The Evolution of GPUs: How Floating-Point Changed Computing

In the 1940s, physicist Richard Feynman famously raced an abacus expert through a series of calculations. The abacus handled basic arithmetic like addition and multiplication with ease. But when it came to division and square roots, Feynman pulled ahead. The abacus was limited to whole-number operations and relied on indirect methods like repeated subtraction and estimation, which made it slower and less precise.

Early computers had similar limitations. Though faster than an abacus, they relied on fixed-point arithmetic, a system that could only represent numbers within narrow ranges and burned through bit space just to show fractions. This made it difficult for them to handle tasks requiring both large and small values, like modeling weather patterns or simulating forces in physics.

Floating-point (FP) arithmetic changed that. It allowed computers to represent a wide range of values with greater precision and paved the way for the powerful GPUs we use today.

From fixed-point to floating-point: A necessary shift

CPUs and GPUs operate using binary representation. Every number, and its corresponding sign, is encoded using a series of 0’s or 1’s. Fixed-point numbers are the simplest form in this binary landscape. Each bit represents a multiple of two.

Similar to the abacus, this format has significant limitations. It doesn’t allow for fractional precision. To work around that, we started dividing binary strings into integer and fractional components, separated by a decimal point.

Floating-point numbers introduced a creative solution. Instead of using standard notation to represent numbers, floating-point uses scientific notation. Each number is defined by a sign, an exponent, and a mantissa:

Value = -1sign* mantissa * 2exponent

This structure allowed computers to represent a much wider range of values. Compared to fixed-point, floating-point offers more range and precision:

32-bit Floating-Point (IEEE 754 Standard)

• • Range: −3.4×10^38 to +3.4×10^38
  • Precision: 7 decimal digits of precision

32-bit Fixed-Point (Q16.16 format)

• • Range: −3.28×10^4 to +3.28×10^4
  • Precision: 5 decimal digits of precision

In binary form, this layout is also efficient. Instead of splitting a bit string into just integer and fractional components, it is now divided into exponent and mantissa components.

The example above shows FP32 representation, which is the common reference point for measuring precision. Other representations, like FP64, offer twice the precision and are referred to as “double precision.” Similarly, FP16 offers less and is referred to as “half precision.”

Programmers often optimize their workloads based on the balance between speed and precision, utilizing different floating-point representations depending on their tasks.

Common floating-point formats and use cases

• • FP64 (Double Precision): Used for high-precision tasks like molecular dynamics or physics simulations, where minor inaccuracies can lead to errors. It uses 1 sign bit, 8 exponent bits, and 23 fraction bits. It’s highly precise but slower to compute.

• • FP16 (Half Precision): Optimized for speed and memory efficiency. Best suited for real-time graphics and AI applications that don’t require high precision. It uses 1 sign bit, 5 exponent bits, and 10 fraction bits. It is faster than FP64 and FP32, but the trade-off is less precision.

• • BF16 (Brain-Float 16): Common in deep learning, where approximate calculations are acceptable. It uses 1 sign bit, 8 exponent bits, and 7 fraction bits. Developed by Google Brain, it expands the range of FP16 while sacrificing some detail because of its limited mantissa precision.

• • TF32 (Tensor Float): Designed by NVIDIA for deep learning, TF32 retains the range of FP32 while reducing precision to improve speed on matrix-heavy tasks like neural network training. It uses 19 bits, trimming the fraction from 23 to 10 bits, and approximates FP32 values to accelerate performance without losing numerical range.

With these floating-point optimizations, we are able to make new advancements in science and engineering. From SpaceX’s reusable rocket, that calculates minute angles from miles away to financial models that forecast economic trends to the smallest decimal point. Computers can now reach numbers that were once unimaginable.

The rise of GPUs: A new era of computation

Floating-point formats may sound like abstract math theory, but their impact is anything but theoretical. Take SpaceX, a company whose success literally depends on precision, speed, and scalability.

SpaceX’s reusable rocket wouldn’t be capable of doing the level of precision required on standard CPUs alone. While CPUs have evolved to support floating-point arithmetic, they’re fundamentally limited in how much they can process in parallel. Tasks like graphics rendering, scientific modeling, and AI often interface with thousands, or even millions, of floating-point operations simultaneously, especially calculations involving matrices.

This demand led to the rise of graphics processing units (GPUs). GPUs are built specifically to accelerate parallel computation. Unlike CPUs, GPUs are equipped with specialized components like Tensor Cores and Fused Multiply-Add (FMA) units that are optimized for floating-point calculations.

Tensor Cores, introduced by NVIDIA, are especially impactful. They perform addition and multiplication at the same time without rounding between operations, a capability that enables fast, accurate matrix multiplication. This function is foundational to modern AI and deep learning, making Tensor Cores a critical driver for today’s AI breakthroughs.

Hardware support for floating-point formats: NVIDIA GPUs

Even with hardware acceleration, not every GPU supports every floating-point format equally.  Some platforms like, the NVIDIA A100 and H100 GPUs support a flexible range of formats including FP64, FP32, FP16, BF16, and TF32, making them highly versatile. Others, like NVIDIA A800, are focused on precision-heavy workloads and optimized primarily for FP64 and FP32. Meanwhile NVIDIA’s Ampere Architecture’s Tesla P100 emphasizes FP32 and FP16 for AI calculations.

This variation makes hardware selection a critical part of system design. For tasks like 3D rendering or physics simulations, a personal computing platform like the Dell Pro Max Tower T2, equipped with NVIDIA RTX PRO 6000 Blackwell GPUs, delivers the performance and precision needed for intensive compute work. In AI research settings, teams often turn to the supports the Tesla P100, and is better for handling data-intensive modeling with FP16 and FP32 optimizations.

SpaceX, for example, uses NVIDIA Jetson Orin NX GPUs within their low-Earth orbit satellites. These NVIDIA GPUs process high-resolution Earth imagery, detect environmental changes, monitor weather systems, and track ship or aircraft locations autonomously. By performing these calculations on-device, the satellites reduce the need to send data back to Earth for processing. Jetson Orin NX is optimized for FP16, enabling real-time processing and speed over the high precision of FP64 or FP32.

SpaceX’s Transporter–11 Payload (Image credit: SpaceX)

The ongoing impact of floating-point precision

The evolution of floating-point numbers is a prime example of the creative problem-solving behind modern computational systems. From the abacus to GPUs, advances in number representation have continually redefined what is possible.

Today, GPUs help power everything from SpaceX’s autonomous satellites to real-time AI interfaces. As our workloads grow more complex, AI continues to demand faster mathematical operations with high precision. Floating-point formats, and the hardware designed to process them efficiently, remain central to that progress.

Platforms like the Dell Pro Max Tower T2 and the flexibility and performance needed to accelerate floating-point calculations across different workloads.

Ready to unlock the power of floating-point precision?

From scientific modeling to deep learning acceleration, Dell Pro Max workstations with NVIDIA RTX PRO GPUs are engineered to handle today’s most demanding compute workloads, locally, efficiently, and with unparalleled flexibility.

Learn More:

• Dell Pro Max – Explore the full portfolio of AI-optimized systems
• NVIDIA RTX PRO Blackwell Workstation Edition – The ultimate AI and graphics performance