Startup GPU maker Bolt Graphics is on track to deliver development kits to lead customers by the end of the year, having taped out its first chip last month. By the end of 2027, the company expects to begin shipping PCIe cards based on a production version of its Zeus GPU. This schedule represents a one-year delay compared with its 2025 plans, but a tape-out is a milestone toward erasing our skepticism about the company.
Bolt is targeting high-performance computing, visualization such as moviemaking, and other applications that have employed data-center GPUs for their FP64 throughput. Those include data-center GPUs from AMD and Nvidia. Market leader Nvidia, however, has redirected its focus toward AI, paring down its GPUs’ FP64 capabilities, creating an opening for a new entrant such as Bolt. Bolt, however, is a small company, having fewer than 100 employees and no publicly disclosed chip-design star known for complex projects such as a GPU. Thus, taping out indicates the company indeed has a chance to challenge the incumbents.
Bolt’s Cost Advantages
Fabricated in a 12 nm process, the taped-out Bolt Zeus GPU is a development vehicle. Bolt plans to build the production version in a 5 nm node, which will put its transistor density and power behind that of GPU rivals but should be less expensive to produce. Per a 2025 XPU.pub article, Zeus is a chiplet-based design, helping it to achieve high throughput despite the older process. Moreover, Bolt places the chiplets on an organic substrate, like AMD has employed with many Epyc and Ryzen models, instead of using a more costly silicon substrate or bridge technology. The chiplets aren’t reticle busters like Nvidia makes but instead are quarter-reticle size, which should reduce product costs by raising yields.
Based on RISC-V, Bolt’s Architecture Differs
Bolt withholds details of the Zeus architecture, but we know that it departs from typical GPU designs. Like the ill-fated Intel Larrabee project, Zeus has CPUs with wide vector units at its heart. The startup’s vector units are much wider, however. Although Bolt withholds details, we speculate that total vector-lane width per CPU could be as high as 16 kbits.
Bolt has licensed a RISC-V core from Akeana, likely one of its high-end 5000-series CPUs. Akeana is a proponent of simultaneous multithreading (SMT), which can hide memory-access latency and speed up context switching. Zeus also has large caches, which, combined with wide vector units and multithreading, could yield a performance profile similar to a traditional single instruction, multiple thread (SIMT) GPU.
Unlike Larrabee, Zeus integrates hardware engines for graphics functions, raising performance and power efficiency compared with a purely CPU-based design. A Bolt goal is to provide much more ray-/path-tracing performance than GPUs offer today, initially to address visualization customers and ultimately to enable solely path-traced games. Hardware will also accelerate HPC tasks inefficiently handled by a CPU and various graphics functions. We expect the latter to include texture mapping and final rasterization.
Zeus’s Programming Model Bridges CPU and GPU Approaches
Zeus’s CPU-centric approach enables programmers to code in the familiar C language instead of using a specialized graphics framework such as Cuda or OpenCL. Developer tools can speed up porting Cuda routines and automatically vectorize C code. Bolt will provide OpenCL emulation and other HPC and graphics libraries. In the near term, however, Bolt must finish drivers and other basic enablement software before it can ship development cards.
Bottom Line
Bolt is a small company taking on large competitors, initially targeting a niche that has become a second priority for those companies. Despite schedule slips, it’s progressing toward having a production-ready chip. The company’s GPU architecture is unusual for a GPU but should still be developer friendly. Bolt’s next step is to bring up the chip, a milestone that will help it secure additional funding. It must then deliver development boards within a year to pull customers away from entrenched incumbents, eroding any lingering skepticism.