@cedanacorp

GPU migration infrastructure for inference and agentic workflows. Increase successful tasks per GPU-hour.

cedana.com
Joined August 2023
  • Tweets11
  • Following3
  • Followers32
New with the NVIDIA Dynamo team: a portfolio of frontier coding models on one 8x B200 node. Dynamo serves. NeMo Switchyard routes and escalates. Cedana switches models in 57-70s, up to 29x faster than a cold start, in-flight sessions intact. cedana.com/blog/coding-model…

Serve a Portfolio of Coding Models on a Single Node with NVIDIA Dynamo, NeMo Switchyard and Cedana...

Host a portfolio of frontier coding models on one 8x B200 node. NVIDIA Dynamo serves, NeMo Switchyard routes and escalates, and Cedana switches models in about a minute with sessions intact.

cedana.com
1
1
63
Native cold start vs. Cedana restore on 8x B200: DeepSeek-V4-Pro: 2051s → 70s (29.3x) GLM-5.2: 1322s → 61s (21.7x) Kimi-K2.6: 1217s → 63s (19.3x) MiniMax-M2.7: 564s → 57s (9.9x) Methodology: cedana.com/benchmarks/llm-co…

LLM Cold Start Benchmarks on 8x B200 | Cedana

Compare native and Cedana restore times for DeepSeek, Kimi, GLM, and MiniMax on 8x NVIDIA B200, with 57–70 second restores and up to 29.3x speedup.

cedana.com
1
60
Our next post will show benchmarks on speedup and cost improvements for tasks.
1
2
Your dashboard is green, but your fleet still delivers fewer tokens than the hardware can. About 500 of 6,200 GPUs sat unusable on one cluster at 90% packing (Weng, ATC '23). Immobility never trips an alarm. #AIInfrastructure
1
49
Full analysis, with a source for every number: cedana.com/post/the-era-of-s…

The Era of Stateful Inference: how to improve cost per token. | Cedana

We are entering the stateful inference era, driven by frontier models with longer context windows, longer in-flight sessions, and single instances spanning 8, 16, or more GPUs.

cedana.com
33
A frontier cold start runs 10 minutes to an hour. Users tolerate seconds. So you hold warm pools. Microsoft's published figure for scheduling around that at O365 scale: up to $2.5M a month (SageServe, SIGMETRICS 2025). #AIInfrastructure
1
2
58
Full analysis, with a source for every number: cedana.com/post/the-era-of-s…

The Era of Stateful Inference: how to improve cost per token. | Cedana

We are entering the stateful inference era, driven by frontier models with longer context windows, longer in-flight sessions, and single instances spanning 8, 16, or more GPUs.

cedana.com
20
For the impatient, snapshot/migrate/resume for vanilla GPU workloads works right now thanks to the good folks @cedanacorp.
Erik Bernhardsson
@bernhardsson9 Jul 2025
The holy trinity of serverless GPU has been achieved internally at @modal_labs: ✅ Custom filesystem optimized for container cold start ✅ CPU snapshot+restore ✅ GPU snapshot+restore Last one is working internally but not released yet – stay tuned!
1
2
810
Caltech's Computational Biology group used Cedana to automatically orchestrate AI and HPC workloads to deliver results ~2x faster and slash costs by ~80%. This is the power of automated GPU workload migration. xitter.cf/garrytan/status/192457…
Garry Tan
@garrytan19 May 2025
YC startup Cedana enables play/stop/rewind on all GPU and CPU compute - the result is 10x faster time to first token. Training jobs automatically resume from failures. Caltech's computational biology team just saved 80% on compute by deploying Cedana 🤯
2
185
Cedana supercharges AI and HPC workloads with live migration of your GPU and CPU workloads. We automatically migrate jobs to optimize SLAs including performance, reliability, cost, and latency. We seamless integrate with K8s, SLURM, Kubeflow, Kueue and Ray.
1
146
Cedana automatically migrates your GPU/CPU workloads when spot instances get revoked or catastrophic failure. ✅ 2-10x faster inference cold starts ✅ Save up to 80% on compute costs ✅ Fault-tolerant, higher throughput agents ✅ Zero code changes cedana.slack.com/archives/C0…
1
94
Load more