The parallel computing applications benefit from such a high allocation of video memory, and that’s why AMD has pushed well forward with its new FirePro graphic card for GPU computing.
AMD has announced a new card FirePro family, specifically targeted for use related to GPU computing and parallel computing with high performance. FirePro S9170, the name of the new card, goes beyond what is offered by the AMD FirePro S9150 model proposing an amount of video memory onboard never seen before in discrete video cards.
We speak of as many as 32 GB of memory GDDR5, type ECC given the intended use of this product. The GPU behind this product is always to Hawaii, also adopted by the boards of the family FirePro 9100 presented earlier by AMD; the choice not to use the chip Fiji recently unveiled by AMD with the Radeon R9 Fury is related to higher-performance calculations with double precision achievable with the chip Hawaii.
Then we find GPU with 2,816 stream processors, capable of a peak performance of 5.24 TFlops in single precision and 2.62 TFlops in double precision, while 32 GB of video memory are capable of a theoretical maximum bandwidth of 320 GB per second thanks the combination between bus from 512 bit and the actual clock frequency of 5 GHz.
Compared to the new card FirePro S9150 boasts a budget of video memory, which is double, from 16 to 32 GB, as well as a slight increase in the clock frequency of the GPU. The TDP said amounted to 275 Watt but you can configure the card in a mode with lower consumption for a final TDP of 235 watts.
The cooling is passive, as clearly visible from the image: this card is designed for installation in which accelerator rack server and is delegated to the server the flow of cooling air essential to the correct operation of the product.
A product designed specifically for parallel computing, with which AMD wants to be in direct competition with NVIDIA Tesla belonging to the family.