Tesla K80 packs dual GK210 GPUs

— 9:50 AM on November 17, 2014

Nvidia unwrapped a new GPU computing card earlier this morning: the Tesla K80, which it touts as the "world's fastest data accelerator" for scientific computing and data analytics. Judging by the hardware under that slick green-and-black cooling shroud, it's hard to dispute that claim.

The Tesla K80 is powered by two "Big Kepler" GK210 graphics processors, each with 2496 stream processors and ungodly amounts of memory bandwidth. We haven't seen the GK210 in a consumer graphics card yet, but the chip's older sibling, the GK110, powers Nvidia's top-of-the-line GeForce Titan Z.

The chart below compares the new Tesla K80 with the Tesla K40, which features a single GK110B chip and was the fastest member of the Tesla family until today:

Product Tesla K80 Tesla K40
GPU 2x Kepler GK210 1 Kepler GK110B
Stream processors 4992 (2496 per GPU) 2880
Peak double precision
floating point performance
2.91 Tflops (GPU Boost)
1.87 Tflops (base speed)
1.66 Tflops (GPU Boost)
1.43 Tflops (base speed)
Peak single precision
floating point performance
8.74 Tflops (GPU Boost)
5.6 Tflops (base speed)
5 Tflops (GPU Boost)
4.29 Tflops (base speed)
Memory bandwidth (ECC off) 480 GB/s (240 GB/s per GPU) 288 GB/s
Memory size (GDDR5) 24 GB (12GB per GPU) 12 GB

Yow. The Tesla K80 has considerably more floating-point power and memory bandwidth than the K40. The total memory capacity is higher, too, although each GPU still has access to 12GB of RAM. Interestingly, the difference in number-crunching power seems to be especially stark when GPU Boost kicks in. According to AnandTech, the K80 has a lower base clock speed but the same peak Boost speed as the K40. The Kepler architecture is slowly being supplanted by Maxwell on the desktop, which makes Nvidia's decision to introduce a new Kepler-based GPU for this Tesla card somewhat surprising.

The GK210 also has double the register file size (512KB) and twice as much L1 cache/shared memory (128KB) per SMX as the GK110B. The additional local storage should allow the SMX to achieve more constant utilization in GPU-computing workloads.

Nvidia says it conceived the Tesla K80 to face the "most difficult computational challenges" in fields like astrophysics, genomics, quantum chemistry, and "advanced deep learning." The card is shipping today, and Nvidia names more than a dozen server vendors that will offer it in their systems, including big names like Cray, Dell, and HP.

Tip: You can use the A/Z keys to walk threads.
View options

This discussion is now closed.