15 years of FP64 segmentation, and why the Blackwell Ultra breaks the pattern

Nvidia limits RTX 5090 double precision

Nvidia’s new RTX 5090 delivers 104.8 TFLOPS of single-precision (FP32) compute but restricts double-precision (FP64) performance to just 1.64 TFLOPS. This 64:1 performance gap represents a deliberate architectural choice rather than a technical limitation of the silicon. The company has spent fifteen years widening the divide between consumer and enterprise hardware to protect its high-margin data center business.

The RTX 5090 serves as the fastest consumer GPU on the market, yet its FP64 capabilities remain remarkably stunted. For researchers and engineers, this means the card excels at gaming and AI but struggles with the high-precision math required for scientific simulations. Nvidia uses this performance floor to ensure that customers who need heavy-duty math buy H100 or B200 enterprise cards instead.

This trend began in 2010 with the Fermi architecture. The GF100 die powered both the GeForce and Tesla product lines, and the hardware natively supported a 1:2 FP64-to-FP32 ratio. However, Nvidia used driver caps to throttle GeForce cards to a 1:8 ratio, creating the first artificial barrier between gamers and professional users.

A fifteen year performance divide

Nvidia eventually stopped using software drivers to limit performance and began changing the physical hardware. Modern consumer GPUs now feature far fewer dedicated FP64 cores than their enterprise counterparts. This structural shift ensures that GeForce cards cannot easily compete with data center hardware in high-performance computing (HPC) environments.

The performance ratio has degraded steadily across every major architecture release since 2010. While data center GPUs maintained high precision for climate modeling and financial simulations, consumer cards fell behind. The gap grew from a manageable difference to a massive chasm over the last decade and a half.

• Fermi (2010): 1:8 ratio
• Kepler (2012): 1:24 ratio
• Maxwell (2014): 1:32 ratio
• Ampere (2020): 1:64 ratio
• Blackwell (2025): 1:64 ratio

The raw numbers tell a stark story of lopsided development. Between the GTX 480 in 2010 and the RTX 5090 in 2025, FP64 performance only increased 9.65x, moving from 0.17 TFLOPS to 1.64 TFLOPS. During that same period, FP32 performance for gaming and general compute jumped 77.63x, rising from 1.35 TFLOPS to 104.8 TFLOPS.

Market segmentation drives hardware design

Nvidia’s strategy relies on the fact that gamers and video editors rarely need double-precision math. FP64 is essential for numerical stability in fields like computational fluid dynamics, quantitative finance, and structural analysis. By weakening this specific feature on GeForce cards, Nvidia creates a clear justification for the massive price premiums on its enterprise hardware.

The price gap between consumer and enterprise GPUs has exploded as a result. In 2010, a top-tier enterprise card cost roughly five times more than a flagship gaming card. By 2022, that multiplier reached 20x, with enterprise cards commanding tens of thousands of dollars. Nvidia justifies these costs through features like ECC memory, NVLink, and superior FP64 throughput.

Nvidia's Ampere GA102 whitepaper explicitly confirmed this design philosophy. The company stated that it includes a "small number" of FP64 hardware units simply to ensure that programs containing double-precision code operate correctly. These units exist for compatibility rather than performance, serving as a safety net rather than a functional feature for heavy workloads.

AI boom breaks the old logic

The rise of generative AI has complicated Nvidia’s long-standing segmentation strategy. Unlike traditional scientific modeling, AI training does not require high-precision FP64 math. Most modern AI models run effectively on FP32, FP16, or even FP8 and FP4 formats. This shift suddenly made consumer GPUs look like viable tools for serious AI research and development.

Startups and hobbyists realized they could train meaningful models on a GTX 1080 Ti or RTX 3090 without spending $10,000 on a Tesla card. This threatened Nvidia’s enterprise margins, as the technical barrier of FP64 no longer prevented professional use of consumer hardware. The company responded by changing its legal terms rather than its silicon.

In 2017, Nvidia updated its GeForce End User License Agreement (EULA) to prohibit the use of consumer GPUs in data centers. This controversial move replaced technical segmentation with contractual restrictions. The Verge reported at the time that the shift was unprecedented, as it used legal threats to prevent customers from using hardware they had already purchased in specific environments.

Emulating precision with tensor cores

Engineers have developed workarounds for the lack of native FP64 hardware on consumer cards. One method involves emulating double-precision math using single-precision floats. This technique dates back to 1971, when T. J. Dekker described a method for double-float arithmetic that splits a 64-bit number into two 32-bit components.

The Dekker method uses a high-term to carry the most significant bits and a low-term to capture the rounding error. While this approach loses roughly 5 bits of precision compared to native FP64, it allows for significantly higher throughput on cards with weak double-precision units. Researchers like Andrew Thall proposed specific algorithms for this type of emulation as early as 2007.

A more modern approach called the Ozaki scheme preserves full 64-bit precision for matrix multiplication. This scheme exploits Tensor Cores, which are specialized hardware units designed for the math used in AI. The Ozaki scheme splits FP64 numbers into lower-precision slices, performs the multiplication, and sums the results back into a 64-bit total.

Nvidia has begun to embrace this emulation approach officially. In October 2025, the company added support for the Ozaki scheme to its cuBLAS library. This allows developers to use fast FP8 and FP4 hardware to perform high-precision math, reducing the need for dedicated, physical FP64 silicon on the die.

The future of enterprise hardware

The latest Blackwell Ultra architecture signals a massive shift in how Nvidia builds enterprise GPUs. The new B300 enterprise card actually reduces physical FP64 performance compared to the previous B200 model. Nvidia is prioritizing NVFP4 tensor cores for AI workloads over traditional scientific math units.

The B300 features an FP64-to-FP32 ratio of 1:64, matching the ratio found in consumer GeForce cards. Peak double-precision performance on the B300 drops to just 1.2 TFLOPS, down from 37 TFLOPS on the B200. This marks the first time enterprise hardware has moved toward the constraints traditionally reserved for consumer products.

Nvidia claims it is not abandoning 64-bit computing entirely. The company plans to use low-precision emulation to supplement hardware units for high-performance computing tasks. This allows Nvidia to pack more AI-focused FP4 and FP8 resources into the same silicon area, maximizing revenue from the AI sector.

The old divide between consumer and enterprise silicon is not disappearing; it is simply moving to a different metric. While FP64 ratios are converging, the gap in low-precision performance is widening. The RTX 5090 offers a 1:1 ratio for FP16-to-FP32, but the B200 offers a staggering 16:1 ratio, ensuring that enterprise cards remain the only choice for massive AI clusters.