How Korea University’s AGI Lab Accelerated LLM Training with Elice Cloud

A GPU Cloud Case Study on Improving Research Speed and Meeting Paper Deadlines

“Even with 8+ A100 or H100 GPUs, single experiments still took over 48 hours.
As deadlines approached, we simply didn’t have time to wait.”

What happens when AI researchers face a backlog of LLM training jobs—while the clock is ticking?
At Korea University’s AGI Lab (PI: Prof. Sungwoong Kim), researchers build multimodal large language models that generate across text, images, and audio.
But limited GPU access and sequential training quickly became a bottleneck.

A Research Lab with Heavy GPU Demand

AGI Lab runs on-prem A100s and L40s and leases H100s from Korea’s AICA infrastructure.
Yet, as experiments grew more complex, resource wait times increased—and parallelism became non-optional.

“We couldn’t run multiple jobs at once. Even simple variation tests had to be queued.”

From Teaching Support to Core Research Infrastructure

Elice Cloud was originally adopted for undergraduate coursework.
But it quickly became core to the lab’s research pipeline.

“We needed an infra partner that was fast, scalable, and reliable—Elice fit.”
Researchers now run multimodal training on Elice’s G-NHHS-640 H100 instance and store data with Elice’s bucket service for easy sharing across jobs.
“We used to waste time re-uploading datasets every time.
With the bucket, everything’s preloaded and instantly accessible from any instance.”

Unlocking Experiment Parallelism

After switching to Elice Cloud, AGI Lab could parallelize high-performance jobs for the first time.

“We no longer had to wait or compromise. Now we can run multiple full-scale experiments at the same time.”

During a recent paper submission, the team tested multiple hyperparameter sets in parallel—right up to two days before the deadline.

“That ability to keep testing late into the cycle helped us submit a better paper.
Without Elice Cloud, we couldn’t have pulled it off.”

Optimized for Experiment-Centric Workflows

Elice Cloud’s interface allows researchers to launch new GPU instances in seconds.
Reusable templates make repeating experiments effortless.

“We once hit a Shared Memory error during training.
Elice Cloud resolved it immediately—no delays, no rescheduling.”

The lab values the fast, local support as much as the compute.

Building a Hybrid Infrastructure Strategy

Going forward, AGI Lab will split its compute across two platforms:

• Local servers for early-stage debugging
• Elice Cloud for deadline-bound or large-scale training

“Elice Cloud is where we go when we need things to just work.”

Designed for Researchers. Built for Results.

“It’s not just about having GPUs. It’s about having the right GPUs, at the right time, with zero setup hassle.”

Elice Cloud helps researchers reduce time-to-result, explore more ideas, and publish faster.

Run more experiments. Finish on time.

Need to train large models under a deadline?
Want infrastructure that scales with your ideas?

Start training smarter with Elice Cloud.

👉 Explore Elice Cloud

👉 Talk to Our Team