Bolt Challenges Nvidia With a Focus on Cutting-Edge Graphics

Darwesh Singh thinks Nvidia has a weakness.

The last decade of Nvidia’s history was among the most consequential stories in technology, ever. The company’s stock price has increased over 200-fold since 2016, and, according to data tracked by Epoch AI, deployed Nvidia AI compute capacity has surged to 225 times greater than the first quarter of 2021.

Yet Nvidia may in some ways be a victim of its own success. Its dominance in AI has led to GPU designs that prioritize tensor units and low precision math. These decisions make sense for AI, but less so for some creative, scientific, and industrial work.

Singh’s five-year-old startup, Bolt Graphics, sees an opportunity to build a GPU specifically for these use cases. Jill Mueller, Bolt’s chief marketing officer, puts it bluntly. Nvidia has “a fundamental lack of understanding of their customer,” she says. “They just throw stuff at you, and there you go.”

Bolt aims for this potential weak spot with Zeus, a GPU that will be sold as both a PCIe card for desktop workstations and, for those who require more performance, a rack-mountable server containing four Zeus GPUs (for up to 96 per rack).

While Nvidia goes low-precision, Bolt goes high

Jacob Feldgoise, senior data research analyst at Georgetown University’s CSET, has also noticed a shift in Nvidia’s recent hardware.

“AI is sucking the computational units used for high-precision workloads out of that hardware,” he says. “If you look at Nvidia’s highest performance GPUs, generation to generation, a greater share of the hardware has been allocated to low-precision compute, as opposed to high-precision compute, which is generally needed for scientific computing.”

Precision refers to how many bits a GPU uses to represent each number. High-precision formats like FP64 (64-bit floating point) preserve more digits and a wider range, while FP16 and INT8 sacrifice precision for speed. Recently, Nvidia introduced a new 4-bit number format, NVFP4, to accelerate AI workloads, which generally tolerate low-precision math.

But some tasks require precision. Singh cited geographical information systems, such as Esri’s ArcGIS, as an example. When rendering the planet on a GPU, low-precision arithmetic applied to large coordinate values can introduce errors that cause objects to drift.

Because Zeus, unlike so many other GPUs, is not designed primarily for AI, its design makes FP64-native vector cores a focus and allocates a large share of silicon to them.

“[Nvidia and AMD] make a conscious trade-off to allocate more die space to matrix multiplication and tensor units and less towards fixed function hardware. We decided to allocate the die space a bit differently,” says Singh.

Rasterization is out, path tracing is in

A focus on FP64 isn’t the only way Bolt differs from the norm. Zeus is also built to render graphics with path tracing instead of rasterization.

Rasterization is the traditional method of high-performance 3D rendering. It projects 3D triangles onto a pixel grid and uses mathematical abstractions to determine the correct color for each pixel. Path tracing instead does the equivalent of shooting rays from a camera to simulate how light should bounce and interact. It delivers more accurate lighting but is computationally expensive.

As with high-precision math, Bolt believes it can find an edge by placing more emphasis on path tracing than do today’s GPUs. Rasterization is supported by Zeus but significantly scaled back; Singh estimates that Zeus’ raster performance is about half that of a comparable Nvidia card.

Bolt’s fresh arrival to the GPU arena also allows the company to take a clean sheet approach unburdened by legacy support. This differs from Nvidia and AMD, which must integrate path tracing alongside rasterization in a way that can support numerous existing application-programming interfaces (APIs) and applications.

Bolt claims that a server rack with 28 Zeus GPUs will deliver real-time path traced performance equivalent to 280 Nvidia RTX 5090 GPUs. The aim is for this configuration of Zeus hardware to support real-time path tracing that simulates up to 20 “bounces”—a reflection or collision of the simulated light—at 4K resolution and 30 frames per second. This is a high degree of accuracy required for professional rendering workloads; for comparison, even the most graphically attractive path traced games simulate just a few bounces.

Can a start-up really launch a new GPU?

There’s a logic to Bolt’s approach. Nvidia and AMD are focused on AI, but GPUs are still useful for many tasks besides AI. However, Bolt will need to overcome two key technical hurdles.

The first is production. Cutting-edge silicon production is in short supply and leaders like Nvidia have most leading-edge production capacity tied up. The Zeus GPU will instead be fabricated on TSMC’s older N5 process node. Bolt is betting that an older process node will keep Zeus competitive with Nvidia on price.

Bolt may also find it challenging to convince users that an unproven GPU is a safe bet. Driver support for software is always a headache in the GPU arena—just ask Intel—and the use cases that might benefit Zeus’ high precision and path tracing will also require reliable drivers.

Bolt plans to address this by launching with support only for specific applications. “We know that PC gaming is a huge segment,” he says. “But our approach is we want to target professional, creative, and high-performance compute first.” The company is working with software companies including Blender, Autodesk, and SideFX.

Bolt

, and it is now focused on bringing the GPU to production by the fourth quarter of 2027.