dstack gives teams a unified way to run and manage GPU-native containers across clouds and on-prem environments — without requiring Kubernetes.
At the same time, many organizations rely on Kubernetes as the foundation of their infrastructure.
To support these users, dstack is releasing the beta of its native Kubernetes integration.
This benchmark investigates whether the Prefill–Decode worker ratio needs to be managed dynamically at runtime, or if a fixed split can deliver the same performance with simpler orchestration.
We evaluate different ratios across workload profiles and concurrency levels to measure their impact on TTFT, ITL, and throughput, and to see whether fixing the ratio in advance is a practical alternative to dynamic adjustment.
dstack is an open-source control plane for orchestrating GPU workloads. It can provision cloud VMs, run on top of Kubernetes, or manage on-prem clusters. If you don’t want to self-host, you can use dstack Sky , the managed version of dstack that also provides access to cloud GPUs via its markfetplace.
With our latest release, we’re excited to announce that Nebius , a purpose-built AI cloud for large scale training and inference, has joined the dstack Sky marketplace
to offer on-demand and spot GPUs, including clusters.
This is a practical map for teams renting GPUs — whether you’re a single project team fine-tuning models or a production-scale team managing thousand-GPU workloads. We’ll break down where providers fit, what actually drives performance, how pricing really works, and how to design a control plane that makes multi-cloud not just possible, but a competitive advantage.
Orchestration automates provisioning, running jobs, and tearing them down. While Kubernetes and Slurm are powerful in their domains, they lack the lightweight, GPU-native focus modern teams need to move faster.
dstack is built entirely around GPUs. Our latest update introduces native integration with DigitalOcean and
AMD Developer Cloud , enabling teams to provision cloud GPUs and run workloads more cost-efficiently.
dstack services are long-running workloads—most often inference endpoints and sometimes web apps—that run continuously on GPU or CPU instances. They can scale across replicas and support rolling deployments.
This release adds HTTP probes inspired by Kubernetes readiness probes. Probes periodically call an endpoint on each replica (for example, /health) to confirm it responds as expected. The result gives clear visibility into startup progress and, during rolling deployments, ensures traffic only shifts to a replacement replica after all configured probes have proven ready.
In large-scale training, a single bad GPU can derail progress. Sometimes the failure is obvious — jobs crash outright. Other times it’s subtle: correctable memory errors, intermittent instability, or thermal throttling that quietly drags down throughput. In big experiments, these issues can go unnoticed for hours or days, wasting compute and delaying results.
dstack already supports GPU telemetry monitoring through NVIDIA DCGM metrics, covering utilization, memory, and temperature. This release extends that capability with passive hardware health checks powered by DCGM background health checks. With these, dstack continuously evaluates fleet GPUs for hardware reliability and displays their status before scheduling workloads.
As the ecosystem around AMD GPUs matures, developers are looking for easier ways to experiment with ROCm, benchmark new architectures, and run cost-effective workloads—without manual infrastructure setup.
dstack is an open-source orchestrator designed for AI workloads, providing a lightweight, container-native alternative to Kubernetes and Slurm.
Today, we’re excited to announce native integration with Hot Aisle , an AMD-only GPU neocloud offering VMs and clusters at highly competitive on-demand pricing.
This is the first in our series of benchmarks exploring the performance of AMD GPUs in virtualized versus bare-metal environments. As cloud infrastructure increasingly relies on virtualization, a key question arises: can VMs match bare-metal performance for GPU-intensive tasks? For this initial investigation, we focus specifically on a single-GPU setup, comparing a containerized workload on a VM against a bare-metal server, both equipped with the powerful AMD MI300X GPU.
Our new benchmark explores two important areas for optimizing AI workloads on AMD GPUs: First, do containers introduce a performance penalty for network-intensive tasks compared to a bare-metal setup? Second, how does partitioning a powerful GPU like the MI300X affect its real-world performance for different types of AI workloads?
