White Papers

Extreme GPU Computing
Page 20
6. Conclusion
In conclusion, the C4130 meets the current challenges of a high-density, accelerator-enabled compute
node. Targeted specifically towards the HPC market, it offers world-class performance and unique
configurability options to fit extreme HPC requirements. The following are the main results:
 The various configurations of the C4130 offer a range of CPU to GPU ratios
 C4130 offers both balanced and switched connections between CPU and GPUs
 The unique placement of GPU in the C4130 ensures optimal GPU performance
 C4130 offers a wide range of GPU boards and coprocessors
 On the K80s, the HPL performance is about 9X compared to CPUs resulting in 4.2 GFLOPS per
watt power efficiency
 Using the K80s can substantially accelerate industry-standard molecular dynamic codes.
Observed acceleration is from 2X to 16X depending on the code and benchmark used. This
improvement is achieved in a power efficient manner, consuming only 2X to 4X more power.
 Compared to the previous generation, the C410X-based GPU solution using M2070s, the C4130
with K80s offers a vastly improved solution. Comparing four GPU boards, the following are the
main enhancements:
o The performance on HPL is 5X to 8X better with reduced power consumption,
resulting in a 9X to 10X performance per watt improvement
o The performance on NAMD is 3X to 4X better with reduced power consumption,
resulting in 6X to 7X performance per watt improvement
o On a “GPU per U” basis the compute density is improved 2.5X to 3.5X. Previously
the C410X based solution required at least 5U to 7U for 16 GPUs (sixteen M2070); with
the latest C4130, users could have access to 16 GPUs (eight K80s) in only 2Us.