# GPU Optimization

GPUs are scarce. Sedai finds the capacity locked in idle allocations and underused GPUs and safely puts it to work, so you get more from what you already have.

## Do More With the GPUs You Have

GPUs are scarce and buying more takes months. Sedai finds the capacity locked inside your fleet, from idle allocations to workloads holding full GPUs they only need a slice of, and safely puts it to work. More workloads per GPU. No disruption in prod.

## Before You Buy Another GPU, Get More From the Ones You Have

Teams hold idle GPUs on purpose, because giving them back can mean waiting months to get them again. Sedai doesn’t ask you to let go. It finds the capacity standard metrics miss, then right-sizes and packs workloads so the GPUs you’re holding take the next job in the queue.

## GPU Workload, Node & Cluster Optimization

Static GPU allocations lead to massive waste. Sedai's proprietary utilization model continuously adapts to real workload behavior, keeping GPU usage optimized even as your AI infrastructure evolves.

### Fit More Workloads per GPU

Most inference and dev workloads don’t need a whole GPU. Sedai shares GPUs through MIG or time-slicing, picks the right configuration, and packs more workloads onto each device.

### Reclaim Idle GPUs

Find GPUs that are allocated but doing nothing. Sedai releases them for other workloads, with the impact shown before and after every change.

### Consolidate GPU Node Pools

Sedai packs scattered workloads onto fewer nodes, raising utilization and cutting node spend. Every scale-down is yours to approve.

## 100% Utilization Doesn’t Mean 100% Used

Most tools rely on standard utilization metrics, such as those reported by NVIDIA System Management Interface (nvidia-smi). However, those metrics only tell you whether a GPU is doing *something*, not whether it's doing something *useful*. A GPU can show 100% utilization while performing zero productive computation.

Sedai approaches this differently:

- Proprietary utilization model infers true GPU usage from multiple telemetry signals

- Models real workload behavior across compute, memory, and throughput dimensions

- Provides a first-class utilization score that drives every optimization decision

- Identifies waste that surface-level metrics consistently miss

## GPU Capacity & Cost Intelligence

Most tools only show you where GPU spend goes. Sedai knows why it's happening and reduces it for you.

### Free Capacity You Already Own

Before procuring new GPUs, reclaim the ones you have. Sedai continuously identifies underutilized devices and frees them for use, reducing procurement delays and queue times for AI teams.

### See Where Your GPUs Actually Go

See exactly where GPU spend lives across workloads, node pools, and clusters. Sedai turns cost drivers into actions for measurable, ongoing savings.

### Waste Detection at Every Layer

Find inefficiencies across workloads, nodes, and clusters. Sedai surfaces idle and over-allocated GPU capacity and removes it, safely and autonomously.

“By having Sedai in place, we’re not just saving money. We’re preventing would-be customer problems, before they become an issue.”

— Matt Duren, VP of Engineering // KnowBe4

## How Sedai Optimizes GPU Infrastructure Safely

Get safe, outcome-driven GPU optimization at scale, designed to act on real workload behavior, with safeguards built into every decision.

### Learns GPU Workload Behavior

### Decides With Context

### Acts Safely

## Autonomy That Delivers

Powered by real app behavior.

GPU capacity reclaimed

Increase in effective utilization

Less manual optimization work

## Optimize Your Entire GPU Stack

Sedai makes your GPU infrastructure smarter and safer.

### Kubernetes Platforms

### Optimize GPU workloads across any Kubernetes distribution

### GPU Right-Sizing

### Slice and pack GPUs to eliminate waste

### Monitoring Integrations

### Plug into your existing observability stack

## Other Tools Automate. Sedai Acts With Real Context.

| Other Solutions | Sedai |
| --- | --- |
| Rely on surface-level GPU utilization metrics | Models true GPU utilization from multiple telemetry signals |
| Stop at dashboards and recommendations | Autonomously executes changes with built-in guardrails |
| Lack workload-level GPU intelligence | Optimizes at the workload, node, and cluster level |
| CPU-focused; GPU optimization is an afterthought | Purpose-built for GPU and AI infrastructure |
| No path from recommendations to safe autonomous action | Datapilot → Copilot → Autopilot progression for safe autonomy |
| Treat GPU efficiency as a cost line item | Turns underused GPUs into capacity for more workloads |

## Get More GPU Capacity
Without the Procurement Wait

## FAQs

### How does Sedai determine true GPU utilization?

Sedai uses a proprietary model that infers real GPU usage from multiple telemetry signals, going beyond the standard nvidia-smi utilization metric, which measures only whether a GPU is active, not whether it's doing productive work.

### Is it safe to let Sedai make autonomous GPU changes?

Yes. All changes are governed by the same safety framework Sedai applies across the platform — incremental execution, continuous validation, and automatic rollback on drift. You can start with Datapilot recommendations and progress toward autonomy at your own pace.

### Which GPU capabilities support autonomous execution today?

GPU Workload Deallocation and MIG Enablement & Packing support both Datapilot and Copilot modes at launch. GPU Node Pool Optimization is available in Datapilot mode, with Copilot support on the roadmap.

### Which Kubernetes platforms are supported?

Sedai GPU Optimization supports any Kubernetes platform and distribution, including EKS, GKE, AKS, OpenShift, Rancher, VMware Tanzu, and more.

### Will GPU optimizations affect my AI workload performance?

Sedai only acts where it has high confidence that changes are safe. Every optimization is validated against workload behavior and performance signals before and after execution.

### Can Sedai help us delay or avoid buying more GPUs?

Yes. Most GPU fleets have significant capacity locked in idle allocations and workloads holding full GPUs they only need a slice of. Sedai finds that capacity and frees it for new workloads. That shortens queue times now and lets you put off new GPU purchases until you’ve gotten full value from the fleet you already own.

### How much GPU capacity can Sedai reclaim?

It depends on your workloads and how your GPUs are allocated today. Sedai has driven up to 3x effective GPU utilization, meaning up to three times the productive work from the same hardware.

### If Sedai frees a GPU, do we lose it?

No. In Kubernetes, freeing a GPU usually means right-sizing requests, bin-packing workloads, or sharing GPUs through MIG or time-slicing. The node stays in your pool, and the scheduler places your next pending workload on it. If your GPUs come through reserved capacity, cloud commitments, neocloud contracts, or your own clusters, that capacity stays yours either way. Node scale-downs are separate recommendations, and you decide whether to approve them.

### How quickly can I see results?

Most customers begin seeing GPU cost savings and utilization improvements within the first few weeks of enabling the capability.

### How does GPU Optimization fit into the broader Sedai platform?

GPU Optimization is built on the same observe → learn → act engine that powers Sedai's CPU, memory, and cluster optimization. Customers already using Sedai can enable GPU Optimization within their existing environment.