GRAPHICS CARDS FOR PROFESSIONALS!
Nvidia has introduced six new professional Quadro graphics cards. The top model uses the large GP100 GPU and HBM2. Up to now the series was largely based on Maxwell. Only the Quadro P6000 with the fully activated GP102 and 24GB GDDR5X and Quadro P5000 with GP104 and 16GB GDDR5X took Pascal.
First GP100 off Tesla
The Quadro GP100 uses the most complex Pascal GPU that only for the high-performance computing market in the form of far Tesla P100 has been available. Again, the Quadro GP100 does not use the full range. Instead of the 3,840 shader units, for example, with the fully GP102 on the Quadro P6000, on the Quadro GP100 only 3.584.
Quadro GP100 and P6000 each have their own special areas
The GP100 calculates FP32 data with full performance and masters Double Precision (FP64) with half the performance. This makes it significantly faster than the GP102 on the Quadro P6000. In addition, FP16 calculations can be performed at twice the speed, which the Quadro P6000 also does not master. For this the GP100 can not accelerate INT8 calculations – the quadro P6000 with the quadruple speed but already. Depending on the application area, either a Quadro GP100 or a Quadro P6000 is required.
In addition, the Quadro GP100, like the Tesla counterpart, uses 16 gigabytes of Samsung’s storage technology HBM2. There are also four HBM2 stacks, each with four gigabytes. The storage bandwidth is 720 gigabytes per second, since Nvidia does not use the full 1,000 MHz of HBM2. The TDP gives Nvidia with 235 watts, 15 watts less than with the Quadro P6000. For the Quadro GP100, Nvidia also offers an NVLink bridge for communication with each other if two should be used instead of a Quadro GP100.
The other new Quadro models use GP104, GP106 and GP107
The remaining five new Quadro products round off the professional line-up. The Quadro P4000, like the Quadro P5000, uses the GP104 GPU, but deactivates some units so that 1,792 shader units remain. The memory expansion is also halved to eight gigabytes and instead of GDDR5X- there are classic GDDR5 memory. The TDP is 105 watts, 75 watts less than the Quadro P6000. With a lower power loss, a single-slot cooling system is sufficient.
Quadro P2000 with 160-bit memory interface
The remaining four Quadro cards are significantly less powerful. The Quadro P2000 is based on the GP106 GPU with 1,024 shader units and five gigabytes of GDDR5 memory. The GPU actually has a 192 bit interface, of which on the P2000 a controller is switched off – 160 bits remain. This also explains the unusual storage expansion. The Quadro P1000 uses a shorted GP107 with 640 shader units and offers four gigabytes of memory with a TDP of only 47 watts. With the Quadro P600 only 384 ALUs remain with two gigabytes and 40 watts, with the smallest Quadro P400 it is only 256 ALUs with two Gigabytes and 30 Watts. All new models will be available in March.