Benefits of Cloud GPU over Consumer GPU

Cloud GPU

Consumer GPU

24/7 Uptime

Guaranteed quality in a server 

Server-Grade Engineering 

> Lower operating voltage for longer reliability 

> Zero-error testing at aggressive clocks > Error Correction Code (ECC) for data integrity

PC-Grade Engineering

Reduced thermal stress for uncompromised reliability

Forced Air Cooling Design

 > Designed for maximum airflow in a server 

> Lower GPU temperature for reliability 

> Lower power consumption

Active Fan Design > Fan works against server airflow 

> GPU runs up to 30-40% hotter, increasing failure rate 

> Higher power consumption

Higher data center availability and serviceability

Dynamic Page Retirement 

> Monitors and removes bad memory with simple reboot 

N/A: Need to physically remove GPU with bad memory

Scalable Performance

Scalable Application performance across nodes

GPU Direct RDMA 

> Direct transfers between GPUs 

> 67% lower latency 

> 5X higher GPU-to-GPU MPI bandwidth

N/A

Strong scaling performance in a node

NVIDIA NVLink™ 

> 5X higher GPU-to-GPU bandwidth 

> Linear strong scaling for lots of GPUs

N/A

Deploy large models 

> Up to 32 GB HBM2 for Tesla V100

4 GB to 12 GB