NVIDIA Tesla K80

What it is

The Tesla K80 is an accelerator board: a piece of hardware that is plugged into a server to do heavy numerical work alongside the main processor. NVIDIA announced it at the SC14 conference in New Orleans in November 2014 ^[1]. It belongs to the company's Tesla line of products, which are sold for data analytics and scientific computing rather than for driving a display.

The basic idea behind such a board is that many scientific and data-processing workloads consist of huge numbers of similar arithmetic operations that can be carried out in parallel. A GPU is built from many small arithmetic units that run at once, so it can chew through that kind of work more efficiently than a general-purpose processor, which is optimised for doing one thing after another.

Two GPUs on one board

The distinguishing feature of the K80 is that it is not one GPU but two. NVIDIA's board specification lists two GK210 GPUs on a single board, with a maximum board power of 300 W ^[2]. Across those two chips the board carries 4,992 CUDA cores ^[1]. CUDA cores are the individual arithmetic lanes inside an NVIDIA GPU; the more of them a chip has, the more operations it can issue in parallel.

Memory is split the same way. The board carries 24 GB of GDDR5 memory, 12 GB per GPU, giving 480 GB/s of memory bandwidth per board ^[1][2]. Memory bandwidth matters because many data-heavy workloads are limited not by how fast the arithmetic units can calculate but by how quickly data can be fed to them from memory.

The performance NVIDIA claimed

NVIDIA said the K80 reached up to 8.74 teraflops of single-precision peak performance and up to 2.91 teraflops of double-precision peak performance ^[1]. These are peak figures, meaning the theoretical maximum the hardware can sustain under ideal conditions rather than what a real application will see.

The gap between the two numbers reflects a deliberate design trade-off. Double precision, which keeps roughly twice as many bits per number as single precision, is the usual requirement in scientific simulation, where rounding error accumulates over long calculations. Hardware that is fast in double precision is more expensive to build, so accelerators are typically much quicker in single precision than in double.

NVIDIA also said the K80 delivered nearly twice the performance and double the memory bandwidth of its predecessor, the Tesla K40 ^[1].

Where it sits

The K80 shows a pattern that became common in data-center accelerators: rather than making a single ever-larger chip, put two GPUs on one board and let software treat them as a pair of devices. That keeps each chip within a manufacturable size while still raising the compute and memory a single slot in a server can provide.

The 300 W ceiling is part of the same story. A server rack can supply only so much power and cooling per card, so generational gains have to come from doing more work inside a similar power envelope — here by doubling the memory bandwidth and the number of processing cores relative to the previous board ^[1][2].