{"slug": "show-hn-we-built-fractional-gpu-slicing-without-nvidia-mig-works-on-amd-too", "title": "Show HN: We built fractional GPU slicing without Nvidia MiG – works on AMD too", "summary": "Podstack launched fractional GPU slicing without Nvidia MiG, supporting AMD GPUs, enabling per-minute billing and dynamic slice resizing from 12.5% to 100% of a GPU. The platform aims to reduce GPU waste and simplify ML infrastructure by offering a unified cloud for training, inference, and billing, with self-hostable options for datacenters.", "body_md": "One full-stack GPU cloud for the entire model lifecycle. Fractional GPUs by the second, one-click training and inference, and a single bill - on infrastructure you can run yourself.\n\nNo credit card required. Spin up your first GPU in minutes.\n\nWorks with your stack\n\nGetting a model to production means fighting scarce, overpriced GPUs and a stack stitched together from half a dozen vendors - burning budget and weeks before you ship. Here's what's broken today, and how Podstack fixes it.\n\nYou rent whole cards on long commitments and pay for capacity you never use.\n\nPodstack: Fractional GPUs, per-minute billing, on-demand - pay only for the slice you use.\n\nCompute, MLOps, inference and billing come from different vendors - many bills, many SLAs, endless glue code.\n\nPodstack: One platform to launch, train, serve and operate. One operator, one SLA, one bill.\n\nIt takes weeks of infrastructure plumbing to get from an idea to a served model.\n\nPodstack: One-click templates and a first-class CLI - from zero to a running model in minutes.\n\nHyperscalers trap your data and workloads behind proprietary APIs and egress charges.\n\nPodstack: Portable and self-hostable, with zero egress. Run it on our cloud - or your own.\n\nQuickPods, TrainPods, and Inference all run on DC Suite - the same platform datacenters license to run their own GPU cloud.\n\nThe ML Infra Stack\n\nEvery AI product runs on the same stack of layers. Podstack builds the five in the middle - developer tools down to GPU virtualization - and licenses them together as DC Suite, so any datacenter can turn GPU hardware into a cloud like this one.\n\nServerless notebooks, Python SDK, CLI, web dashboard, and one-click templates - the developer entry point.\n\nShips inside DC Suite - run it on our cloud, or license it for yours.\n\nProprietary control plane and scheduler, scale-to-zero, container/OCI compatible.\n\nSlice any GPU 12.5-100% - NVIDIA and AMD, no vendor SDK lock-in.\n\nPer-minute metering, billing, and invoicing for every tenant.\n\nCost tracking and utilisation insight across the fleet.\n\nOut-of-the-box hooks into your datacenter infrastructure management.\n\nPer-tenant isolation and audit logging, ready for regulated customers.\n\nYour buyers sign up, launch, and pay without a ticket queue.\n\nUtilisation, health, and capacity monitoring across every card.\n\nGPU capacity sits idle. Whole cards get sold to workloads that only need a fraction of the VRAM.\n\nDC Suite: PodVirt slices any GPU from 12.5% to 100%, so every gigabyte of VRAM is sellable and utilisation becomes revenue.\n\nHardware partitioning (MIG) is rigid: fixed slice sizes, top-end cards only, firmware changes and node restarts to reconfigure.\n\nDC Suite: Software-defined slicing works across NVIDIA and AMD - any slice size, resized dynamically, no vendor SDK lock-in.\n\nOwning GPUs does not make you a cloud. You still need scheduling, metering, billing, and a customer portal.\n\nDC Suite: Orchestration, BillOps metering, billing and invoicing, and a self-serve portal ship in the box - capacity earns from day one.\n\nEnterprise customers demand isolation, audit trails, and data residency you can prove.\n\nDC Suite: Per-tenant isolation, audit logging, and residency controls are built into the platform - the same controls behind our own cloud.\n\nMost datacenters worldwide still have no platform to sell and maintain their GPU inventory - raw GPU rental leaves value on the table.\n\nDC Suite: Out-of-the-box DCIM integration turns inventory into a managed, value-added catalogue, so DCs upsell platform services on top of raw GPU - not just rack space.\n\nExact free capacity is opaque. A buyer who needs GPUs today has no way to see which datacenter has precisely that capacity available.\n\nDC Suite: The Podstack network connects buyers in need to supplier DCs with exact, live capacity - demand lands on the right GPU at the right datacenter.\n\npartner DCs expand their fleets once idle capacity starts earning\n\n85-87% across datacenters running Podstack\n\nSOC 1, SOC 2, GDPR, ISO and the other certifications a DC needs\n\nPodstack runs its own GPU fleet and onboards partner capacity only after a stringent vetting process. Every datacenter in the network carries SOC 1, SOC 2, GDPR, ISO and the other certifications enterprise workloads demand.\n\nDescribe your workload and get GPUs ranked from live pricing and availability.\n\nCustom & Enterprise Solutions\n\nReserved capacity, high-performance storage, managed labs, and production inference - planned with our team and sized to your workload.\n\nSDK, CLI, API, or YAML. Your choice.\n\n``` python\nimport podstack\nfrom sklearn.datasets import load_iris\nfrom sklearn.tree import DecisionTreeClassifier\nfrom sklearn.model_selection import train_test_split\n\n# Initialize Podstack\npodstack.init(api_key=\"your-api-key\")\n\n# T\n```\n\nOur own control plane and scheduler - not open-source patchwork.\n\nAudited security controls for regulated workloads.\n\nChoose where your workloads run.\n\nOne operator, one SLA, one bill.\n\nGet started in minutes. No credit card required.\n\nWe use cookies to enhance your browsing experience, analyze site traffic, and personalize content. By clicking \"Accept All\", you consent to our use of cookies.\n\nRead our [Privacy Policy](/privacy) for more information.", "url": "https://wpnews.pro/news/show-hn-we-built-fractional-gpu-slicing-without-nvidia-mig-works-on-amd-too", "canonical_source": "https://podstack.ai/", "published_at": "2026-07-18 11:35:01+00:00", "updated_at": "2026-07-18 11:51:18.872011+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-infrastructure", "ai-products", "ai-tools", "developer-tools"], "entities": ["Podstack", "Nvidia", "AMD", "DC Suite", "PodVirt"], "alternates": {"html": "https://wpnews.pro/news/show-hn-we-built-fractional-gpu-slicing-without-nvidia-mig-works-on-amd-too", "markdown": "https://wpnews.pro/news/show-hn-we-built-fractional-gpu-slicing-without-nvidia-mig-works-on-amd-too.md", "text": "https://wpnews.pro/news/show-hn-we-built-fractional-gpu-slicing-without-nvidia-mig-works-on-amd-too.txt", "jsonld": "https://wpnews.pro/news/show-hn-we-built-fractional-gpu-slicing-without-nvidia-mig-works-on-amd-too.jsonld"}}