Bolt Graphics Zeus Architecture
We are not getting a full block diagram yet, but this is a RISC-V RVA23 out-of-order scalar core. There are then vector cores and accelerators. Something that is at least slightly reminiscent of some other startups we have seen is the use of RISC-V. Instead of building its own ground-up ISA, it can use RISC-V plus then add its own special sauce.
It feels a bit like the Zeus 1c26-032 is a single chiplet part. We have DP 2.1a, HDMI 2.1b, and then some really neat features. Namely, there is a RJ45 BMC port as we have seen on DPUs, servers, and other components, but less commonly on GPUs. There is a QSFP-DD port for 400GbE, and two PCIe Gen5 x16 slots. Bolt Graphics can apparently do PCIe Gen5 / CXL 3.0 and Ethernet for a lot of flexibility. On the memory side, there is 32GB of LPDDR5X and up to 128GB of DDR5. The LPDDR5X here is being used because it is cheaper and higher capacity than either HBM3 or GDDR7.
The Zeus 2c26-064/128 is the two chiplet module. The 64 or 128 seems to be the capacity of the LPDDR5X.
Scaling up, and very interesting here, is the Zeus 4c26-256. Each chiplet gets an I/O chiplet and 64GB Of LPDDR5X. So there is a lot of OSFP 800GbE connectivity along with PCIe Gen5 and memory. Something to keep in mind here is that we often say that PCIe Gen5 x16 is around a 400GbE port. So with two Gen5 x16 ports, and the ability to do either PCIe/ CXL or Ethernet, that seems to track the diagram here. With 256GB of LPDDR5X and 512GB of DDR5 DIMMs per chiplet for 2TB, that gives us the 2.2TB of memory total on the GPU.
Taking a look at the smaller chiplet, this is the Zeus 1c26-032 card. You can see the two DDR5 SODIMM slots on the card as part of the expandable memory.
We can also see the BMC port, HDMI port, DisplayPort, and the QSFP-DD 400GbE port.
Here are the specs. We can see that Bolt is aiming at a relatively lower power envelope and has more local cache.
Here are the specs on the larger 2c26 platforms. Worth also noting here is that Bolt Graphics is targeting lower memory bandwidth, but higher memory capacity in its designs.
Taking a step back on the big 500W Zeus 4c26 part, the 6x 800GBE is really interesting. That is a lot of connectivity coming out of a GPU. Putting that into perspective, that is ~12x PCIe Gen5 x16 lanes worth of bandwidth. It also means many fewer NICs being sold since these GPUs have them built-in.
Software is going to be a big part of this. It feels like the Bolt Graphics folks have a vision, but they need to work on the software side.
It also feels a lot like these are less meant to be deployed as a single GPU in a system, and instead are designed to be clustered. A small cluster might be two GPUs.
Adding a switch would mean a large memory footprint in a smaller power budget.
On the larger scale, the goal is to scale-out using 800GbE. We have looked at 800GbE switches like the Marvell Teralynx 10 51.2T 64-port 800GbE switch and this year we will have the 102.8T switches for 128x 800GbE links.
In terms of availability, early access to developer kits is scheduled for Q4 2025 and then scaling in Q4 2026. That gives folks like NVIDIA quite a decent amount of time to react if they feel it is necessary.
For our readers, it looks like maybe we will see more of these in the future.
Final Words
This feels like an announcement where we really want to see product. The fact that we are still a few months away from early developer kits makes this feel a bit early. On the other hand, if Bolt Graphics has Zeus competing with a combination of NIC and consumer GPU, while offering more memory then it could be a really neat combination. I am always a bit skeptical of products until we see them live. Hopefully, we can show you this in action as the hardware becomes available.