0

The chip’s huge memory could offer a unique niche for AI inference workloads

In sum – what we know:

• New contender: Startup Bolt Graphics is developing Zeus, a RISC-V GPU designed specifically for professional rendering and scientific computing rather than gaming.
• Memory advantage: The architecture prioritizes capacity with expandable memory up to 2.25TB — which is ideal for large AI models.
• Availability: No physical silicon has been released — developer kits will roll out this year, with mass production targeted for 2027.

Nvidia’s leadership on AI silicon has looked pretty much unshakeable for years now, but a startup called Bolt Graphics thinks it’s spotted some vulnerabilities. The company is gearing up to launch Zeus, a RISC-V GPU that approaches graphics processing from a different angle — and that difference might carve out real territory in a market hungry for anything that isn’t Nvidia.

Bolt is generally aiming Zeus at professional rendering, scientific computing, and film production, steering clear of mainstream AI training as an explicit goal. The chip’s architecture is unusual, though, particularly around memory expansion, and that could open up unexpected doors for certain AI applications. Of course, whether any of this translates to actual adoption is another matter entirely, especially since no physical silicon exists yet to back up what the company is promising.

Memory versus muscle

Bolt isn’t pitching Zeus as an AI chip exactly, but the architecture does offer some compelling advantages for specific workloads, especially when it comes to large language model inference. The headline feature here is memory capacity — Zeus supports up to 384GB in a standard PCIe form factor, scaling all the way to 2.25TB per unit in a 2U server setup. The concept of aGPU with an expandable memory architecture is pretty unusual — a design decision that could unlock real value for organizations running inference on massive models where capacity trumps raw compute.

For enterprises wrestling with how to cram large language models into GPU memory, this kind of headroom could be very helpful. Today’s high-end professional and consumer GPUs hit memory ceilings well below these numbers, which forces organizations into distributing models across multiple cards or simply running smaller models. Zeus could theoretically hold an entire massive LLM in memory on a single unit, streamlining deployment and potentially cutting latency for inference.

There is a catch, though. Zeus makes a significant bandwidth trade-off that cuts against the memory advantage. The Zeus 2c variant delivers roughly 725 GB/s of memory bandwidth, versus 1.8 TB/s for the RTX 5090. Bolt’s counter is that Zeus actually offers better bandwidth per core than both the RTX 5090 and AMD’s 7900 XTX, but that lower aggregate throughput could hurt in bandwidth-heavy AI workloads.

The chip targets 20 FP64 TFLOPS for scientific computing, which plants it squarely in high-performance computing territory rather than the AI training space where Nvidia and AMD are fighting with tensor-optimized designs. 

Claims vs reality

Bolt’s marketing has certainly turned heads, especially claims that Zeus runs “10x faster” than Nvidia’s RTX 5090. But these numbers deserve serious scrutiny. Every performance metric published so far comes from internal simulations — so far, there’s no physical chip, no independent testing. The distance between simulated and real-world performance can be enormous, and mass production is still two years out, according to the company.

The performance gains also attach to highly specific workloads. Bolt says Zeus delivers 13x faster ray tracing than the RTX 5090 and 300x faster electromagnetic wave simulations compared to Nvidia’s B200 Blackwell. The company is touting 4K path tracing at 120 fps with over 25 samples per pixel — numbers that would be impressive for film production and professional visualization.

On traditional shader performance, though, Zeus is modest. The chip offers 10-20 TFLOPS across variants, stacked against 105 TFLOPS for the RTX 5090. This is a deliberate architectural choice — sacrificing general-purpose shader throughput for specialized path-tracing and memory capacity.

The timeline adds another layer of risk. Developer kits were expected in late 2025, but that deadline has come and gone, and developer kits are now slated for sometime in 2026, with mass production targeted for 2027. By then, Zeus will be competing against next-generation silicon from Nvidia, AMD, and Intel — not the current hardware it’s being benchmarked against today. Two years is a long time in semiconductors, and the competition won’t be sitting still.

The ecosystem challenge

Hardware performance may not even be Bolt’s biggest hurdle. The real obstacle is software ecosystem maturity. While Nvidia’s decision to support RISC-V as a host architecture validates the instruction set, it doesn’t solve the library gap. Bolt cannot run CUDA natively. Instead it’ll need to rely on translation layers or its own toolchain, neither of which has the decades of optimization that Nvidia’s proprietary stack enjoys. AMD’s ROCm, though less entrenched than CUDA, still offers a far more mature alternative for developers.

The chip uses a RISC-V multi-chiplet design with integrated CPU cores capable of running Linux, making Zeus essentially a self-contained computing platform rather than a mere peripheral. By running the OS directly on the card, Zeus bypasses typical PCIe bottlenecks. It allows Bolt to tap into the open-source RISC-V tooling that Nvidia is now inadvertently funding, while attempting to bridge the compute gap with “Glowstick,” their proprietary path-tracing renderer.

The question is execution. A stable RISC-V host environment is one thing, but a fully optimized compute pipeline is another. For professional rendering and HPC adoption, Zeus needs to work seamlessly with existing workflows. If developers have to rewrite their kernels to move away from CUDA, they likely won’t. Intel’s recent GPU struggles prove that hardware specs are irrelevant without the infrastructure underneath.