| |  | bullx B505 accelerator blade |
|
| | • Embedded GPUs eliminating bottlenecks | | |
|
| bullx B505 accelerator blades are designed for full bandwidth between each GPU and host CPU, and for double interconnect bandwidth between blades.
|
| |
| Double-width blade |
| 2 NVIDIA Tesla M2050/M2070/M2070Q/M2090 GPUs |
| 2 Intel® Xeon® 5600 CPUs |
| 1 dedicated PCI-e 16x connection for each GPU |
| Double InfiniBand QDR connections between blades |
|
| |
| | |
| |
| | | | | | |
The bullx chassis can host indifferently 18 single-width compute blades (B500 model) each containing two CPUs, or 9 double-width accelerator blades (B505 model) containing two CPUs and two GPUs - or any combination of those two types of blades. |
|
The bullx B505 accelerator blade packs two Intel® Xeon® 5600 quad-core CPUs and two NVIDIA® Tesla™ 20-Series GPUs in a double-width blade. Each NVIDIA Tesla computing card offers 488 cores, up to 1.15GHz (3GB or 6GB total dedicated memory per GPU, depending on model). |
The bullx B505 accelerator blades are the only blades on the market designed for full bandwidth between each GPU and host CPU, and for double interconnect bandwidth between blades |
|
To take the most out of GPU accelerators, they should be integrated as much as possible within the host server:
A key feature to optimize application efficiency is the bandwidth between the accelerator card and the host processor. To maximize the performance of the two GPUs, Bull has therefore chosen to have one dedicated PCI-e 16x connection for each GPU.
For applications parallelized over several blades, such as Reverse Time Migration algorithms, it is important to minimize data transfer time between GPUs, and thus to support a large bandwidth between blades. The bullx B505 blade includes two InfiniBand QDR network connections, so that each GPU has access to QDR bandwidth. The two InfiniBand QDR ports are connected to the first level switch integrated in the bullx chassis.
|
|
|
| | |
I 
International [
