Industry executives and experts share their predictions for 2021. Read them in this 13th annual VMblog.com series exclusive.
AI Will Accelerate Its Move to the Cloud
By Yaniv Romem, CEO, Excelero
In this year like no
other, the more things change - the more they stay the same. In enterprise IT, the
bar keeps moving ever higher - and despite an incredible amount of innovation, organizations
still need greater performance, efficiency and ROI. Here are our thoughts on the
industry's directions.
1. AI, ML and DL will continue to transform the
way data centers are architected.
The fast-growing use of GPU's, and with that the move to high-performance
networking, is a given. AI, ML and DL workloads
also have raised the bar
significantly for all aspects of storage performance. The market will continue
to address often unmet demands for IOPs, bandwidth and latency in 2021.
2. Deep learning, as distinct from AI, will
continue to command innovation.
Deep learning requires processing of very large amounts of annotated data to
get to useful accuracy close to 100% - since your Alexa smart assistant is not
very useful if it only recognizes 8 out of 10 words (80% accuracy) and your
self-driving car is not very useful if it recognizes only 8 out of 10 speed
limit signs. The problem often is the inability to load this data from the
storage system into the GPUs at reasonable speed, and for DL, the definition of
reasonable speed is significantly faster than traditional storage systems. Expect to see continued innovation in this
space.
3. AI
/ HPC, data analytics and database workloads will continue to move to the
cloud at an accelerated pace. The super-efficient, low latency, high IOPs, high
performance storage required by AI, ML and DL are difficult to achieve with
present cloud technology. Without them, applications perform less than
acceptably, or costly GPUs are underutilized and thus infrastructure efficiency
lags - just as is the case with on premise infrastructure.
Enterprises moving such workloads to the cloud will continue to drive new
hardware and service offerings. Some will move to cloud-specific offerings, but
many will build best of breed solutions by combining readily
available services, such as elastic storage solutions that are
cloud-oriented in nature, easy to consume with consistent price/performance
benefits and give them flexibility of choice and avoid constraining them to a
single cloud service provider.
4. Chip innovations that propel price/performance
will make NVMe an option for more
applications. Intel's recent is debut of 2nd generation Optane is
just one example.
5. The price/performance ratio of storage systems
will receive greater focus. While many
storage solutions support performance scale-out, when deployed to achieve an
AI/ML/DL-compatible level of IOPs, the cost of the storage often exceeds that
of the GPU servers. From the customer's perspective, the major part of the
available budget should of course be spent on the compute power, i.e. the GPU
servers, and the storage system should consume only a fraction of the budget that
gets spent on the GPU servers. Watch
this space in 2021.
##
About
the Author
Yaniv
Romem is CEO and co-founder of Excelero Storage, market leader in distributed block storage software
whose Elastic NVMe software powers AI training and analytics workloads at any
size and performance scale. Yaniv has been a technology evangelist at
disruptive startups for the better part of 20 years, and a founder at several
including Xeround and Picatel. His
passions are in the domains of high performance distributed computing, storage,
databases and networking.