Silicon Graphics Rackable C2110G-RP5, System User'S Manual, Page: 20 / 78

6 007-5841-001
1: Introduction
Important: Check with your SGI sales or service representative prior to using any GPU not
sourced at the SGI factory.
Any combination of these cards (up to a total of four) may have come bundled with the system.
Power for the GPU cards is provided via a GPU power cable from each of the GPUs to JPW3, 4,
6 and 7 on the serverboard (one cable for each card).
Figure 1-2 on page 7 shows a general block
diagram of the C2110G-RP5 server’s processor and I/O chipset.
GPU Features
Each of the GPUs will feature some or all of the following:
• Hundreds of GPU cores in each card that can deliver more than 660 Gigaflops of
double-precision and over one Teraflop of single-precision calculations.
• ECC protected internal register files, L1/L2 caches, shared memory, and external DRAM.
• Up to 6 GB of GDDR5 memory per GPU enhances performance and reduces data transfers
by keeping larger data sets in local memory attached directly to the GPU.
• Integrates the GPU subsystem with the C2110G-RP5 server’s monitoring and management
capabilities such as IPMI.
• Onboard L1 and L2 caches that accelerate algorithms and sparse-matrix multiplication.
• Provides faster context switching, concurrent kernel execution and improved thread block
scheduling.
• Enhances overall system performance by transferring data over the PCIe bus while the
computing cores are processing other data.
• Provides a flexible programming environment with broad support for various programming
languages and APIs.
Cooling System
The 2U server chassis has a cooling design that includes five internal 8-cm heavy-duty system
cooling fans located in the chassis. An air shroud channels the airflow from the system fans to
efficiently cool the processor and GPU areas of the system. Each power supply module also
includes an internal cooling fan. All chassis and power supply fans operate continuously.