Bolt Graphics Zeus The New GPU Architecture with up to 2.25TB of Memory and 800GbE

In the world of GPUs, it has been NVIDIA, AMD, and Intel for some time. There are other projects in China especially, but companies breaking into the US market has been rough. Instead of just building low-end graphics and hoping to scale, Bolt Graphics is solving a specific challenge when it comes to high-end GPU computing and is doing so elegantly.

Here is one of the craziest overview slides you will see for GPUs in 2025. It does not have a single mention of AI.

Instead, this is designed for rendering, gaming, and HPC. As you might be thinking, that is a lot of area to cover even leaving out AI. For example, companies like NVIDIA have outreach to game developers at scale so when new titles come out, they (hopefully) run well on GPU architectures. Still, perhaps the message here is that if you want to focus on higher-precision compute, then a different architecture can be used.

Bolt Graphics is trying to do Path Tracing in hardware.

One of the benefits to having better hardware is that artists can see much closer representations of what final renderings will look like. If you have ever heard Pixar, or other studios discuss what they do, they generally have huge render farms that take a long time to render each frame accurately.

We will go into architectures later, but the 2.5x speedup for the 1c part is the single chiplet GPU with 400GbE and around 120W TDP apparently.

Of course, the idea is that you can scale the number of GPUs and get to higher performance levels. Something that is notable here is the comparison to the NVIDIA GeForce RTX 5090, not a NVIDIA RTX 6000 Ada, H200, or something similar.

Bolt Graphics has a real time path tracer called Glowstick.

This may not be a completely fair statement, but in the world of GPUs, the software development and support side is a huge deal. Something like a path tracer is a specific area that Bolt could develop for customers.

We are going to let folks read through the Glowstick features on these slides.

This is one of Bolt’s key features for customers.

Here are some demo renders on an Alpha build of Glowstick.

Another use case for Zeus is HPC simulations.

This FP64 performance slide we need to take into some context. NVIDIA years ago decided that FP64 was a data center GPU feature not a consumer GPU feature.

One application is electromagnetic wave simulation where Bolt is claiming Zeus is much faster.

Aside from speed, Bolt is saying that it can handle larger simulation spaces, especially using the 4c model.

In the announcement deck, Bolt Graphics had a number of simulations.

Here is another one.

And another.

Here is a FFT simulation. FFT simulations are a big area.

Here is a silicon photonics crystal waveguide simulation.

Here is another way to look at Zeus in the HPC space. Bolt Graphics has a focus on FP64 and FP32 compute and is looking to do things at a different performance, price, and capacity space than many of the GPUs today.

Next, let us get to the architecture.