cedana / solutions · pharma & biotech

Your science shouldn't
restart from zero.

A maintenance window or a failed H100 sends a week-long run back to the start. Cedana checkpoints every GPU workload continuously, so an outage costs a minute.
effective utilization
42 80%+
Fortune 100 pharma R&D cluster
work lost
0
when a GPU fails or drains
time to first migration
<30min
to deploy on your existing stack
code changes
0
for any researcher
The workloads · stateful by nature

Drug discovery runs for days. The state lives on the GPU.

Each one keeps its state in GPU memory, not a database. When the GPU goes away, the work goes with it.

AlphaFold 3 · RoseTTAFold · OpenFold · Boltz

Structure prediction

The first step in nearly every modern discovery pipeline, with state held on the GPU across long multi-GPU runs.

at riskPrediction and training runs span hours to weeks. One failed node restarts the whole synchronized job from zero.
with cedana
  • Coordinated checkpoints across all GPUs (NCCL/RDMA)
  • A failed node resumes the run, never restarts it
  • Drain for maintenance without losing GPU-weeks
alphafold · foldingpredicting…
pLDDT 42 · 13 residues~50k GPU-hours
gpu · 8× H100SXM5 · NVLink
util█████████████░░░82%
vram█████████████░░░66/80GB
The problem · why the GPU goes away

Six ways a shared cluster loses the work.

01
DRAIN.WINDOW

Maintenance windows

A reservation holds every job near a quarterly patch. The cluster idles on both sides; jobs that slip through die.

02
FATAL.NODE

Hardware failures, compounding

H100s fail roughly 8-9% a year, and bigger runs fail faster. One failure loses the whole run.

03
WASTE.ALLOCATION

Over-allocation defense

Scientists oversize jobs to dodge an OOM crash. The dashboard looks busy; the cluster isn't.

04
FRAG.STRAND

Fragmentation & queue waits

Odd shapes don't fit the standard 8-GPU node. GPUs strand and researchers wait hours.

05
EVICT.2MIN

Spot & cloud-burst risk

Spot is up to 90% cheaper but evicts in two minutes. Long runs can't risk it.

06
KILL.PRIORITY

Preemption for priority

A high-priority job arrives and the scheduler kills a long run. The work is lost, not paused.

A Vrije Universiteit Amsterdam study found up to 65% of compute time lost to cancellations, timeouts, and failures. The bill is paid in scientist-hours and program timelines, not just GPU dollars.
Why cedana · what your cluster gains

More science from the same GPUs, nothing for researchers to change.

Six benefits, each one earned by a specific piece of the architecture, not a slogan.

continuous transparent checkpointing

No work lost to a failure or a patch

run · save · resumerunning…
startdone
work lost: 0resumes at last save, not zero
migrate to idle nodes · resize · automate maintenance

Twice the science from the same cluster

cluster · effective util42%
baseline 42%reclaimed +0
42% effective · was 42%same hardware
transparent at the OS level · no code change

Nothing changes for your researchers

workflow · unchangedtyping…
researcher@cluster : ~
0 code changessame scheduler
distributed multi-GPU coordinated (NCCL/RDMA)

Training is protected from failures

8× H100 · all-reduce (NCCL)syncing…
0102030405060708step 12,340 / 50,000
job never restarts~50k GPU-hours kept
workload portability · spot-safe migration

Access cheaper, burstable capacity

workload · portablerunning
$$$on-prem$spot$$cloud
up to 90% cheapersame checkpoint, either side
on-premise & air-gapped · provable recovery

Fits your regulatory posture

air-gapped perimeterrecording…
on-premcheckpoint ledger
provable recoveryappend-only · in your perimeter
One architectural move sits under all six: transparent OS-level capture, resumed on healthy hardware. The compound result on a Fortune 100 pharma R&D cluster: effective utilization from 42% to 80%+, the same GPUs, zero code changes.
How it works · save → migrate → resume

Your run survives the GPU underneath it.

Cedana runs at the OS and kernel level and works transparently with Slurm or Kubernetes, keeping every run moving across the GPUs underneath it.

GPU-A● runningmd · trajectoryT 310 K · NVTutil82%vram51/80Gstep 3,000 / 50,000GPU-B○ healthy · standbymd · trajectoryutil0%vram0/80Gstandby↳ live migration · the analysis never stops
01

Running

A run lives in GPU memory for days.

02

Save

Cedana snapshots full GPU state, below the app.

03

Migrate

A fault or drain triggers a live migration.

04

Resume

It resumes at the same step. Nothing lost.

vMotion for GPU workloads. Researchers never touch it; their runs just stop dying.

Runs on your stack · any cloud, any scheduler

Runs everywhere your GPUs already live.

AWSGoogle CloudNVIDIASLURMKubernetesLambdaCoreWeaveAWSGoogle CloudNVIDIASLURMKubernetesLambdaCoreWeave
FAQ · what infra teams ask

Questions we hear first.

01How does this work, and where does it sit in your stack?
Cedana is transparent and runs at the OS and kernel level, beneath your applications. That means it works transparently with Slurm and Kubernetes, so it runs with any of your workloads with no modifications. See the installation instructions →
02Will this disrupt our researchers?
No. Cedana is transparent. Same Slurm and Kubernetes commands, same notebooks. Jobs simply stop dying.
03Does it pass IT and security review?
Yes. Cedana is SOC 2 compliant, ships on-premise, and supports fully air-gapped deployments, so it stays inside your existing security perimeter.
04How do we know the savings are real at our scale?
We run a utilization-and-maintenance audit on your own workload mix to produce a comparable number first.
05What about our Schrödinger and HPC stack?
Cedana integrates beneath ISV job managers and works with Slurm, Kubernetes, Kueue, Ray, Armada, and Nvidia Dynamo.
06What about cloud bursting?
Cedana supports portability between on-premise and cloud. The same checkpoint resumes on either.
~ / cedana / deploy ready

Command your
compute.

Your compute, liquid — checkpoint, migrate, and resume live GPU jobs across the fleet.

deployK8s helm chart · SLURM plug-in
first migration<30 min
code changes0
workloadstraining · inference · HPC
01
AWS
02
Google Cloud
03
NVIDIA
04
K8s
05
SLURM
06
Nvidia Dynamo