.  
- Additionally, the cost of GPU accelerators has decreased, while gold prices have surged, highlighting the evolving economics of high-performance computing.

GPUs aren't worth their weight in gold – it just feels like they are