This past month, NVIDIA has launched new cards from what used to be the Tesla and Quadro range, that is, professionals, with very reasonable prices and such as the A10, A30, A4000 and A5000, which complement the existing A100 and A6000.
All of them with PCI-E 4.0 connectivity, a key element now that both Intel ICELake, AMD Rome and AMD Milan have this technology, twice as fast as the PCI-E 3.0 bus.
We want to put special emphasis on the NVIDIA A30 card. This product is very revolutionary in several aspects:
1 º After NVIDIA had only left the A100 as a card with the ability to do double precision, this card comes out with more than half the Tflops of its older sister and with a price much lower than half.
2 º Meets the new standard NVIDIA MIG (Multi-Instance GPU) that allows virtualizing this card at the hardware level, without the need to resort to additional software and with Open Source tools.
3 º Its consumption of 165 W makes it one of the lowest on the market. It offers 5,2 Tflops in double precision. If we compare it with the famous K40, with a consumption of 235 W, which offered a performance of 1,43 Tflops, the progress has been spectacular. Even if we compare it with the Tesla V100, which offered 7 Tflops and a consumption of 250 W, which was only released 2 years ago, the energy efficiency is much higher.
Therefore, the new Tesla Ampere range, and especially the A30, becomes a very flexible solution that allows (depending on the models):
A) Dividing the card to be able to attend to various processes in the SLURM queues and in the virtual machines with Open Stack, amortizing purchase and consumption costs (only for the A100 and A30 models with the MIG standard)
B) Add cards, to address Deep Learning and Artificial Intelligence solutions, through NVLINK connectivity (available on A5000, A6000 and A40 models on PCI-E 4.0 and A100 40 and 80 GB models on SXM4 models)
C) Being able to face calculations in double precision (A100 and A30 models only), single precision and Tensorflops, in all models.
Therefore, with this new range, as NVIDIA has accustomed us, from one generation to another, there are improvements of around 50% in performance. Of course, all these new cards already work on Cuda 11.x and it has an ever-growing software ecosystem to take advantage of them.
We do not want to finish this analysis, without pointing out all the software that NVIDIA has released, to virtualize GPUs, both for the graphical environment and for computing power. In addition, software such as Parabricks allows genetic calculations to be carried out on two GPUs, which until now took 1 week, in a single hour.
SIE is NVIDIA's Preferred Partner in the Compute and visualization specialty, in addition to selling the entire range of this company's software products, offering its customers a quality advisory service, as well as being able to remotely test different cards to see the real performance with customer applications.