Nvidia’s online GTC event was last Friday, and Nvidia introduced some beefy GPU called the Nvidia A100. This isn’t a consumer card; The Nvidia A100 is a high-end graphics card for AI computing and supercomputers. We don’t look at this hardware as it is out of our reach, but we will be summarizing the keynote and looking at Nvidia’s DGX A100 influence on the upcoming RTX Nvidia GPUs.
Nvidia introduces the new Nvidia DGX A100 GPU, the world’s largest GPU
AI learning and training space is a vast sector. All the tasks done there have used immense computing power and faster hardware. That is where the big guns like the Volta GPU or the upcoming DGX A100 GPU come into play.
They have a significant processing power, which allows them to perform extensive integer value calculations. Graphics cards provide the additional computing power added to a system for better performance but a specification set of tasks. Not all the graphics cards are tailored towards gaming, which we see in the retail space.
Nvidia’s newer Ampere architecture based A100 graphics card is the best card in the market as dubbed by Nvidia. The newer Ampere card is 20 times faster than, the older Volta V100 card. The A100 is based on TSMC’s 7nm die and packs in a 54 billion transistor on an 826mm2 die size. Yes, there is a huge technological advancement, and that is not only that, but the smaller die size and 7nm process technology also help lower the power consumption of the card.
The DGX A100 will have 8 Ampere graphics cards inside it, with each graphics card sharing 40GB of HBM2 memory. The HBM2 memory for faster memory access while computing. Nvidia increased the speed of the NVlink to up to 600 Gb/s. Don’t confuse the NVLink to the consumer-based NVLink. The faster NVLink allows the cards to 8 graphics cards in the A100 to communicate faster, improving the workflow. There are 6 NVLink on the graphics card.
Nvidia is not joking about the computing prowess of the DGX A100.
The Nvidia DGX A100 has 19.5 TFLOPS of computing power for FP64, 312 TFLOPS for FP32 training, and 1,248 TOPS for INT8 Inference. So these applications are AI training for physical simulation etc. These are massively powerful graphics cards that consume lower power thanks to the die size shrink and tailored architecture.
As for the price, each of the DGX A100 will be priced at $200K. Compared to the older V100, the pricing seems compelling for the computing power Nvidia is giving to its customers. The slide presented a bigger buyer’s remorse. The buyer’s remorse in the RTX 2070 Super and RTX 2070 looks puny compared to the supercomputers cost.
So what does it state for the RTX 3000 Series graphics card?
The Ampere architecture might not fully make it into the RTX 3000 series card. The RTX 3000 will use a 7nm process node from TSMC. The new tensor cores in the A100 also hints at more modern tensor cores for the RTX 3000 series.
There will be no changes to the RAM specification; All the RTX 3000 series cards come with DDR6 memory. Transistor counts will increase in the newer RTX 3080Ti by 30%, and we can see a performance uplift up to 40%. With the shrunk die, the RTX 3000 series card will be able to outperform or perform equally to the RTX 2000 series with the same TDP or even lower.
We still don’t know how much Nvidia will charge us for the newer RTX 3000 series card, but we hope to see it remain the same as the RTX 2000 series. If better competition between AMD and Nvidia starts to happen, we might even see a price drop. The cards are expected to release from September until then we will only get hints and a small teaser of the cards.