GPUs aren’t worth the weight in gold, but it feels like they are.

Celebrating Supercomputing Month: A Journey Through HPC and GPU Value

The Legacy and Growth of Supercomputing Conferences

November has long been synonymous with supercomputing milestones. The inaugural Supercomputing Conference (SC) took place in Orlando, Florida, back in 1988-years before the TOP500 list was introduced in June 1993. At that time, the event attracted just 36 companies and fewer than 1,500 attendees. Fast forward to SC25 in St. Louis, Missouri, where the conference drew over 16,500 participants and featured 559 exhibitors showcasing cutting-edge high-performance computing (HPC) technologies.

This event has consistently served as a pivotal platform for advancements in HPC, driving innovation and collaboration across the industry for more than three decades.

Precision Computing in HPC: The Role of FP64

High-precision 64-bit floating-point (FP64) calculations remain fundamental to HPC workloads, powering complex simulations and scientific models. While modern GPUs incorporate lower-precision vector and tensor cores optimized for AI training and inference, adapting traditional FP64 solvers to these units is still uncommon and largely experimental at scale. HPC centers often face the challenge of balancing the high computational cost of FP64 with the need for performance, sometimes opting to modify applications to leverage faster, lower-precision hardware.

Despite decades of architectural shifts since the era of Seymour Cray’s CDC 6000 series, HPC facilities continue to embrace disruptive hardware changes to maximize performance-an approach that mainstream enterprise computing typically avoids due to stability and compatibility concerns.

GPU Accelerators: Comparing Their Value to Precious Metals

At The Next Platform, we have previously likened the cost of GPUs to that of gold to highlight their immense value. For this Supercomputing Month, we expanded this comparison by analyzing GPU accelerators alongside various precious metals commonly used in semiconductor manufacturing.

Since the GPU acceleration surge in HPC began around 2013 with Nvidia’s Kepler K40 GPUs, we estimated the weight of Nvidia’s SXM cards-despite limited official data-by referencing resale information and known system weights such as those of DGX units. This allowed us to calculate the price per ounce for each generation, including the latest “Blackwell” B200 GPUs, while focusing primarily on double-precision (FP64) performance metrics.

Price Trends: GPUs Versus Gold and Other Metals

Our analysis reveals that although GPUs are costly, gold remains significantly more expensive per ounce. In 2024, Nvidia’s “Hopper” GPUs peaked at nearly $650 per ounce, whereas gold closed the year at approximately $2,624 per ounce-about four times higher. By late 2025, GPU prices have declined, with the Blackwell B200 trading near $330 per ounce, while gold prices have surged past $4,000 per ounce, widening the price gap to over 12 times.

Other precious metals such as palladium and platinum also command higher prices per ounce compared to GPUs. Conversely, germanium, despite a recent price spike, costs roughly half as much as GPU accelerators. Silver, gallium, and copper remain more affordable, with copper’s cost being negligible despite the small expense involved in minting pennies in the U.S.

Putting GPU Costs in Perspective: An Unconventional Comparison

For a bit of levity, we compared GPU prices to the wholesale cost of live cattle, converted to ounces, to provide a relatable benchmark. While premium beef at grocery stores typically costs between 55 and 60 cents per ounce, dining at upscale steakhouses can multiply that price several times over. Unlike GPUs or precious metals, purchasing beef also supports local economies and agricultural industries, illustrating how value can extend beyond mere cost.

Looking Ahead: The Future of HPC and GPU Innovation

As HPC continues to evolve, the interplay between hardware cost, precision requirements, and performance optimization remains a dynamic challenge. The ongoing development of GPU architectures and their integration into supercomputing environments will shape the trajectory of scientific discovery and AI advancements in the years to come.

More from this stream

Recomended