When you're playing games at 4K resolution, the Titan X's 12GB of memory is overkill, but hardcore high-performance computing tasks can gobble up all the RAM you're able to throw at it. AMD's new FirePro S9170 satiates the professional graphics crowd by dropping a mammoth 32GB of GDDR5 memory onto the board, practically sneering at the 24GB of RAM that comes standard on Nvidia's dual-GPU Tesla K80.
AMD says the FirePro S9170 is "the world's first and fastest 32GB single-GPU server card." Nvidia's next-gen Pascal GPUs, due in 2016, will also be able to support 32GB of RAM.
AMD's press release stays mum on processor specifics besides saying it's based on the second-generation Graphics Core Next (GCN) GPU architecture. That said, the lack of cutting-edge high-bandwidth memory technology, paired with the 512-bit memory interface and 320GBps of total memory bandwidth, suggests the FirePro S9170 is based off the Grenada core found in AMD's R9 390X, rather than the company's hulking new Fiji GPU.
While memory's a major concern for compute uses, it's not the only one. The FirePro S9170 offers an eyebrow-raising 2.62 teraflops of peak double precision performance--far more than the 1.66 TFlops Nvidia's single-GPU Tesla K40 offers, and even more than the dual-GPU Tesla K80 pumps out when humming along at base speed. (The Tesla K80's boost clock capabilities can give it an edge over the FirePro S9170 when needed though, topping out at 2.91 TFlops.)
AMD's new 275W card hits 5.24 TFlops of peak single precision compute performance, and includes full OpenCL 2.0 support. Further details have yet to be disclosed. Look for the FirePro S9170 to land at the top of AMD's professional graphics product stack sometime in the third quarter.
Sign up for Computerworld eNewsletters.