A Brief Overview of The NVIDIA GTC Keynote

Can You Smell What The Cook Is Rocking

NVIDIA’s Jen-Hsun Huang offered a very different style of keynote for this year’s GTC, presenting NVIDIA’s newest products from his kitchen.  You will already be familiar with the A100 and how five of their DGX A100 systems are able to do the work of a data centre with 50 DGX-1 systems and do so with a mere 28 kilowatts of power as opposed to the 630 kilowatts the DGX-1’s would consume. 

There is also a smaller version, the Redstone HGX which consists of four A100 GPUs on a single add-in card.  The GPUs each have 12 NVlinks per GPU but instead of being set up to use the NVIDIA NVSwitch Redstone takes those 12 NVLinks and splits them into three group.  This mesh design is similar to previous designs such as how Tesla V100s were connected and means lower cost and lower power consumption which makes it attractive for smaller enterprises.  if you are curious how the HGX A100 is set up, you can drop by Serve The Home who took a deep look into it’s architecture.

NVIDIA’s Isaac robotics platform has been adopted by the BMW Group, a big win for NVIDIA as they will be used for training the assembly line robots at their factories.  If it is networking you are more interested in then check out their Mellanox ConnectX-6 Lx SmartNIC, a dual 25GBps or single 50GBps NIC with a variety of security features built right in.  It will be compatible with PCIe 4.0 as well as PCIe 3.0 and supports GPUDirect RDMA acceleration for NVMe over Fabrics.

Last but not least is the announcement of the NVIDIA Clara healthcare platform, which among other things could turn some sensors currently being used into smart sensors.  Considering the state of the security of most hospital’s hardware, this might be a very good idea if NVIDIA’s security can mitigate many of the issues we have seen with smart devices.  Clara Parabricks will also vastly increase the speed at which computational genomics can be accomplished, it currently holds a record for analyzing the whole human genome DNA sequence in under 20 minutes.  Licenses to use that specific feature are being offered to COVID-19 researchers free of charge for 90 days.

Expect to hear more on the podcast.