Bolt Graphics unveils Zeus, a specialized GPU targeting path tracing, CAD, and HPC applications with expandable memory up to 384GB and integrated RISC-V cores. The company claims Zeus 4c delivers 13× RTX 5090 ray-tracing performance, while Zeus 1c provides 3.25× improvement. However, these claims rely entirely on internal simulations. Zeus trades traditional shader performance (10–20 TFLOPS versus RTX 5090’s 105 TFLOPS) for ray-tracing optimization. Developer kits arrive in 2026, with mass production in 2027, creating uncertainty about real-world performance versus ambitious projections.
Bolt Graphics introduces Zeus, a specialized GPU targeting path tracing, CAD workloads, HPC simulations, and film/TV production. The company positions this architecture as fundamentally different from traditional consumer graphics cards, focusing on areas where conventional GPUs struggle to deliver adequate performance.
(Source: Bolt Graphics)
Memory and connectivity architecture
The Zeus GPU features an expandable memory system that combines soldered LPDDR5X memory ranging from 32GB to 128GB, with additional DDR5 SO-DIMM slots. Depending on the configuration, users can install either two or four SO-DIMM modules to achieve a maximum memory capacity of 384GB. The Zeus 2c configuration delivers 725 GB/s of memory bandwidth through this hybrid approach.
Beyond traditional video outputs, Zeus incorporates a 400 GbE QSFP-DD networking port that enables direct GPU-to-GPU communication in render farms without requiring separate network interface cards. The design also integrates RISC-V CPU cores capable of running Linux, creating a self-contained computing platform. Power delivery utilizes the standard eight-pin PCIe connector, while maintaining consumption under 400W.
Performance specifications and claims
Bolt Graphics claims Zeus can deliver 4K path tracing at 120 fps with over 25 samples per pixel, addressing the visual quality compromises developers typically make when implementing real-time ray tracing. For scientific computing applications, the GPU targets 20 FP64 TFLOPS while claiming 300 times faster performance than the B200 in electromagnetic wave simulations without accuracy reduction.
Figure 1. Ray-triangle benchmark comparison. (Source: Bolt Graphics)
The company plans three architectural variants: Zeus 1c uses a single chiplet design suitable for add-in cards, Zeus 2c employs dual chiplets with 128GB onboard memory, and Zeus 4c creates a quad-chiplet server platform for high-performance computing workloads.
Ray-tracing performance analysis
Bolt Graphics recently published benchmark data comparing Zeus against current hardware using ray-triangle intersection measurements. Their microbenchmark evaluates the number of ray-triangle intersections each GPU can process per pixel per frame while maintaining 120 fps at 4K resolution. This metric indicates the complexity ceiling for geometry and lighting that each GPU can handle in ray-traced or path-traced rendering scenarios.
According to their internal simulations, Zeus 4c delivers 13 times the ray-tracing performance of the GeForce RTX 5090, while Zeus 1c provides 3.25 times the performance of Nvidia’s current flagship. Even their entry-level configuration maintains over 25 samples per pixel at 120 fps in 4K resolution.
However, Bolt Graphics does disclose its benchmark methodology and explains how it simulates competitive hardware performance, making independent verification impossible. The company relies entirely on internal simulations rather than actual hardware testing for these performance claims.
Bolt has disclosed its methodology in our public documentation. It’s running on real hardware using industry-standard methodologies also used by Nvidia, Intel, AMD, Qualcomm, Arm, Tenstorrent, SambaNova, Groq, Graphcore, and NextSilicon, and they have public demos available on YouTube. There is also a demo of an Apollo Electromagnetic Wave Simulator.
Real-world performance considerations
The synthetic nature of Bolt’s benchmarks creates important limitations when extrapolating to practical applications. Their controlled test environment uses predictable ray patterns against static triangle sets with optimized acceleration structures, producing clean measurements that reflect theoretical throughput under ideal conditions. But, you need to establish some ground truths, and they’ve done that.
Real-time game engines introduce numerous variables that typically reduce practical performance below synthetic benchmark results. Dynamic objects require constant acceleration structure updates; reflections and refractions create incoherent ray patterns; and triangle density varies significantly between frames. Engine-specific traversal algorithms, shading pipelines, and memory layouts further influence how many triangles each ray must test, creating fluctuating performance that differs substantially from synthetic measurements.
Compute performance trade-offs
While Zeus excels in ray-tracing operations, traditional shader performance appears more modest compared to current hardware. Zeus 1c delivers 10 FP32 TFLOPS, Zeus 2c doubles this to 20 FP32 TFLOPS, while the GeForce RTX 5090 provides 105 FP32 TFLOPS. This difference suggests Zeus optimizes specifically for ray-tracing workloads rather than general-purpose graphics computation. Its increased amount of cache per core helps to offset the decreased off-chip memory bandwidth.
The memory architecture presents similar trade-offs. Zeus 2c offers substantial capacity with 128GB compared to the RTX 5090’s 32GB of GDDR7 memory, but achieves lower bandwidth at 725 GB/s versus 1.8 TB/s. This configuration favors workloads requiring large memory footprints over those demanding high memory throughput. However, Bolt’s memory bandwidth per core is actually beyond those of the 5090 and 7900 XTX.
Market timeline and competitive context
Bolt Graphics targets developer kit availability in 2026, with mass production following in 2027. This timeline means Zeus will compete against next-generation architectures from AMD and Nvidia rather than current 2025 hardware, making present performance comparisons less relevant for actual market positioning.
The two-year development window also introduces uncertainty about whether Bolt Graphics can deliver on their performance claims when transitioning from simulation to actual silicon. Without independent hardware testing, the gap between projected and realized performance remains unknown.
Technical assessment
Zeus represents an intriguing approach to specialized GPU design, prioritizing ray-tracing throughput and memory capacity over traditional rasterization performance. The expandable memory system and integrated networking capabilities target professional rendering workflows where memory capacity and inter-GPU communication matter more than raw shader throughput.
The architecture’s focus on path tracing addresses genuine limitations in current GPU designs, where developers must compromise visual quality to maintain acceptable frame rates. If Bolt Graphics delivers on their performance claims, Zeus could enable new levels of real-time ray-traced visual fidelity.
The architecture’s ultimate success will depend on how well actual hardware matches these ambitious projections when independent testing becomes possible in 2026.There’s a great video of the AIB here. They also shared how others can build renderers to achieve the performance they achieved.
YOU LIKE THIS KIND OF STUFF? WE HAVE LOTS MORE. TELL YOUR FRIENDS, WE LOVE MEETING NEW PEOPLE.