Integration Guide

CoreWeave + Meridian

Deploy Meridian's GPU-accelerated inference pipeline on CoreWeave's Kubernetes-native infrastructure. Bare-metal NVIDIA H100s, InfiniBand interconnects, and zero cold-start scaling — all orchestrated through a single Helm chart.

Quickstart

Provision a CoreWeave Kubernetes cluster with GPU node pools, then deploy Meridian in under five minutes.

Step 1 — Cluster provisioning

coreweave kubernetes cluster create meridian-prod \
  --node-pool gpu-h100-x8:2 \
  --region US-EAST-02 \
  --k8s-version 1.29

Step 2 — Add Meridian Helm repository

helm repo add meridian https://charts.getnimbus.net
helm repo update

Step 3 — Deploy

helm install meridian meridian/meridian \
  --namespace meridian --create-namespace \
  --set gpu.count=8 \
  --set gpu.model=H100 \
  --set licenseKey=$MERIDIAN_LICENSE_KEY

Architecture

Meridian runs as a StatefulSet on CoreWeave, with each pod pinned to a dedicated H100 via GPU device plugin scheduling. InfiniBand SR-IOV provides pod-to-pod throughput exceeding 400 GB/s for tensor-parallel inference across nodes.

H100 Bare Metal

No virtualization layer. Direct GPU access with NVIDIA device plugin and MIG partitioning for multi-tenant workloads.

InfiniBand NDR400

400 Gb/s inter-node fabric. Meridian uses NCCL over IB for gradient reduction and tensor parallelism.

PVC-backed Model Cache

CoreWeave NVMe-backed PersistentVolumeClaims cache model weights. Sub-second cold starts on pod reschedule.

Horizontal Pod Autoscaling

Custom metrics from Meridian's Prometheus endpoint drive HPA. Scale from 1 to 64 GPUs based on queue depth.

VPC Peering

CoreWeave VPC peers directly with your AWS/GCP/Azure VPC. Inference endpoints stay on private RFC 1918 addresses.

Spot Instance Fallback

Configure spot priority with on-demand fallback. Meridian checkpoints inference state every 30s for graceful preemption.

Ready to deploy?

CoreWeave clusters are available on-demand. Contact sales for reserved-instance pricing and dedicated InfiniBand fabrics.