Hard to imagine Google, Facebook building AI without (checks notes) Dell EMC's Data Science Provisioning Portal
If you want to do some ML, and you've got a fat budget, they've got some tech to sell you
Hoping to take advantage of the alleged any-minute-now boom in artificially intelligent software for enterprises and service providers, Dell EMC has announced a pair of "AI Ready Solutions."
The new kit – poised to power all that machine-learning code everyone and their dog has been talking up – consists of a pair of rack-scale converged system architectures: one with Cloudera and Hadoop nodes, the other with Nvidia GPU nodes and Isilon all-flash NAS.
Dell EMC said both are validated stacks of hardware and software. The Cloudera-Hadoop one is for standard machine learning and big-data analytics, while the Nvidia-Isilon one is for deep learning. The latter will compete against Pure’s AIRI product and the NetApp ONTAP AI system.
The two Dell EMC systems aim to provide a self-service for data scientists, with a selection of AI frameworks and libraries, and a single Dell EMC throat to choke for support. Both feature Dell servers, storage, and networking.
Colourful AI rack cabinet sides. Look what's in Dell EMCs AI rack garden.
Machine Learning Cloudera-Hadoop system
The acronym-tastic DERSAIML – Dell EMC Ready Solution for AI: Machine Learning – stack is formed thus:
- Cloudera Data Science Cluster: head node and two worker nodes with 960GB to 1.92TB direct-attached SSD storage.
- Hadoop nodes: starting with three infrastructure nodes and seven worker nodes, scaling out to thousands of nodes.
- 25GbitE Ethernet networking with Dell EMC Open Network Switches.
- Software stack: Cloudera Manager, Data Science Workbench, Enterprise Data Hub, Spark, and Dell EMC Data Science Provisioning Engine.
- Frameworks/libraries: BigDL.
Deep Learning Nvidia-Isilon system
The DERSAIDL – Dell EMC Ready Solution for AI: Deep Learning – with Nvidia and Isilon components in its stack encompasses:
- PowerEdge R740xd cluster management: a dual-processor head node with 12 x 10TB direct attached SAS drives.
- PowerEdge C4140 1U scale-out worker nodes with dual Intel Xeon Gold 6148 20-core processors, 348GB RAM and four Nvidia V100 Tesla GPUs.
- Dell EMC S3048-ON Ethernet switch to connect to an Isilon filer.
- Mellanox SB7800 InfiniBand switch to interconnect servers.
- Isilon F800 all-flash scale-out NAS with eight 40GbitE links.
- Software stack: Bright Cluster Manager for Data Science, and Dell EMC Data Science Provisioning Portal.
- Frameworks/libraries: Caffe 2, MXNET, TensorFLow, NVIDIA's CUDA Deep Neural Network library (cuDNN), and CUDA basic linear algebra subroutines (cuBLAS).
Compute and storage resources can be scaled independently, with storage added non-disruptively and compute nodes needing a few mouse clicks, we're told. There can be thousands of worker nodes, Dell EMC boasts. The F800 NAS can reach up to 250,000 IOPS with a 15GB/sec bandwidth, and 96 to 924TB capacity per chassis. It has eight 40GbitE networking ports per chassis, and scales out to 33PB, and up to 540GB/sec bandwidth per cluster.
We have partial – we emphasize the "partial" – vendor-supplied comparative performance data for this system versus Pure's AIRI and NetApp's A700/Nvidia and A800/Nvidia systems, using deep-learning neural networks Resnet-50 and AlexNet also in vendor-run benchmarks:
|4 GPU||8 GPU||16 GPU||32 GPU||4 GPU||8 GPU||16 GPU|
|NetApp A700 Nvidia||1131||2048||4870||4243||4929|
|NetApp A800 Nvidia||6000||11200||22500|
|Dell EMC Ready Solution - Deep Learning||2931||5590||11120||7691||14259||2735|
The higher the score, the better. Blank spaces indicate no available data. With Resnet-50, the Dell EMC kit beats the NetApp A700/Nvidia systemwide, and is slightly slower than NetApp's A800/Nvidia ONTAP AI offering. It also out-performs the Pure AIRI product.
This is your four-minute warning: Boffins train ImageNet-based AI classifier in just 240sREAD MORE
ESG, which ran the Resnet-50 and AlexNet benchmarks for Dell EMC, said AlexNet is an image classifier that can label objects in pictures from 1,000 categories, such as keyboard, mouse, pencil, and so on. The benchmark trains the AlexNet model using the ImageNet data set, a de-facto standard for deep learning training. The 143GB ImageNet data set contains 14,197,122 images from 21,841 distinct categories.
It is less computationally intensive than Resnet-50, which is also an image classifier.
With Alexnet, the Dell EMC system is faster than the NetApp A700/Nvidia system but there are no NetApp A800/Nvdia numbers for this benchmark. Based on the Resnet-50 results, we might assume it would be in the same area as the Dell EMC deep-learning offering.
Portal and services
The idea behind both AI Ready Solutions is to simplify AI system deployment and operation. Dell EMC said it has enabled "self-service" by data scientists, with no need for written scripts to be developed.
It has set up a Deep Learning Institute with Nvidia, which is based in Austin, Texas, where customers can learn how to build AI-accelerated applications with hands-on training. It also provides AI consulting services covering installation and deployment, data engineering, and data science best practices and processes.
With this pair of Ready Solutions, Dell EMC has entered the AI rack garden and intends to flourish there. The systems are said to be available now in the US, and will roll out in Brazil, Canada, Mexico, France, Germany, UK, Australia, China, India and Japan within the next two months. ®