Hello everyone,
I’m currently in the planning stages of building a high-performance CUDA system and seeking advice on selecting the most robust server motherboard to support multiple GPUs. My aim is to create a system capable of handling demanding computational tasks, potentially utilizing up to 4x 295s GPUs or similar high-end cards. I’m particularly interested in insights relevant to the current hardware landscape as of late 2009.
In my research, I came across a previous discussion highlighting the ASUS P6T7 WS SuperComputer motherboard as a viable option for custom-built systems.
http://usa.asus.com/product.aspx?P_ID=9ca8hJfGz483noLk&templete=2
However, I am also considering pre-fabricated server solutions from manufacturers like SuperMicro, specifically the SuperServer 7046GT-TRF.
http://www.supermicro.com/products/system/4U/7046/SYS-7046GT-TRF.cfm
SuperMicro also offers the X8DTH-6F motherboard, which boasts an impressive 7 PCI-E 2.0 x8 slots:
http://www.supermicro.com/products/motherboard/QPI/5500/X8DTH-6F.cfm
My primary concerns revolve around the practicalities of implementing such a system. Firstly, with the SuperMicro X8DTH-6F, it’s unclear whether installing double-width GPUs will obstruct adjacent slots. The product images suggest wider spacing, but confirmation on whether it can truly accommodate 7 double-width cards, or if it’s limited to 4, would be greatly appreciated. Powering 7 high-end GPUs also raises questions about power supply requirements and overall system stability.
Secondly, the X8DTH-6F utilizes PCI-E 2.0 x8 slots instead of x16. I need to understand if this difference in bandwidth significantly impacts CUDA performance. For compute-intensive tasks, is the x8 bandwidth sufficient, or is x16 crucial for maximizing GPU utilization?
Finally, considering the imminent release of Fermi architecture GPUs, I am keen to know how these current Server Motherboards and systems are expected to perform with the new generation of cards. Will these platforms remain relevant and effective for Fermi-based CUDA systems?
Ultimately, I’m looking for guidance on the optimal hardware platform currently available to maximize the number of GPUs in a CUDA system. Are the SuperMicro options, particularly the motherboards mentioned, at the cutting edge of what’s achievable today, and are they future-proofed for the transition to Fermi GPUs?
Any insights, experiences, or recommendations on choosing the best server motherboard for a multi-GPU CUDA system would be invaluable. Thank you in advance for your help.
