One hour. 1,081 nodes successfully connected. 1,008 nodes online at the same time. GPUWorks' first public "showcase" after launch — the 1,000-Node Challenge — delivered this live scorecard. After the livestream ended, we sat down with Wang Wei, CTO of ZStack, to talk more. Why take on a thousand-node challenge? How was a 45 MB installation package built? What extra work do engineers put in behind the scenes so users can save a little time? The following conversation is based on the interview, with questions lightly condensed and content abridged and reordered for clarity.
01 / SCALE IN ACTION
"The installer is small, but we wanted to prove it has real muscle"
Host: Why launch this challenge? Why 1,000 nodes?
Wang Wei: Long-time ZStack users may know that when we first released our product, we emphasized scalability — the ability to support 1,000 servers. But back then, doing cloud platforms and virtualization management meant that to demo this scenario you'd actually need 1,000 physical servers, which came with real constraints. From day one, GPUWorks was designed with the idea that users might have GPU servers, GPU virtual machines, or GPU containers — different forms of compute could all be brought in. So we thought: if we claim to support 1,000 GPU nodes, why not demonstrate it in the cloud?
There's another reason. Many people might assume this product is only 45 MB and installs in seconds, which doesn't fit the traditional image of enterprise software. In the past, installing a piece of software might mean reading a long manual and spending dozens of minutes, even days. So we wanted to prove that such a small footprint can still carry serious capability. That's how this livestream came about.
Host: What does thousand-node scale mean to users just starting out with a few machines?
Wang Wei: When users first get their machines, they might just be testing the waters with three or five units, and can quickly install GPUWorks. When it's time for real production and three or five isn't enough, scaling to 10, 100, or even 1,000 nodes doesn't require migration or redeployment — you simply keep adding nodes to the existing cluster.
02 / LIGHTWEIGHT DEPLOYMENT
"GPU server time is precious"
Host: The launch specifically highlighted 45 MB and 60 seconds. Why those two numbers?
Wang Wei: GPUWorks actually has a lot of features, and the community has had plenty of discussion lately. But for the first release, we picked a few points we most wanted people to remember. One is that the installer is small and self-contained. Users don't need to install Docker or a database just to run it. Once the machine's OS is ready, you drop in the binary and you're done.
The other is that installation is simple. The binary you download is both the installer and the main program — a bit like the "portable software" people used to love: extract it and it just works. What I dread most is an install that drags on, throws an error midway, walks me through step after step, and in the end I'm not even sure where it landed. We wanted to make this process as effortless as possible.
Host: Why invest so much effort into the installation step?
Wang Wei: Because GPU server time is precious. By the time users order, wait for delivery, rack, power on, and connect the network, they've already spent a lot of time. If installing software adds days or even a week on top of that, it's a huge waste. Renting is even worse — machines are billed by the hour, and if all that time goes into installing and deploying software, it's genuinely agonizing. So we wanted it to install within 60 seconds.
03 / ENGINEERING CHOICES
"These are really some obsessions we technologists carry"
Host: What went into shrinking the installer to 45 MB?
Wang Wei: When I joined ZStack in 2016 doing backend development, we reported to Zhang Xin. He often said: for a computer, one second is a long time. If an operation has to wait a second, you have to ask why — can it be done in 0.1 seconds, or less? The same goes for space. Today we download a song or a movie, and no one bats an eye at tens of megabytes or a few gigabytes. But think of old game cartridges — tiny capacity, yet they could hold rich content. We ask the same: can this space be used more precisely?
Host: With disk capacity so large now, why still care about installer size?
Wang Wei: These are really some obsessions we technologists carry. For many users, the difference between a few, tens, or hundreds of megabytes may not matter much. But our thinking is that a sufficiently small installer is in some ways proof that the engineers and architects did careful design — how the package should be built, what it should contain, all of it was thought through.
Take the bandwidth test in network management. We found that some users have RDMA networks and want to know how the bandwidth is doing and whether the network is healthy, but they're not familiar with what software or dependencies to install. So we thought, can we put the test code inside the product? That involves trade-offs. Leave dependencies outside, and users hit a wall mid-use when they're suddenly told to install another package; bring everything in, and the product keeps growing. So it takes a lot of weighing and trimming.
04 / FIT YOUR ENVIRONMENT
"We want users to download it and use it right away"
Host: How does not needing to set up K8s first help users get started?
Wang Wei: First, technically speaking, no single approach is absolutely superior. We don't reject K8s — in fact we've borrowed a lot of good designs from it, like tag-based batch management. But if the platform itself must depend on K8s, users have to prepare management nodes and related dependencies first, and may also need to learn some K8s operations. We want the product to be as simple as possible, with a low barrier to entry — download it and use it.
Host: What if a user's GPUs are already in K8s or Slurm?
Wang Wei: That's on our roadmap as well. For GPU devices already in K8s or Slurm, we hope to provide ways to connect to and use those resources. We've talked with many universities and research institutions, and a lot of their GPU devices are deployed in Slurm clusters. We want to accommodate these real-world usage scenarios too.
Note: K8s and Slurm resource integration is described in this interview as a future roadmap item.
05 / FROM GPUWORKS TO FLITAI
"Light and nimble, but still powerful enough"
Host: From GPUWorks to FlitAI, what product philosophy do you want to carry forward?
Wang Wei: "Flit" means light and nimble. One of the things ZStack was first known for was simple installation and ease of use. We want to carry that into the AI era and AI scenarios, while also giving people something fresh. Beyond GPUWorks' compute and model management, there will be other products down the line. For example, Zentrix — mentioned in the livestream — includes AI gateway capabilities and may see extensions later, or products around observability. Whatever product we build, we want it to be easy to install and deploy, as easy as possible to use, and powerful at the same time.
The 1,000-Node Challenge let everyone see GPUWorks perform live. This conversation also showed how the team understands "user-friendly": understanding the devices users have, the environments they already run, and the steps that actually cost people time — then bringing those questions back into product design. Feel free to try GPUWorks for yourself, and share your experience with us.
06 / COMMUNITY
Join the community. GPUWorks' open-source journey has only just begun. Whether you're an AI developer, an infrastructure engineer, or a hardware vendor, you're welcome to join us — try different model and compute combinations, share your experience, report issues, and help us improve GPUWorks together.
Try GPUWorks and join open-source collaboration!
Welcome to the first community practice program! GPUWorks' first community practice program is now open!
GitHub: github.com/flitai-io/gpuworks
Docs: gpuworks.readthedocs.io
Discord:https://discord.gg/9m3gryGwzg