Current slide {CURRENT_SLIDE} of {TOTAL_SLIDES}- Best-selling in Graphics/Video Cards
Current slide {CURRENT_SLIDE} of {TOTAL_SLIDES}- Save on Graphics/Video Cards
In my opinion, this is the most cost-effective board for GPU computing in the market. If you compare price/performance of this GPU to Pascal or Volta, there is a clear advantage. The K20 greatest strength is its high double precision performance. Many of the newer GPU are optimized for AI and single or half-precision. This board includes 5 GB of GDDR5 RAM and has the peak double precision 1.175 TFLOPS (3.524 TFLOPS in single precision). Pascal is perhaps 3x faster but it also costs 15x more. The greatest limitation of this board is its support for only PCIe 2.0, not 3.0 as in the later GPU such as K40. This throttles host-to-device transfers at about 6.1-6.5 GB/s but for many compute tasks this is acceptable. The GPU I received were like-new, probably never actually used. All worked as expected. One thing to be aware is the RAM requirement by this server version of the K20. It allocates 16 GB of pinned memory per board from the main RAM. If you have less than 64 GB with two boards installed, the driver won't load and the device won't be usable. So make sure you have tons of RAM for this GPU. I am using these boards with DELL CloudEdge C410X external PCIe2 enclosure. It allows me to hook as many as 8 GPUs per server. But a more optimal setup is to assign 2 or 4 GPU per server because of the PCIe bandwidth limitation. With 2 boards per server I am seeing 71% efficiency in HPL 2.0 for CUDA, and with 4 boards per server (with 2 HICs) the efficiency drops to 61%. All in all, this is an excellent entry-level board for GPU computing. It will allow you to write and debug your programs and get ready to run them on the upcoming generations of even bigger CUDA boards.Read full review
Verified purchase: Yes | Condition: Pre-owned
Verified purchase: Yes | Condition: Pre-owned