When Bolt Graphics formally announced its Zeus GPU platform earlier this year, the company briefly stated its upcoming flagship graphics processor can be around 10 times faster than Nvidia's GeForce RTX 5090 in ray tracing workloads. But the startup never previously demonstrated actual benchmark results. Recently, the company quietly added a graph showing the relative ray tracing performance of Zeus GPUs compared to existing graphics cards, which appears to be quite impressive. However, there are a number of things to note about these simulated test results.
The graph that Bolt Graphics shows is the ray-triangle intersection budget, which is measured in ray-triangles (tris) per pixel per frame. This expresses how much raw ray tracing work a GPU can do in terms of ray–triangle intersection tests for every pixel in a single rendered frame, while maintaining a 120 FPS framerate, at 3840x2160 resolution. This number is useful as a theoretical ceiling for geometry and lighting complexity a GPU can handle in ray- or path-traced rendering. And it's in line with Bolt's marketing message, that since modern GPUs do not have enough ray tracing and path tracing performance, game developers do not use these technologies extensively.
(Image credit: Bolt Graphics)
Based on Bolt's internal simulations using its own micro-benchmark, its forthcoming quad-chiplet Zeus 4c (which is not a graphics card, but is a server) should be 13 times faster than Nvidia's existing flagship GeForce RTX 5090 (which is the best graphics card today), whereas a single-chiplet Zeus 1c (which is a card) should be 3.25 times faster than Nvidia's range-topping offering. In fact, even the entry-level Zeus can enable over 25 samples per pixel while sustaining 120 FPS in a 4K resolution.
A higher value in this micro-benchmark means the GPU can sustain more samples, denser geometry, or both, without dropping below the target frame rate (120 FPS in this case). Bolt does not disclose anything about its micro-benchmark or how it simulates the performance of its hardware and how it gets comparative performance for AMD, Intel, and Nvidia GPUs, so we can't really say much about these results.
Typically, the workload in synthetic tests is controlled: Rays are cast in predictable patterns against fixed triangle sets, and the acceleration structures are static and well-optimized. This test, however, produces a clean, high number that reflects the GPU's raw intersection throughput under ideal conditions, which is good enough if one only wants to learn about the theoretical limits of a GPU in this specific workload.
In a real game engine, that number is affected by many variables. Acceleration structures may need to be rebuilt or updated for dynamic objects. Rays may be incoherent (due to reflections and refractions), and scene triangle density will vary by frame. The engine's traversal algorithms, shading pipeline, and memory layout also influence how many triangles each ray must test, which changes the effective rate. As a result, the practical ray triangle rate in games is lower than in synthetic benchmarks, and can fluctuate heavily depending on scene content and rendering settings.
As for actual performance in real games, this depends on multiple factors, like shader and memory performance. The shading performance of Bolt's Zeus 1c is 10 FP32 TFLOPS, while Zeus 2c doubles it to 20 FP32 TFLOPS, which is dramatically lower than the 105 FP32 TFLOPS offered by the GeForce RTX 5090.
As for memory, on the one hand, Bolt's Zeus 2c (which is the best the company can offer in an add-in-card form-factor) can carry 128 GB of LPDDR5X memory onboard, which is significantly more than GeForce RTX 5090's 32 GB of GDDR7 memory. But on the other hand, Nvidia's graphics board boasts a memory bandwidth of 1.8 TB/s, whereas Bolt's Zeus 2c can only offer 725 GB/s.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Keeping in mind that Bolt Graphics itself uses simulations to test the capabilities and performance of its GPUs, it's impossible to estimate the performance of Zeus 1c, Zeus 2c or Zeus 4c in real-world applications. Then again, these graphics solutions are projected to be available as samples for developers in 2026 and mass-produced in 2027. By that time, AMD and Nvidia will certainly have released their next-generation GPUs. So it will make more sense to compare Zeus 1c and 2c to the performance of 2027's GeForce and Radeon graphics cards, not the best GPUs of 2025.
Follow Tom's Hardware on Google News to get our up-to-date news, analysis, and reviews in your feeds. Make sure to click the Follow button.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
FWIW, I did some digging into what info I could find, back in the comment thread of the prior announcement. The later posts in this thread are the more informed takes:
https://forums.tomshardware.com/threads/startup-claims-its-zeus-gpu-is-10x-faster-than-nvidias-rtx-5090-bolts-first-gpu-coming-in-2026.3875286/post-23471706
Basically, they've got some RISC-V cores, with vector extensions, and some dedicated RT engines.
Their claims might have some legitimacy, but I think Anton's skepticism is well-placed. It's very hard for a chip startup to beat the established players, due to how fast the industry moves and how many more resources the big companies have to throw at the venture.
