Up to 2.5 times faster than NVIDIA RTX 5090 in traceability: Bolt Graphics announces Zeus graphics cards

Bolt Graphics has unveiled the technical specifications of its Zeus series graphics cards. The documentation describes four models that, according to the developers, are capable of significantly outperforming NVIDIA’s flagship GPUs in ray tracing tasks.

The base model of the line, Zeus 1c26-32, is actually the foundation for all the others. Its GPU consists of two separate chipsets, one of which is responsible for I/O interfaces (including PCIe and network connections), and the other — for computing tasks.

The performance part of the GPU, in turn, is built on clusters with the following components:

• One scalar processor (one operation per clock cycle)
• FP64 cores for double precision operations
• Unnamed graphics accelerators
• Unified cache of 128 MB
• Separate router for data transfer between clusters

The memory subsystem is based on LPDDR5X and, depending on the model, can be soldered on the board from 32 GB to 256 GB. If this is not enough, the video memory can be expanded via two, four or eight SODIMM slots, depending on the card model. The other models in the — Zeus line, the 2c26-64, 2c26-128, and 4c26-256, are scalable versions of the base model, with two or four computing chips. In the case of the oldest model, they are also served by four I/O chips.

Despite the unusual architecture, the real interest is in the declared performance of the video cards. According to Bolt Graphics, even the base Zeus model is 2.5 times faster than the RTX 5090 performance in ray tracing: 77 Gigarays versus 32 Gigarays. And that’s with 80% lower power consumption — 120 watts versus 575 watts. It looks fantastic, but it’s worth noting that all the claimed characteristics currently exist only on paper. No tests have been conducted with real hardware, and all the data is obtained from software simulations.

Probably, Zeus graphics cards are focused on highly specialized tasks, because despite the advantages in FP64 and ray tracing, modern games are heavily dependent on FP32 performance. And here, the performance of prototypes is quite modest: even the top-end 4c26-256 has only 40 TFLOPS, while the RTX 5090 provides 105 TFLOPS. It is reported that the first hardware prototypes of the cards will appear in the fourth quarter of this year.

Source: PCGamer