Thursday, March 6th 2025
Bolt Graphics Announces Zeus GPU for High Performance Workloads
Bolt Graphics announces Zeus, a completely new GPU design for high performance workloads including rendering, HPC, and gaming. Zeus addresses performance, efficiency, and functionality limitations with legacy GPUs.
Zeus is orders of magnitude faster than any other GPU in key workloads. Users can gain 10x in rendering performance, 6x in FP64 HPC workload performance, and 300x in electromagnetic wave simulations. Users running these types of demanding workloads need access to large amounts of memory. Bolt brings expandable memory to GPUs, for the first time, which allows users to increase their memory up to 384 GB in a PCIe card, and up to 2.25 TB per Zeus in a 2U server. A rack of Zeus 2U servers can be configured with up to 180 TB of memory, 8x larger than legacy GPUs.Historically, increasing performance requires more energy usage. Zeus reverses this long-term trend by reducing energy consumption while increasing performance. Reducing the impact GPUs have on the environment is one of Bolt's core values as a new entrant in the GPU space.
Zeus will be available in multiple form factors including PCIe cards, servers, and the cloud. Bolt plans on expanding Zeus into smartphones, tablets, laptops, consoles, and cars over the coming years, providing a unified GPU architecture across various platforms.
Glowstick: Real Time Path Tracing
Along with Zeus, Bolt Graphics is announcing Glowstick, a real time path tracer for rendering customers. Path tracing simulates complex light interactions that happen in the real world, but are too computationally intensive for legacy GPUs to simulate in real time. Glowstick will revolutionize industries like film, architecture, product design, and game development by enabling real time path tracing so users can see their work immediately and collaborate with their customers or coworkers on the spot.
A single Zeus PCIe card enables path tracing in real time at 4K 120 FPS for workloads like games, architecture, and product design without upscaling or frame generation workarounds. An architect working with their client can show off their photorealistic design and make changes with their clients immediately - reducing the need for hours-long re-renders or crashes from memory constraints. Film customers working on massive worlds with the highest quality textures can achieve real time path tracing with only 28 Zeus GPUs instead of 280 of the highest performing legacy GPUs. Not only will Zeus users be able to build smaller data centers or render farms, but they will require drastically less power from the city.
Glowstick will be included with Zeus at no additional cost and supports industry standard OpenUSD, MaterialX, OSL, and Deadline for seamless integration across platforms and render scheduling. Glowstick will also have its own texture library starting with 5,000 textures, the largest of any GPU vendor library.
Apollo: Electromagnetic Wave Simulation
Various industries, including scientific research, part and product engineering, pharmaceuticals, defense, energy, aerospace, weather modelling, photonics and optics design depend on accurate physics simulations.
Physics simulations were historically performed on large CPU clusters, but various limitations with legacy GPUs hindered wider adoption outside of exotic, expensive datacenter-class GPGPUs.
Electromagnetic wave simulations are key to designing modern high-technology products, including radar sensors, silicon photonics chips, medical equipment, consumer electronics products, and lenses. Faster and bigger simulations enable faster time to market and better parts, products, and systems.
Zeus resets performance and scale expectations around highly accurate physics simulations with over 300x performance in electromagnetic wave simulation, key for designing consumer electronics, optical lenses and waveguides, CT and X-Ray scanners, stealth materials, and more.
Zeus does not compromise accuracy for performance, maintaining IEEE-754 FP64 accuracy.
Users can take advantage of Zeus's enormous 2.25 TB of memory capacity to run 40x larger simulations.
Early Access: Join the Party
Bolt Graphics will be doing live demos at the Game Developers Conference (GDC) in San Francisco from March 18 to March 21. Users can sign up for early access to Zeus, Glowstick, and Apollo on the Bolt Graphics website (www.bolt.graphics). Developer kits will be available later in 2025, with mass production beginning in late 2026.
Source:
Bolt Graphics
Zeus is orders of magnitude faster than any other GPU in key workloads. Users can gain 10x in rendering performance, 6x in FP64 HPC workload performance, and 300x in electromagnetic wave simulations. Users running these types of demanding workloads need access to large amounts of memory. Bolt brings expandable memory to GPUs, for the first time, which allows users to increase their memory up to 384 GB in a PCIe card, and up to 2.25 TB per Zeus in a 2U server. A rack of Zeus 2U servers can be configured with up to 180 TB of memory, 8x larger than legacy GPUs.Historically, increasing performance requires more energy usage. Zeus reverses this long-term trend by reducing energy consumption while increasing performance. Reducing the impact GPUs have on the environment is one of Bolt's core values as a new entrant in the GPU space.
"Zeus increases performance while simultaneously reducing power consumption," said Darwesh Singh, Founder and CEO of Bolt Graphics. "I'm proud of the Bolt team's dedicated effort to create a solution that addresses key customer pain points, enabling them to be more productive and bring their ideas to life."Another first for GPUs is native integration of high-speed 400 GbE and 800 GbE Ethernet interfaces directly into the GPU eliminating the need for expensive, high latency, and power hungry network interface cards. Now users can directly connect Zeus GPUs to each other at massive scale without any of those downsides.
Zeus will be available in multiple form factors including PCIe cards, servers, and the cloud. Bolt plans on expanding Zeus into smartphones, tablets, laptops, consoles, and cars over the coming years, providing a unified GPU architecture across various platforms.
Glowstick: Real Time Path Tracing
Along with Zeus, Bolt Graphics is announcing Glowstick, a real time path tracer for rendering customers. Path tracing simulates complex light interactions that happen in the real world, but are too computationally intensive for legacy GPUs to simulate in real time. Glowstick will revolutionize industries like film, architecture, product design, and game development by enabling real time path tracing so users can see their work immediately and collaborate with their customers or coworkers on the spot.
A single Zeus PCIe card enables path tracing in real time at 4K 120 FPS for workloads like games, architecture, and product design without upscaling or frame generation workarounds. An architect working with their client can show off their photorealistic design and make changes with their clients immediately - reducing the need for hours-long re-renders or crashes from memory constraints. Film customers working on massive worlds with the highest quality textures can achieve real time path tracing with only 28 Zeus GPUs instead of 280 of the highest performing legacy GPUs. Not only will Zeus users be able to build smaller data centers or render farms, but they will require drastically less power from the city.
Glowstick will be included with Zeus at no additional cost and supports industry standard OpenUSD, MaterialX, OSL, and Deadline for seamless integration across platforms and render scheduling. Glowstick will also have its own texture library starting with 5,000 textures, the largest of any GPU vendor library.
Apollo: Electromagnetic Wave Simulation
Various industries, including scientific research, part and product engineering, pharmaceuticals, defense, energy, aerospace, weather modelling, photonics and optics design depend on accurate physics simulations.
Physics simulations were historically performed on large CPU clusters, but various limitations with legacy GPUs hindered wider adoption outside of exotic, expensive datacenter-class GPGPUs.
Electromagnetic wave simulations are key to designing modern high-technology products, including radar sensors, silicon photonics chips, medical equipment, consumer electronics products, and lenses. Faster and bigger simulations enable faster time to market and better parts, products, and systems.
Zeus resets performance and scale expectations around highly accurate physics simulations with over 300x performance in electromagnetic wave simulation, key for designing consumer electronics, optical lenses and waveguides, CT and X-Ray scanners, stealth materials, and more.
Zeus does not compromise accuracy for performance, maintaining IEEE-754 FP64 accuracy.
Users can take advantage of Zeus's enormous 2.25 TB of memory capacity to run 40x larger simulations.
Early Access: Join the Party
Bolt Graphics will be doing live demos at the Game Developers Conference (GDC) in San Francisco from March 18 to March 21. Users can sign up for early access to Zeus, Glowstick, and Apollo on the Bolt Graphics website (www.bolt.graphics). Developer kits will be available later in 2025, with mass production beginning in late 2026.
14 Comments on Bolt Graphics Announces Zeus GPU for High Performance Workloads
They'll be in Hot Chips in August.
Hope to see something real by then.
I can also seem them not wanting to be bought up for a quick cash grab, when they could potentially make more money by themselves if they actually managed to pull off what they are talking about. Or eventually raise the value of the acquisition is the world see them as a legit threat to Nvidia/AMD for CG. Their plan is probably to make money with the professionals first, and then target the conssumer market. What's interesting is how "niche" their GPU sounds like compared to what we have now where they are slowly turning into A.I accelerator tuned for gaming. It's sounds like they went for a design that is very good in specific task, but would rather let the A.I market to the other
That said, it's such a weird card design, even if we remove the top PCIe interface (which is most likely for debugging), there are so many things that don't add up, even with the high-speed fibre interface. Even things like the power regulation on the board is unique, at least the part on the rear of the PCB, which is unlike anything I've seen on an add-in card. Then there's the two SO-DIMM slots, which may very well be for some kind of local buffer, but would it really make sense compared to sharing system memory?
Sure, it's a render, so maybe the final card won't look anything like this and it could just be a ruse to confuse the competition, but it's all very strange.
www.servethehome.com/bolt-graphics-zeus-the-new-gpu-architecture-with-up-to-2-25tb-of-memory-and-800gbe/2/
Each board can have 1, 2 or 4 chiplets, with each chiplet having access to 32 or 64GB LPDDR5X, plus DDR5 SO-DIMMs, as seen in the above pictures for the 1c config.
Their LPDDR5 is rated at 273GB/s, so I will infer that it has a 256-bit bus at 8533MT/s (similar to Strix Halo, but with faster clocks). The DDR5 part is rated at 90GB/s, so I will infer they're using 128-bit (your DIY equivalent of "dual channel") at 5600MT/s.
Their total memory seems to be under the assumption of having 64GB LPDDR5X + 2x48GB SODIMM modules, with the bandwidth being the aggregated value from both of these.
All in all, seems like a DPU with a CPU that has a really beefed up SIMD unit. The memory bandwidth is really subpar compared to any other accelerator, and even their FP64 numbers are far from impressive compared to other GPUs that have it enabled.
It may be interesting for really specific scenarios, such as large path-traced renderings, but I don't think it'll be replacing any GPU farm. I have no idea why those are being planned as dedicated "accelerators". They are meant to be clustered among themselves:
If you are going to have multiple of those communicating among themselves, and totally ignoring the host, why even have a host? IMO they could be their own thing as in a box full of standalone devices with just a single host controller sitting in a different box.
Heck, you already have networking, display output, memory and everything else in there. Why not just slap an extra RISC-V CPU at this point and have it be totally standalone?
- at least 32 GB VRAM
- Plug & Play
Who knew it is that easy, it only needs to make it to the market now.
;)