Silicon Data has created SiliconMark as a benchmarking tool for rented decentralized GPUs. You can think of it as Carfax for GPUs.
Renting graphics processing units (GPUs) for AI and cloud gaming computing purposes is becoming a thing. You can tell because Silicon Data is making a living benchmarking these GPUs with Silicon Mark.
The GPU rental market is a place where companies that want GPUs but can’t get their hands on them can turn. There are a lot of companies that have purchased GPUs for their cloud-connected data centers to handle tasks like AI and cloud gaming. But those GPUs aren’t busy all of the time. They can rent them out, said Carmen Li, CEO of Silicon Data, in an interview with GamesBeat.
“We’re in conversation with the exchanges. I’m amazed at how convoluted the market is and how non-transparent it is. For SiliconMark, I think it will be very critical to be the Carfax for GPUs,” Li said.
And there are also a lot of companies that want GPUs but cannot get their hands on them. This is why Nvidia is valued at nearly $4 trillion and why its revenues rose 75% last quarter. So using decentralized networks or other deals, those with excess GPUs rent them to those who need them. The supply and demand can fluctuate fast. If a game suddenly becomes popular, a company will need a lot more of them to satisfy gamers. Or if AI applications take off, another company will need a lot of GPUs to do the computing.
Silicon Data can help benchmark these GPUs and whether they’re really worth renting or not. You can kind of think of this part of what Silicon Data does as Carfax for cloud-based GPUs.
It’s all happening very fast. The idea of renting GPUs hasn’t been around that long.
“We have had very exciting ride the past year and a half,” Li said.
The GPU rental index tracks GPU usage per hour, and it’s the precursor to financial products such as futures and other data analysis.
“While we were developing the index, we realized that” there were different factors that could affect pricing, like how new the GPUs are, whether they are configured in certain ways, have different quality levels, Li said.
“We saw there was no benchmark fee service. There is no transparency,” Li said. “So we stepped in just to do GPU benchmarking.”
In April, the company published its first index for prices for the GPU rental market.
Li began her career as a trader. She traded options for natural gas in Chicago in financial services. She found it to be a bit boring, and moved on to work on global data integration at Bloomberg.
Then, at the start of 2024, she started Silicon Data to work on a new idea.
Silicon Data, the market intelligence company that wants to transform how compute infrastructure is priced and benchmarked, announced the launch of SiliconMark, the industry’s first real-world, cross-platform GPU performance benchmark designed specifically for AI infrastructure buyers.
SiliconMark is now available in private beta to select customers. Built for AI developers, infrastructure teams, and data center operators, SiliconMark enables users to test the true performance of rented GPUs, including NVIDIA H100, A100, AMD MI300, and Google TPUs, before deploying expensive training jobs.
In under 90 seconds, it surfaces how chips perform in real-world conditions across major rental platforms, highlighting performance drift, pricing inconsistencies, and deployment readiness.
SiliconMark integrates with Nvidia’s DGX Cloud Benchmarking , ensuring performance
comparability across both cloud and high-end on-prem systems.
“Spec sheets don’t tell the whole story,” said Carmen Li, CEO of Silicon Data, in a statement. “SiliconMark is about visibility. If you’re spending hundreds of thousands on GPU time, you should know if you’re getting what you paid for, and if you’re overpaying for
underperforming infrastructure.”
While its doing the benchmarking, there’s also a lot of information that Silicon Data can surface on the state of the GPU economy. That’s where I thought things were getting really interesting.
The company has six people and it has raised $4.7 million.
The release follows a wave of momentum for Silicon Data:
● The company’s GPU Rental Price Index (SDH100RT), the first daily benchmark for
Nvidia H100 rentals, was recently published on Bloomberg terminals and highlighted in
Bloomberg’s AI infrastructure coverage.
● IEEE Spectrum profiled the index in its upcoming July issue, noting its role in turning
GPU capacity into a measurable and tradeable commodity.
Now, with SiliconMark, the company is expanding from pricing transparency into performance validation, helping buyers benchmark infrastructure like traders benchmark financial products.
Key Benefits of SiliconMark:
● Vendor-neutral, fast benchmark (<5 minutes)
● Tests actual runtime behavior across GPU types and clouds
● Helps users negotiate pricing, spot performance drift, and optimize workloads
● Designed to empower renters, not hardware vendors
The SiliconMark beta is currently in use by AI infrastructure teams, LLM developers, and compute resellers navigating volatile GPU supply chains and opaque pricing environments. With training costs often exceeding six figures, these teams need ground truth performance data, not theoretical benchmarks, to make infrastructure decisions.
SiliconMark works across on-demand rental platforms, dedicated GPU hosts, and colocation setups. It is cloud-agnostic, infrastructure-light, and built with reproducibility in mind.
“This is about turning compute into a measurable asset class,” added Li. “We want to be the S&P for GPUs.”
Founded by former Bloomberg data executive Carmen Li, Silicon Data is building the
infrastructure layer for the compute economy. The company offers GPU rental indices, carbon intensity ratings, and performance benchmarks to help AI builders, financial institutions, and infrastructure providers understand the real cost and value of compute. Backed by leading investors including DRW and Jump Trading, Silicon Data is on a mission to make compute pricing as transparent and actionable as traditional commodities and capital markets.
Jason Cornick, head of infrastructure at Silicon Data, said in an interview with GamesBeat, “The challenge is the buyers [of GPUs] look in the obvious places: the hyperscalers, or AWS-like firms in the world, and they have no capacity. That’s when the market becomes inefficient, and it takes data from someone like Silicon Data to make the market more efficient again.
“They’re having a terrible time finding the buyers because there’s no trust,” Li said. “If there is more trust, then the market moves faster and you will feel more comfortable renting. The whole ecosystem can benefit from the transparency we bring to the market. Hopefully we can restore the trust.”
More analysis on GPU performance deviations and economic impacts
GPUs are powerful processors designed to accelerate data-intensive workloads, but they rely heavily on surrounding system components to maintain high throughput. Without the right supporting architecture, even top-tier GPUs can be underutilized, reducing performance and ROI.
Silicon Data said it is seeing maturity in the GPU space as hardware standards like Nvidia DGX and HGX platforms provide guidance to manufacturers on what is truly required to maximize GPU performance. As GPUs have become the most expensive component of modern accelerated computing systems it is essential to maximize performance for efficient capital application.
The company has been publishing its index since last September. From September through June, the GPU rental market showed a 23% decrease in pricing.
Silicon Data also sees the effectiveness of Nvidia’s engineering standards programs when it looks at the change over time. The following chart shows the newer H100 GPU with the smallest Spec differential/Pflop and as the GPUs get older the differential significantly increases, with an average Tesla V100 GPU only performing at 26.23% of what could be expected from the Specs.
The following table shows how different NVIDIA GPU models perform in practice versus their manufacturer specifications, alongside the economic implications per petaFLOP of compute.
Spec differential/Pflop indicates how much performance is lost per unit cost compared to
manufacturer specs. This has the obvious direct cost impact – if you lose 20% of your GPU processing to sub-optimal configuration that is an immediate loss, but there are indirect impacts associated as well. This performance loss doesn’t just increase costs — it also slows down data processing, delays model training, and reduces organizational agility.
There is also an additional hypothesis that we are testing, and that is that GPUs deteriorate much more rapidly than other components, like CPUs and RAM. Historically servers have been acquired with 5-7 year service expectations, but there appear to be issues with GPUs experiencing thermal degradation and having significant performance loss over a much shorter time frame, reducing the effective lifecycle.
SiliconMark uses GPU serial numbers to track individual GPU performance over time, and as it gathers more data, it seeks to test that hypothesis and understand the new server lifecycle for accelerated systems.
This kind of work can provide third-party validation and transparency for the services that one company is giving to another in lightning-fast transactions where time is of the essence.
“Imagine you’re any company, right? You have to finance your GPU cluster. You go to a bank and tell them, ‘Hey, I have 100 nodes. Bank says great. ‘Who can help me verify that they are of good quality? Or do they even exist?”
Both the parties that rent out the GPUs and those doing the renting can use this information. Those with GPUs can assess the pricing across the spectrum, and those who need GPUs can use the info to get the best deals in a transparent way.
“If you have access to a GPU instance, you can run the benchmark yourself,” Li said. “You see the results.”
SiliconMark captures enough data on the machines to have a good fingerprint on what the vendors offer, with serial numbers and hardware specifics that the end users will want to know. Since there hasn’t been much transparency in the market, it is tough for buyers to be good at their jobs. The biggest source of the secondary GPUs is often hyperscalers.
The market has its share of those who exaggerate what they have or are speculators. They may not be able to handle the toughest workloads, Li said.
“We help those resellers validate the inventory they’re being given by third parties so they have confidence that the [hardware] performs as expected,” Li said.
Li added, “There’s a lot of GPUs. People have jumped on the bandwagon, they’ve invested in them, and they want to resell it now,” Li said. “They may have speculated or may not have a fit for the resources they bought. They want to recover some value from their investment.”
They turn these GPUs over to resources who then rent them out or auction them off. Different kinds of parties like private equity firms can tap the data to get a more transparent view of a market opportunity.
The company is agnostic when it comes to the chip designers, whether they’re from Nvidia, AMD or Intel.
“We are very excited to work with a few largest financing projects right now to help them get more clarity,” Li said. “There’s so many blockchains out there. And again, there’s so many scams.”
