At the CES 2026 technology expo, a California-based graphics processing unit (GPU) startup named Bolt Graphics drew significant attention when showcasing their latest GPU platform, Zeus. This product targets gaming tasks, Computer-Aided Design (CAD), and High-Performance Computing (HPC) simulations. First announced a year ago, Zeus is built on the open RISC-V instruction set architecture and makes an incredibly bold promise: delivering path tracing performance up to 10 times higher than Nvidia’s “monster” RTX 5090. However, the tech community is still waiting for real-world evidence to verify this shocking claim.
Hardware configuration and connectivity of the Zeus platform
The prototype card displayed at CES made a strong impression with its support for a massive combined memory capacity of up to 384GB, including both LPDDR5X and DDR5 standards. Within this, the video memory (VRAM) capacity soldered directly on the board can reach up to 128GB. Furthermore, the device is equipped with up to four DDR5 standard SO-DIMM slots, allowing for flexible memory expansion and a high-speed 800Gbps memory interface. Despite possessing such “monstrous” specifications, the card’s maximum power consumption is capped at 225W, powered through a single 8-pin PCIe connector.
Regarding scalability and connectivity, Bolt Zeus is equipped with two PCI-Express 5.0 x16 slots supporting the modern CXL 3.0 standard. The displayed connectivity options include one DisplayPort 2.1a output port and one HDMI 2.1b port, meeting high-resolution display output needs. Notably, the card also features integrated ultra-high-speed network ports including 400GbE and 800GbE QSFP-DD. This equipment provides advanced networking capabilities, serving large-scale rendering tasks and complex data simulations effectively.
Additionally, an integrated BMC/IPMI controller on the board allows data centers to easily deploy multiple GPU units in rack-mount servers or specialized workstations. Bolt Graphics also confirmed that the Zeus architecture includes Texture Mapping Units (TMUs) and Raster Operations Pipelines (ROPs), although the company has not yet disclosed the exact technical specifications of these components.
Real-world performance and different configuration versions
According to the FP64 mathematical benchmarks released by Bolt Graphics itself, the entry-level version is the Zeus 1C, with a single processing unit capable of delivering path tracing performance 2.5 times higher than an RTX 5090. This version is equipped with 32GB of LPDDR5X memory providing 273GB/s bandwidth and can be expanded with up to an additional 128GB of DDR5 memory via two SO-DIMM modules at 80GB/s.
In the higher segment, the Zeus 2C version uses dual silicon with up to 128GB of LPDDR5X memory, claiming to provide 5 times the path tracing performance of Nvidia’s flagship GPU. Meanwhile, the highest-end version is the Zeus 4C, which uses four silicon chips—designed as a server platform rather than an independent discrete card—capable of delivering ten times the performance. The 4C configuration includes four I/O chiplets, up to 256GB of LPDDR5X memory, and a massive up to 2TB of DDR5 memory.
Bolt’s FP64 electromagnetic (EM) simulation benchmarks show that the Zeus 4C is up to 300 times faster than Nvidia’s B200 Blackwell GPU under controlled testing conditions. This is an extremely impressive figure for intensive computational tasks. However, it should be noted that Bolt Graphics has not yet released any benchmarks regarding rasterization capabilities or traditional rendering tasks. The company has also not announced an exact launch date, although they previously stated that Zeus is expected to be available on the market sometime in 2026. Therefore, users and experts still need more time to verify whether the actual product can meet the massive expectations set by the manufacturer.
