On Monday Nvidia announced a major deal with Fujitsu to help build an AI supercomputer for RIKEN using 24 DGX-1 servers. Midweek at the Open Compute Project (OCP) Summit in Santa Clara, Calif., the GPU technology leader unveiled blueprints for a new open source Tesla P100-based accelerator – HGX-1 – developed for clouds with Microsoft under its Project Olympus. (We’ll make an educated guess that the D in DGX-1 stands for Deep Learning and the H in HGX-1 for Hyperscale.) At roughly the same time, Facebook introduced Big Basin, the successor to its Big Sur GPU server, which also uses Nvidia P100s (in a similar 8-way configuration, which we’ll get into in a moment). And in the embedded world, Nvidia announced the Jetson TX2, billed as a “drop-in supercomputer,” with an ARM-based CPU supporting Pascal graphics.
That’s a productive week by any standard and there are multiple threads to follow here. Most of the activity was driven by artificial intelligence/deep learning’s continued drive into upper-end HPC and the cloud. Nvidia has been striving to leverage its GPU strength in both traditional scientific computing as well as in AI/DL whose applications often require lower precision (32-, 16-, and even 8-bit) computation.
Roy Kim, director Tesla Product Management, described the adoption of AI/DL as a revolution gathering speed fast. “The deep learning and AI revolution, even though it is huge, is also fairly young. A few years ago people were still asking the question, what is deep learning. Now every cloud vendor is asking how it can be AI-ready,” said Kim. A standardized HGX-1 design will make that possible, he contends.
The emergence of open source hardware for the cloud via OCP and Olympus is reminiscent of the emergence of the ATX ‘standard’ for PCs. The HGX-1 will be used as part of a standard AI/DL reference platform and enable cloud providers to rapidly develop AI/DL offerings, according to Kim.
Here’s a brief summary of Nvidia’s busy news week:
- HGX-1. Think DGX-1, without the CPUs. It’s an accelerator box with eight Tesla P100s, connected in the same hypercube mesh as the DGX-1 and also leveraging the NVLink interconnect. The HGX-1 hooks to servers via PCIe interface. Developed under the Olympus program guidelines, the design is open source such that users could easily take the files to their preferred ODM for manufacture. It will be interesting to see how cloud providers respond and whether significant tweaking takes place to optimize the HGX design for particular AI/Dl workloads.
- Big Basin. Facebook says Big Basin trains models that are 30 percent larger because of enhanced throughput and an increase in memory from 12 GB to 16 GB. “In tests with popular image classification models like ResNet-50, we were able to reach almost 100 percent improvement in throughput compared with Big Sur,” according to Arlene Gabriana Murillo’s FB blog. Designed as a JBOG (just a bunch of GPUs) to allow for the complete disaggregation of the CPU compute from the GPUs, it does not have compute and networking built in, so it requires an external server head node. “By designing [Big Basin] this way, we can connect our Open Compute servers as a separate building block from the Big Basin unit and scale each block independently as new CPUs and GPUs are released,” says FB in a blog Built in collaboration with ODM Quanta Cloud Technology, the Big Basin system also features Tesla P100 GPU accelerators.
Fujitsu AI Supercomputer. The new RIKEN machine will include 24 DGX-1 systems as well as 32 Fujitsu PRIMERGY servers and is expected to reach 4 petaflops peak performance when running half-precision floating point calculations. The new supercomputer is scheduled to go online next month and will be used to accelerate AI research in medicine, manufacturing, healthcare and disaster preparedness.
- Jetson TX2. A replacement for the Jetson TX1, the embedded module (SoC) features Pascal graphics with 256 CUDA cores while its CPU is an HMP (Heterogeneous Multi-Processor Architecture) Dual Denver plus a quad ARM Cortex-A57. Nvidia, like others, seems to be doing more with ARM, which though strong in the embedded and mobile space has struggled to penetrate the datacenter. That may be changing. Microsoft announced an ARM initiative on cloud workflows this week. “We have been running evaluations side by side with our production workloads and what we see is quite compelling. The high Instruction Per Cycle (IPC) counts, high core and thread counts, the connectivity options and the integration that we see across the ARM ecosystem is very exciting and continue to improve,” wrote Leendert van Doorn of Microsoft in a blog.
The FB Big Basin and Microsoft embrace of HGX-1 suggest some of different ways in which Nvidia GPU technology may be deployed by cloud vendors. The Microsoft HGX-1, built by Ingrasys (a Foxconn subsidiary), is flexible in the sense that the HGX-1 is deliberately designed to accommodate differing AI/DL workloads.
“[For Facebook], it’s really about their particular workloads. They talk about natural language processing, image processing, and all of this is really core to the services they provide their users. So they built a system that is best suited for their workload. The topology is very similar to the HGX-1 in that it has the same hypercube mesh and has eight Tesla P100s in the box with NVLink. The only difference is that it has been optimized and tuned for deep learning training, which means it has been hardened for DL training as opposed to HGX-1 which is highly configurable,” said Kim.
Interesting, when Microsoft describes the Olympus philosophy it says: “Project Olympus applies a model of open source collaboration that has been embraced for software but has historically been at odds with the physical demands of developing hardware. We’re taking a very different approach by contributing our next generation cloud hardware designs when they are approx. 50% complete – much earlier in the cycle than any previous OCP project. By sharing designs that are actively in development, Project Olympus will allow the community to contribute to the ecosystem by downloading, modifying, and forking the hardware design just like open source software,” wrote Kushagra Vaid, GM, Azure Hardware Infrastructure in a fall 2016 blog.
The HGX-1 is a complete design, said Kim but that doesn’t preclude optimization. “The design itself is complete and so you can go to Foxconn and give them this design and file and say can you manufacture this for us. It’s been tested and it works. Certainly because it is open source I can imagine other cloud vendors going in and saying I could tweak this to be more efficient specifically for the target market that I am going after and that one of the benefits. I wouldn’t be surprised if that happens. I think it gives each cloud provider an ability to optimize the system for their particular workload.”
Will there be an HGX-2? “That’s a good question. The idea is that the standards do evolve to meet the needs of the evolving workloads. We are going to continue to work with our cloud vendors to provide the best answers for that. Without giving you any roadmap, we do expect it to evolve,” said Kim.