ANZ Bank · NVIDIA
0.82 — Gini coefficient for assessing risk
Independently reportedBanking & finance · Other · Classical ML · Pilot
Other
SF Technology · HAMi
Vendor-reportedLogistics & transportChinaIn productionEnterpriseLLM
Vendor-reported. The customer is named and the numbers are quoted from the source page, but the account comes from the vendor. No independent confirmation.
Reported by dynamia.ai (vendor self-report). Checked against the source page on 2026-09-06. 4 of 4 figures below appear on that page word for word.
The problem
Traditional GPU usage patterns led to underutilization (below 30%), resource waste, coarse scheduling, and difficulties adapting to heterogeneous devices.
What was deployed
- Vendor
- HAMi
- Products
- EffectiveGPU, HAMi
- Technique
- LLM
- Build or buy
- Bought and customised
- Deployment
- Custom build
- Scale
- 65 services on 28 GPUs; 19 services on 6 test GPUs
What changed
Each row is quoted from the source. Figures we could not find on the page in those words are marked — they are kept, not deleted, so you can judge them.
37 cardsGPU cards saved for large model inference services
“Deployed 65 services using 28 GPU cards, saving 37 cards”
Quoted word for word from the source
13 cardsGPU cards saved for testing service cluster
“Deployed 19 services using 6 test GPU cards, saving 13 cards”
Quoted word for word from the source
0.5%Performance degradation after adding pooling layer
“Minimum performance decrease of only 0.5% after adding pooling layer”
Quoted word for word from the source
up to 200%Memory overcommitment ratio
“Introduced dual-dimension overcommitment technology for memory and computing power (up to 200% memory overcommitment ratio)”
Quoted word for word from the source
EnsuredReal-time performance for critical tasks
“Ensured real-time performance for critical tasks through priority scheduling and resource overcommitment”
Quoted word for word from the source
“Through close collaboration with the HAMi open-source community and secondary innovation based on its framework, EffectiveGPU has helped us significantly improve GPU resource efficiency and reduce operational costs. This is an exemplary case of win-win cooperation between open-source collaboration and enterprise practice.”
Difficulties and limits
- Low Resource Utilization
- Coarse Scheduling Granularity
- Heterogeneous Adaptation Difficulties
- Impact on ROI
Sources
- dynamia.ai — vendor self-report
Similar deployments
Unilever · Google Cloud
15-35% increase — retailer sales
Independently reportedOther · Other · Classical ML · Scaled
BHP Billiton
zero — accidents due to driver drowsiness
Independently reportedOther · Other · Classical ML · In production · 2022
DSST: College View · Playlab, Class Companion, Manta AI
Vendor-reportedEducation · Other · NLP · In production · 2024
Cubo Ai · Google Cloud
more than 10X — user growth supported with same IT workforce
Vendor-reportedOther · Other · Computer vision · Scaled · 2019
Precision HVAC & Cooling · AutoScale My Biz
15+ — High-Value After-Hours Leads Captured
Vendor-reportedOther · Other · LLM · In production