HPC cloud specialist Nimbix this week announced the addition of IBM’s Power9 technology to the Nimbix cloud starting with an early access “gated program” available now to be followed by broader access in the spring. Timing of the announcement coincided with IBM Think 2018 Conference being held in Las Vegas and Nvidia’s GPU Technology Conference (GTC) scheduled for next week in San Jose. IBM also announced expanded access to Power9-based servers on IBM’s Cloud (read our coverage here).

Roughly a year and half ago Nimbix jumped onto the IBM/OpenPOWER bandwagon by adding Power8-based Minsky machines from IBM, including the IBM PowerAI bundle of optimized AI tools (e.g., frameworks, databases, libraries, etc.). This follow-on commitment, say both Nimbix and IBM, is a sign of strengthening traction for Power-based platforms.

“When we made the first investment, which to some extent was a risk because we were all x86, we partnered with IBM because we felt it was a strong offering and to showcase [our] heterogeneous computing strength,” Nimbix CEO Steve Hebert told HPCwire at Think 2018, “[We also wanted] to show that no matter what the hardware is, whether it is Nervana chips from Intel or TPUs from Google, our software stack – JARVICE – doesn’t care.”

Nimbix has long positioned itself as a leader in providing cloud-based heterogeneous computing but centered on using FPGAs and GPUs not alternative CPUs. Nimbix does not currently support the Arm processor, which might now seem like a possible next step. Regardless, Nimbix remains a strong Intel shop – it is currently undergoing a massive processor refresh swapping in Skylakes and around 80 percent of workloads run on Nimbix are on x86 – but its willingness to offer Power-based systems reflects debate in the HPC and enterprise community around which technologies will win in the emerging AI era. Hebert’s characterization of the advantages of each platform seems understated, perhaps predictably since Nimbix supports both.

“For us, quantifiable advantages generally come down to which ‘machine’ produces more results per unit time at the lowest cost. I think the Power+GPU with NVLink systems have some strong advantages for various software stacks that can take full advantage of the system bandwidth available from those systems. Nimbix continues to support many workloads on x86+PCI GPU and x86+Nvlink GPU as well since so many software stacks “just work” in those environments,” said Hebert.

“We try not to get too religious on system architectures, speeds and feeds and in fact work hard to hide all this complexity from our customers to deliver the fastest and most cost-effective time-to-results.”

The market will ultimately decide. There’s clearly a major push by IBM/OpenPOWER to lay claim to AI computational infrastructure landscape. It is being positioned as virgin territory better served by Power9 architecture, not dominated by x86, and on the edge of a massive adoption/deployment inflection point. This was certainly the central theme for IBM Think 2018.

The IBM Power9 Pitch

IBM says the Power9 processor, introduced in December, was designed from the ground up around AI/deep learning/machine learning workload requirements; its high memory bandwidth, enhanced connectivity (NVLink, OpenCapi), and Power instruction set are all tailored for accelerated computing and AI.

The new Nimbix offering is IBM’s Power System AC922 based on the IBM Power9 processor, introduced in December. IBM says Power9 was designed from the ground up with AI workloads in mind; its high memory bandwidth, enhanced connectivity (NVLink, OpenCapi, PCIe4), and Power instruction set are all tailored for accelerated computing and AI, says IBM.

Not surprisingly, IBM is forcefully marketing the AC922 as mini-Summit node. Summit is the DOE-funded pre-exascale supercomputer being built at Oak Ridge National Laboratory. It is based on IBM Power9 chips and Nvidia V100 GPUs and is anticipated to reach 200 petaflops peak. If stood up in time to be benchmarked for the June Top500 List, many expect Summit to top the list, the first time a U.S. machine has done that in several years.

Nimbix’s AC922s are essentially the same as the Summit nodes but have four instead of six GPUs (Nvidia V100s) and two Power9 processors per node. Porting the cognitive software platform to the Nimbix AC922 systems is ongoing. Nimbix CTO Leo Reiter said the PowerAI suite, available on the Power8 Minsky machines on Nimbix, is not fully ready on the Power9 platform but would be soon. “It’s early days so we are in the midst of fostering all of the cognitive software platform, all of the accelerated databases [that] we’re looking to bring up on Power9. There are some things IBM has to do but that’s all being handled very rapidly,” according Reiter.

Speaking at a press conference during Think 2018, Bob Picciano, Senior Vice President of Cognitive Systems, said “the capabilities for PowerAI will be made present [in the IBM Cloud] in the second quarter.” It seems likely loose ends will be tidied up quickly.

Nimbix finds itself in an interesting position. As the big hyperscalers wade deeper into providing advanced computing technologies (GPU, FPGA, et al) and AI/data analytics services, Nimbix’s claims of advantage built around heterogeneous computing/HPC expertise – accelerators, large memory systems, InfiniBand EDR throughout – loose some of their luster.

Hebert is aware: “It’s a mixed blessing. It’s sort of “the tide that raises all boats.” It’s also good validation. It doesn’t look like we are off on a little niche island. So that’s the good news. The bad news is what happens to a little guy like Nimbix and the big guys start doing what they are doing. The reality is we have a very unique technology platform.”

“We started Nimbix because you couldn’t get an FPGA in the cloud. So we built it. Same with GPUs. So we built it. Now as other infrastructure providers are offering those things, for us the value of our stack is [key] –  when you can take 500-plus point-and-click workloads that are our ISV partners have built and that we have built and send that into an enterprise where they can literally launch hundreds and thousands of jobs. Google, Amazon, Microsoft, none of them have that capability yet.”

Think of a pyramid sliced into layers from bottom to top, said Hebert. “With Microsoft and Google, you’ve got the infrastructure layer, you’ve got server OS, hypervisor, server OS, containers at the top. We go the other way, we have applications, containers, bare metal. For us, the point of the customers’ engagement is the apps, the ecosystem of apps. It’s very light weight. It’s agile and what that means is we will always have higher performance. If you were to take amazon’s latest metal or Microsoft’s and put in your hypervisor and your instance type, then you run an equivalent node class in the Nimbix cloud, you are going to get better performance and better economics.”

Container Technology Leader?

Nimbix’s so-called “container native” technology is a key differentiator, said Hebert, “We have been using Linux containers in HPC since 2012, well before containers were popular there. We have some very interesting intellectual property in the area of reconfigurable cloud computing. We think about cloud computing from the workload down to the metal, not from infrastructure up to the applications stack.”

“[Container technology] is pretty much going to be the way you deliver HPC apps now because of its agility. Instead of having to set up big clunky NFS (network file server) shares where if you have to update the code, you have now trashed the entire cluster if you misconfigure something so this solves all of these problems. You can run Open MPI on one stack and within seconds when that job is finished your next stack that absorbs that piece of the cluster can be run OpenMP or something completely different across InfiniBand fabric with accelerators in a containerized fashion.”

“I would also argue we are the price-performance leader in HPC and deep learning and cognitive workloads. We are going to be publishing some benchmarking studies with the world too, probably second half of the year. We want it to be a benchmark that is easily replicable by third parties.”

Being drawn into a technology arms race with large cloud providers would be counterproductive, Hebert recognizes, but growth is necessary. “In the early days we were a sand box because we were tiny but our strategy all along was build a global supercomputing platform that scales to infinity, just like the other guys. We’ve gone from handfuls of sand box machines to datacenters full of machines and serve users in 68 countries. Not only are the larger customers growing their spend with us, but the number of new logos that we are winning is increasing,” according to Hebert.

The point, he said, is to balance technology refreshes because you have to monetize your prior generation assets. “For example, we have made investments in P100 (GPUs). We didn’t make huge investments there. Thank goodness. Because that would have been a bad decision. The other cloud providers, the same thing. Microsoft and Google came in a little late here with the K80s,” said Hebert. Indeed, Nvidia’s launch of the Volta V100 just one year after the P100, and then positioning the V100 as a ‘do everything’ high-end GPU has the potential to quickly devalue P100-based systems.

AI is New Opportunity

AI writ large perhaps presents an opportunity for Nimbix to make AI a point of emphasis given its dependence on accelerated computing. AI is a tiny market at the moment – market watcher Gartner suggests only about four percent of enterprises have deployed AI in a meaningful way. The combination of new technology and tools needed to deliver AI combined with Nimbix’s ability to mitigate some of the complexity-to-use obstacles through its container-native infrastructure gives Nimbix an edge according to Hebert.

Collaborating closely with IBM could be a critical wedge. “We were the first in the world to offer a container built service for Power. In other words if you wanted to build a container for IBM Power, you can actually do that as a service on our platform. This was last year. That’s been a hugely impactful thing for the ecosystem because it allows ISVs to start to make your applications portable between architectures,” said Hebert.

“IBM is running a big portion of its cognitive training classes and data sciences experiences on our cloud, globally. They are rolling these out to thousands and thousands of data scientists around the world. There are enterprises but I would argue they are not in the same state. It’s not like they have these well-vetted AI pipelines. They are all just beginning to invest.” For Nimbix, the IBM training contract is a significant mutually-beneficial piece of early business on Power platforms, and dealing with AI issues.

Cloud providers sometimes asked to distinguish between development and so-called production AI projects they support with growth in production projects seen as evidence of ‘real’ AI traction versus development efforts that may or may not lead to volume activity. Yet in practical terms most enterprises are far from ready to deploy AI.

“It’s sort of a semantic term because most ‘production’ is still in a zone of innovation not mainstream, what I call a full AI pipeline. So if you were to use one of the most mature cases, autonomous vehicles, you’ve got huge clusters of GPU machines training neural networks that are then being pushed into inference engines in the car, and then those cars came in after running, they do a data dump, they filter data, feed it up, tune the model, this is a full pipeline. Not many are doing this.

“We talked about this with IBM’s DDL – distributed deep learning with Tensorflow – in which DDL is scaling linearly with GPUs. Enterprises are nowhere near ready to consume this technology. Maybe Uber and Ford and some of the big social media guys are [but not many others].”

Building the AI Workforce

One hurdle is the lack of data scientists. “I was in Houston with the oil and gas guys and everyone is investing in these data science teams and what they are doing is transforming their analytics guys, the traditional statistical analytics, and having to morph those guys and collide with deep learning type data science. That takes a long time. Furthermore, they are not even sure of the petabytes of data that they have, how they’d begin to filter data to feed into a neural network to spit out something that will help run their reservoirs more efficiently. We are seeing these issues in every major enterprise,” said Hebert.

What’s more, said Hebert, for every AI use case there is a whole collection of startups that are innovating on the AI software side. They are working to call on the big oil companies or retailers or wherever their expertise is: “Our message to those guys is ease of deployment with our container native platform using push to compute. You have your app and it propagates through our whole cloud and you are going to have better performance and costs than the other. That said most of these guys are going multi-cloud. They want choice. We’re okay with that. I think you’re a going to see winners emerge in each of the vertical use cases, but in our case we want to be that platform layer that helps be part of that processing engine our customers use.”

Traditional HPC, of course, remains an important part of Nimbix business, but that world is changing said Hebert, “There’s a couple of camps, the purists that want to control everything down to the silicon and if you are going for exaxcale and all that stuff. That world exists but it is an increasingly small user base whereas the enterprise guys just have to get stuff done.”

Hebert contends Nimbix technology has broader applicability. “We have some interesting patents which we think position us pretty well. We’ve already starting working with licensing our software technology to enterprises that want the stack, that vertical integration, as an option for on-prem HPC and AI that can also seamlessly burst to cloud. So that’s a key sort strategic evolutions of where we are taking are technology.”

It’s a grow or die world. Hebert said, “We measure. We want to know what the market growth rates are because if we are not gaining share as a smaller player, that’s not a good problem I want to face.”

Is acquisition a strategy? Hebert doesn’t dismiss the idea. “I mean it would be speculative. I think we have incredibly innovative technology and I think the big guys should certainly take a look at that. We feel like we have invested a lot to help users with this capability that now is in exciting times. We were probably started five years early but it is a good thing we did. I don’t think you could enter the market easily today; you would have to raise a lot [of capital.]”