GTC 2018 Highlights Captured by Microway HPC Experts, Scroll down to experience the event

GTC 2018 Keynote

Jen-Hsun Huang impressed with another compelling keynote. The latest developments included

  • A New 32GB Quadro GV100 GPU designed for professional visualization & realtime ray tracing (pairs in SLI to yield 64GB of GPU memory)
  • A New 32GB Tesla V100 GPU, shipping immediately in NVIDIA DGX platforms and shipping imminently elsewhere
  • A new design for 16 Tesla V100 GPUs in a single addressable 512GB fabric, dubbed the "World's Largest GPU". Enabled by a new NVSwitch for NVLink.
  • NVIDIA DGX-2, a new 16-GPU appliance with the world's first implementation of the above
  • TensorRT 4, a massive update to the popular library for high performance DL Inference
  • Kubernetes for NVIDIA GPUs
NVIDIA GTC 2018 Keynote Highlights

Watch the Entire Keynote or 15 Min Supercut

Expo Floor at GTC2018

The new DGX-2 was the hottest feature in NVIDIA's GTC booth. Pictured here is the DGX-2 with each of its two 8 GPU sleds visible. The host server, with 2 Xeon Scalable Platinum Processors and 1.5TB of memory, sits below.

Expect the world's leading AI frameworks to see the first optimization for DGX-2's 512GB of GPU memory. This should appear as a single space for DL training sets. Other large scale simulation challenges may map well to this architecture as well.

At 10U, 350lbs, and 10.5kW of power consumption, DGX-2 will be a beast for any datacenter.

NVSwitch and 32GB Tesla V100 GPU in DGX-2 Sled

The new NVSwitch is the glue that makes DGX-2 work . On the left side (blue PCB) of this photo is the new NVSwitch. 12 NVSwitches are implemented in the DGX-2 design. It's a whopping 14TB of GPU-GPU bandwidth.

DGX-2 ships in Q3 and is priced at $399k USD.

Tesla V100 GPUs were seen across the GTC Expo floor. If SC17 was the initial technology debut, GTC felt like the coming out party for Tesla V100.

More platforms than ever are available, and system designs have finally begun shipping with the latest Xeon Scalable Processor host-CPUs.

NumberSmasher 1U GPU Server with NVLink

Microway's NumberSmasher 1U GPU Server with NVLink, on display in our Booth 625, is available now with the new 32GB Tesla V100 GPU.

GPU Sled in the Microway Octoputer™ 8-GPU Server with NVLink

Updates for Xeon Scalable Processors and 32GB Tesla V100 aren't just for 1Us. Here's the 8 NVIDIA Tesla GPU with NVLink sled from our Octoputer with NVLink (now available with the latest Xeon hosts). It features the same GPU connectivity as a DGX-1.

Block Diagram of DGX-1 and NumberSmasher 4U Octoputer with NVLink
Power Systems AC922

Walk by the IBM booth? You may have have missed the Power Systems AC922, the platform that makes CORAL tick. It stands out as the:

  • First-ever system with CPU:GPU Coherence, for world's simplest GPU programming (note: same simple "1 big memory space" programming experience as a DGX-2, and available today)
  • Only platform with NVLink from CPU:GPU
  • 5X the CPU:GPU bandwidth of x86 CPUs
  • Only platform with PCI-E 4.0

Learn more about the architecture of GPU-accelerated platforms with our Tesla V100 GPU Solutions Guide.

DGX Station is your AI workstation appliance, and it attracted a lot of eyes at SC17. It brings NVIDIA AI tools, an easy-to-use containerized software platform, cloud bursting capability, and automatic framework updates to deep-learning and AI professionals.

Most Popular Session Highlights

Architects for the Intersection of HPC & AI: Microway

Need help navigating the world of HPC & AI? Microway experts can: provide you more insight on what went on at GTC, help you with a solution design, or simply a provide quote on a hardware configuration.


Microway, Inc.

Report Abuse

If you feel that this video content violates the Adobe Terms of Use, you may report this content by filling out this quick form.

To report a Copyright Violation, please follow Section 17 in the Terms of Use.