AMD's Fiji GPUs are not exactly stellar when it comes to double-precision FP64 computing performance but also offers impressive single-precision FP32 compute performance, with desktop Radeon Pro Duo dual-GPU reaching 16TFLOPs of compute performance.

With that in mind, the new FirePro S9300 X2 targets HPC applications that need maximum single-precision FP32 compute performance, including deep neural networks, machine learning, geoscience, molecular dynamics, data processing and analysis and development platforms for Exascale computing. Based on two fully-enabled Fiji GPUs, each clocked at 850MHz, the new FirePro S9300 X2 provides 13.9 TFLOPs of single-precision compute performance. It also comes with 4GB of non-ECC High Bandwidth Memory (HBM). 

AMD lowered the GPU clock in order to keep the TDP at 300W, which also allowed it to pair the FirePro S9300 X2 with a passive cooling, something that is really important in the HPC market. 

According to AMD's own testing, the new FirePro S9300 X2 is is up to 3.5x faster than Tesla K40 and up to 2x faster than Nvidia's flagship Tesla K80. It is also significantly faster than AMD's Hawaii-based FirePro S9170, which will still remain the fastest double-precision compute performance FirePro graphics card in AMD's lineup.

AMD's new FirePro S9300 X2 will be available in the second quarter of this year with a price set at US $5999.