AMD announces new dual-GPU FirePro S9300 x2 designed for HPC and scientific computing

Not long ago, AMD revealed its first double GPU in light of its most astounding end Fiji GPU. The Radeon Pro is obviously a specialty card with a $1500 cost and a particular spotlight on creating for VR — yet AMD clearly has outlines on more than one business sector. Today the organization reported its FirePro S9300 x2 — the first and likely just Fiji-inferred GPU the organization will dispatch for the HPC space.

The S3200 x2's customer roots are obvious in its determinations. Each GPU is constrained to 4GB of memory in a business sector where 8GB RAM supports are viewed as low-end and top-end cards can pack as much as 32GB of RAM. Not at all like the Hawaii-based S9170, the S9300 x2 needs bolster for ECC memory and its twofold exactness gliding guide execution is restricted toward 1/16 of FP32 (0.8 TFLOPS). The S9170, interestingly, presents to 2.6TFLOPS of twofold exactness execution.

The S9300 X2's essential quality is its single-exactness drifting point execution. At 13.9TFLOPS, it's the most effective single-accuracy card AMD has ever discharged and hypothetically quicker than anything in Nvidia's present stable of equipment also.

The S9300 x2 is intended for uninvolved cooling and a 300W force envelope, which implies clock rates are significantly lower than what we normally find in a card of this sort. AMD hasn't uncovered the normal clock speeds on the Radeon Pro Duo, however the GPUs it demoed at GDC had three eight-pin PCI Express power connectors. While organizations normally overprovision their GPUs to guarantee they can deal with sharp transient spikes, the Radeon Pro Duo is obviously intended to force more than 300W.

On the other hand, this matches what we saw with the Radeon Nano. Fiji's energy utilization rises forcefully as its clock speed expands; keeping the GPU at a generally low 850MHz guarantees it can fit inside of its normal force envelope

The HBM bottleneck 

At the point when AMD chose to utilize high data transmission memory (HBM) for its Fiji GPU, it made a cognizant tradeoff between memory limit and crude transfer speed. The benefit of HBM is that it gives Fiji a 512GB/s memory transport in a much littler impression while utilizing significantly less power. The burden of original HBM is that it's restricted to a 4GB memory support.

Following the time when AMD reported Fiji, there's been hypothesis that the organization may utilize some sort of workaround to construct a GPU with more than 4GB RAM. On the off chance that such an innovation was always going to show up, the S9300 x2 is the place it would be conveyed — the card's $6000 sticker price gives it space to hold up under the expense of advancement.

The other side to this, be that as it may, is that the HPC field has a tendency to be ruled by industry specialists who know about the memory necessities of their own product. AMD has officially declared it will give a HPC bunch to handling information accumulated by the Chime telescope, as appeared above, and it likewise trusts it can make advances in the oil and gas business. These are all workloads where single-exactness skimming point execution is above all else, and the S9300 x2 ought to exceed expectations in these zones.
 
A niche product, but an important step
 
The S9300 x2 is a specialty item even inside of the restrictions the HPC space, however it's still great to see AMD venturing up to the plate. One reason Nvidia has delighted in much more advantageous edges contrasted with AMD is that it holds between 75-85% of the expert GPU market (the rates move to some degree relying upon the quarter) and the greater part of the HPC space. Of the Top 500 frameworks that utilization some sort of co-processor, 66 utilize a Nvidia GPU, 29 depend on Xeon Phi, and only three use AMD Radeon.

Given that AMD GPUs now can keep running in any event some CUDA code, on account of the organization's Boltzmann activity, the contention for utilizing AMD might have quite recently gotten more grounded than it has ever been. Generally AMD's top of the line workstation and expert GPUs are less expensive than Nvidia's (and yes, $6000 is shabby in this business sector). We'll check whether these progressions yield organic product in the quarters to come.