AMD Announces FirePro S9300 x2 Server GPU -13.9 TFLOPS Of Compute

AMD launches a new compute card that aims to take on and topple NVIDIA’s Tesla GPU line. The new S9300 x2 is a dual GPU configuration (as hinted by the name), that sports some eye-watering numbers that’ll leave most enthusiasts salivating. We’ll start with the headline, 13.9 TFLOPS of single-precision floating point calculations per second (FP32), setting the S9300 x2 as the single fastest dual GPU card on the market. On top of that is an impressive 1TB/s (that’s a terabytes) of memory bandwidth per card.

These compute cards are not like normal workstation cards you’ll find in desktops, but headless and passively cooled beasts that will find their way into server farms and various supercomputers. There are no display adapters on the back, and they require a constant supply of cool air from an air-conditioned server room.

The S9300 x2 is AMD’s first HBM equipped professional compute card; yes, AMD is starting off with a dual GPU configuration first, rather than a single, and it’s probably for a rather simple reason. Since these cards are based on the same third-generation GCN architecture as the Fury desktop counterpart (Fiji), it has the same limitation – 4GB of VRAM per GPU. This means the S9300 x2 has a grand total of 8GB of VRAM to play with, which does limit its appeal. However, not all workloads require large pools of memory, and one such workload is the Canadian Hydrogen Intensity Mapping Experiment – or CHIME project. Collecting radio waves with four 100-metre-long cylindrical reflectors and feeding the values into a large supercomputer, CHIME intends to map out the rate of expansion of the universe.

In many regards, the S9300 x2 is a very specialized card, with huge amounts of bandwidth and compute, but it does have its Achilles’ heel (or two in this case); the traditionally low amount of VRAM as mentioned above, and its double-precision floating point calculations (FP64).

Most workstation cards typically have half the FP64 compared to FP32 compute, but with the S9300 x2, it’s a lot lower unfortunately. While FP32 is a staggering 13.9 TFLOPS, it has just 870 GFLOPS FP64 – or 0.87 TFLOPS. AMD’s older generation S9170 could handle 2.62 TFLOPS FP64, with an impressive 32GB of VRAM too. However, not everyone needs double-precision, so it all comes back to workload.

Where the S9300 x2 will really shine is its various API support and AMD’s first real workstation card to take advantage of its new GPUOpen initiative. Heterogeneous-compute Interface for Portability – or HIP, will allow AMD to convert existing code, including NVIDIA’s CUDA, over to C++. AMD is also working on its open-source drivers for Linux system.

The S9300 x2 is a very niche card, but it’s that niche that will give it the edge as a large part of the oil and gas industry make heavy use of FP16 and FP32 calculations. It’s a brute-force compute card for small data sets that far exceeds anything else on the market currently. Initial pricing puts the card at $5999.