NVIDIA is now shipping their 4.58 TFLOPS single-precision floating point GPUs. The Tesla K10 GPU Accelerators, based upon the Kepler GK104 architecture, are the first Teslas available from this new generation of products. They are designed for single-precision float-point applications, so double-precision users will need to wait until winter for the Tesla K20 GPU Accelerators.
Although Tesla K10 is weak in double-precision performance, many users will find it has a lot to offer:
- PCI-Express x16 generation 3.0 link
- Dual GK104 GPUs, each with 4GB GDDR5 Memory
- 8GB Total Memory (2 x 160 GB/sec peak bandwidth)
- 4.58 TFLOPS peak single-precision floating point
- 190 GFLOPS peak double-precision floating point
One significant drawback is that all Tesla K10 GPUs are passively cooled, so you won’t be able to install them in your existing workstations. You may be able to install them in your existing GPU servers and you may be able to purchase a new workstation with Tesla K10, but there are compatibility/noise issues. Contact us to discuss your options.
It’s also worth noting that there are different versions of the product depending on which direction of airflow you require (back-to-front or front-to-back). Here’s a shot of the Tesla K10 options:
The NVIDIA product documentation does not offer an extremely clear description of the airflow. In fact, many people I’ve spoken to assumed the air flowed in exactly the opposite direction! Given the way the heatsinks are designed (as shown below), installing a K10 backwards makes the cooling situation twice as bad.
If a tower/workstation form-factor is required, we do have one available. Unfortunately, it’s rather noisy. There is no quiet configuration on the market for Tesla K10. NVIDIA has already announced a workstation version of the Tesla K20, but it won’t have the same level of single-precision performance.
For those curious as to why there are two separate product versions, I’ve taken a close-up shot of the heatsinks. There are two different versions of the heatsink, with varying width between the fins. This prevents the first GPU from pre-heating the air too much before it reaches the second GPU.
Despite the compatibility warnings above, the Tesla K10 is compatible with the majority of our Tesla GPU-accelerated servers and clusters. We’re happy to help if you’re not certain which GPU suits you best or simply to make sure you purchase the right version of the Tesla K10!