Hewlett Packard Enterprise (HPE) for the second straight week introduced new server products with today’s launches aimed at high performance computing. The centerpiece is a GPU-accelerated deep learning platform (Apollo 6500) whose announcement coincides with start of NVIDIA’s GPU Technology Conference (GTC) being held this week in San Jose. HPE also announced a Lustre storage solution (Apollo 4520) and three ‘purpose-built’ servers for financial services (FS) based on Apollo and Moonshot products.
There’s little question that deep learning and its various siblings (cognitive computing, machine learning, et al.) are increasingly important in HPC and the enterprise. The Apollo 6500 is a dense CPU/GPU machine in a relatively small footprint that can accommodate up to eight GPU cards in a single chassis. “We had a product a few years back, the SL6500, and this is the latest version of that. It will debut with the Intel Broadwell processor,” said Bill Mannel, vice president and GM of HPC, Big Data, and IoT at HPE.
Broadwell was formally launched just last week – it’s the Xeon E5-2600 processor v4 and Intel’s first processor to be offered within Intel’s Scalable System Framework. The new chip integrates processor, fabric, storage and memory technologies and was designed for use in two-socket motherboards, which is what the Apollo 6500 has. Broadwell is expected to accelerate many HPC codes, perhaps in the high 40-percent range. (For more on Broadwell, see ‘Broadwell’ Launch Brings Intel’s Advanced Scale Framework Into Focus, HPCwire)
The Apollo 6500 will run NVIDIA GPUs, AMD GPUs, and the Intel Xeon Phi, said Mannel. To help drive adoption of the 6500 (and deep learning more generally) HPE is providing three sets of tools intended to support a wide variety of users. The main three are:
- Cognitive Computing Toolkit. “If you are a scientist or engineer familiar with using MatLab or R then the cognitive computing toolkit will look very familiar and allow you to get up to speed very quickly on deep learning. The plan is to release this as an open source product,” said Mannel. Components are derived from HPE Software (including Vertica) and recent work in HP Labs, according to Mannel. The CCT is focused mainly on structured and semi-structured data and intended “to help ramp customers up in to memory centric computing and deep learning,” Mannel said.
- HPE IDOL (Version 11. 2). This is HPE’s platform for advanced enterprise search and data analytics for unstructured data with machine learning. HPE says it lets you search and analyze text, image, audio, and video from virtually any source. Mannel described the offering as well-suited for use with newer network technology being applied to a variety of deep learning type workloads.
- Haven OnDemand. Think of this as “deep learning as a software-as-a-service for different workloads,” said Mannel. There are on the order of 60 APIs for a wide variety tools such as for audio-video analytics, format conversion, graph analysis, image analysis, policy, prediction, and text analysis, to name a few.
HPE reports that when used with GPU computing platforms like the NVIDIA Tesla Accelerated Computing Platform, the 6500 “provides maximum GPU processing capacity” across a broad ecosystem of tools. The 6500 is designed to support deep learning computing platforms and application programming interface models such as Caffe, CUDA, Torch, Theano, Tensorflow, the NVIDIA Deep Learning SDK.
According to Mannel, demand for ‘deep learning’ is currently greatest in video and audio analysis applications. He noted there are many similar workloads in government intelligence. Demand is also growing in startup organizations (public and private) doing AI research and development that’s connected to deep learning, such as self-driving cars and drone aircraft. However narrow the market is today, the expectation is for the number of deep learning and related applications to grow significantly.
Availability of the 6500 is scheduled for this summer but pricing hasn’t been set said Mannel.
Pivoting to Storage Technology
It’s not surprising that HPE is ramping up its offerings here – HPC storage demand is expected to outpace HPC compute demand. The Lustre solution being announced is the Apollo 4520; it fits between 4510 (single motherboard) announced about this time last year and the 4530 (three motherboards.) Mannel said the latter is finding heavy use in object storage types of applications as well as in Hadoop environments.
Intel Enterprise Edition Lustre comes with the box. “The 4520 has two motherboards and will allows customer to have a failover capability and to help spread out the IO capability of the box,” he said. Currently GPFS (IBM Spectrum Scale) isn’t supported but is on the technology roadmap, according to Mannel. The new solution will be available April 18 with starting prices around $8500.
It is roughly a year since the HPE-Intel alliance around engineered solutions was formed, said Mannel, and the Apollo/Lustre solution is an early fruit.
“We have a good roadmap with engineering teams engaged. Basically we are creating a product that is a more flexible and open solution for a fat file cluster system for HPC users. So there’s a lot of appliances out there, and we sell them, but there are a class of users that want to be able to get into the platform itself, be able to configure for their particular environment and be able to tweak parameters, etc., and by combining IEEL with Apollo we believe we have an solution for those folks allows scaling in multiple directions (performance and size),” said Mannel.
Purpose-built HPC Solutions
Like all HPC technology suppliers, HPE is betting heavily on HPC’s wider adoption throughout the enterprise. A major goal of its Intel alliance is to identify and deliver engineered solutions targeting “specific verticals and specific customers in those verticals to create solutions that target a particular workload. The first segment is FS market because it’s a very large growing market,” said Mannel.
Take for example, the trader workstation world, he said. Currently it’s a sea of traders facing multiple screens supported by a hodgepodge of equipment. With Moonshot, he contends, “we can take all that video capability that is being delivered by mostly high-end VDI machines or workstations and PCs under desks and pull them into a tight package and distribute very good quality graphics across a networks.”
The other two FS offerings announced include:
- HPE Trade and Match Server. It’s an optimized Apollo 2000 server with a two-socket motherboard with a single socket populated. HPE says the processor can be over-clocked to achieve high trading performance and that some customers report being able to get 4Ghz. There are a series of processor options according to Mannel.
- HPE Risk Compliant Archive. It’s composed of iTernity iCAS software verified for archiving standards and Scality RING file and object storage on Apollo 4000 servers.
HPE also released an edition of Vertica for Hadoop-configured Apollo and Proliant platforms. The company will issue reference architectures for optimizing Vertica on SQL. Plans to support Spark are also on the technology roadmap.