Which Applications Benefit?
Many application workloads require the use of large datasets, and need rapid access to those datasets. The types of workloads which benefit the most from our disaggregated memory technology are those that:
- Large amounts of memory, often larger than is economical to deploy in each server within the compute system
- Fast access to that memory, such that it can be treated like local RAM rather than like network-attached storage
- Are operating on systems where it is challenging to deploy very large amounts of RAM, such as on most servers
Enabling New Applications
We believe that the disaggregated memory technology that TORmem is creating will be vital to enabling new classes of technologies which build on the use cases beginning to become commonplace today, which will be used by almost every business in the years ahead. These use cases must be able to operate cost-effectively and at high performance. Today’s limited local memory approach does not provide the level of flexibility that will be needed.
The amount of data being generated continues to grow, with 463 exabytes of data predicted to be generated per day by 2025. An exabyte is equal to one quintillion (1,000,000,000,000,000,000) bytes, and much of this data will be processed for artificial intelligence, machine learning and big data analytics by compute systems operating in the data center. With this sheer volume of new data combined with the desire to perform more detailed processing on that data over time, it is clear that performance and cost enhancements can be achieved by disaggregating memory from the rest of the server in compute systems deployed in the data center ranging from single rack systems all the way up to hyperscale deployments by cloud operators.
Artificial Intelligence and Machine Learning
Both model training and inferencing can require the use of very large datasets, and each benefit from the ability to process that data faster. Even the fastest AI accelerator can only process data which it has fetched from memory; where memory is either too small or too slow, very expensive accelerator resources can be left idle, decreasing performance and increasing the cost to perform each task. Whether CPU, GPU, TPU or another accelerator; they can all benefit from a fast and disaggregated memory bank in the rack which performs just like their own local system memory.
Big Data Analytics
Much of the data generated across the world today is analyzed not just by itself, but in the context of other data in order for businesses to extract more valuable insights from it than has been possible in the past. Sentiment analysis of social media activity, transaction analysis for credit card processing and real-time financial market analysis are all examples of valuable use cases which stand to be effectively accelerated using our technology. With TORmem, the very large datasets used for these applications can be kept in fast, disaggregated memory, making large-scale in memory databases a reality for a wide range of users, from global banks and large enterprises to smaller deployments.
Although it may sound a little hyperbolic, in many cases the primary differentiator between a supercomputer and the large server clusters of an enterprise or a hyperscale cloud is in their memory architecture. By bringing high-speed, economical banks of disaggregated memory to the data center, TORmem is closing the gap between today’s data center compute systems and the world of supercomputing. Tasks such as weather simulation, structural analysis and climate modelling can become more possible and economical with our memory technology at scale.
Accelerating Current Workloads
Many current workloads can also be accelerated by the addition of disaggregated memory. Whether in new or existing compute systems, it can be challenging to balance the need for more memory to keep CPUs, GPUs and other compute infrastructure fed with data, with the capacity of servers and the inflexibility and cost that comes with loading up individual servers with large quantities of RAM. TORmem’s approach is to add a bank of high-speed disaggregated memory to the rack, shared between multiple servers, which can greatly alleviate this problem and speed up current workloads.
Why Disaggregated Memory?
One memory for all. A single disaggregated bank of high-speed memory for the whole rack.
Our technology enables In-Memory Computing (IMC) at data center scale.
Optimize cost, accelerate current workloads, and enable new ones with us.