Modularity and Disaggregation : why and how it will change the game in the datacenter

Future servers, in the years to come, are expected to be different from today’s servers, when it comes to the architecture: they will make more efficient use of computing resources and memory by relying on a modular architecture through disaggregation. This important development will enable applications to benefit from this new approach. I covered this topic in my presentation (slides) during the workshop on "Trends in supercomputing architectures in the exascale era: heterogeneity, modularity, disaggregation" chaired by?Jean-Philippe Nominé?(ETP4HPC /?Commissariat a l'Energie Atomique et aux Energies Alternatives) and?Pascale Bernier-Bruna?(ETP4HPC /?Eviden) together with?Enguerrand Petit?(Intel Corporation) at TERATEC 2023 in Paris on Jun 01, 2023. Here are the key points to remember.

Why disaggregation and modularity in a server ?

Multiple technological advances in the design of microprocessors have enabled more modular chips, based on 'tiles' or 'chiplets'. This modularity will in turn enable new designs and more performance. Disaggregation will also increasingly be implemented at the node level as well as at the system level, particularly through fast networks and low latency protocols like CXL?. This will allow, for example, on-demand integration of various memory levels or of more efficient coherent discrete accelerators. Modularity and disaggregation will therefore increase the heterogeneity of systems. In addition, development of standard programming environments (such as oneAPI) becomes critical, that will allow users to have a transparent use of these systems, while maintaining the highest levels of performance.

How will it work ?

This new approach impacts hardware, software and connectivity, bringing improvements at every level. Let me explain how :

• Modularity at the chip level will be achieved through assembling chiplets (or tiles) of different types, different sources, with different IP, with different capabilities, to respond to the various requirements in the workloads. Applications make calls to scalar instructions, vector instructions, more and more matrix instructions (for AI inference and training, already available in the integrated accelerator enabling Advanced Matrix eXtensions in the Intel? 4th generation Xeon? Scalable Processor), and also spatial instructions (in FPGAs typically). The concept is to have different tiles in a package, each tile with a specific balance of the scalar, vector, matrix, spatial capabilities depending on the targeted workloads. Intel develops and supports standards for enabling interoperability in the ecosystem, and UCIe? (Unified Chiplet Interconnect express) is the new standard for connecting chiplets. UCIe is supported by the leaders in semiconductor, packaging, IP, as well as cloud service providers.
• At the next level of connectivity, Integrated Optical I/O is a key innovation answering the needs in growing data rates, energy efficiency, and channel loss minimization.
• At the system level, disaggregation is made possible with CXL (Compute eXpress Link) initially proposed by Intel, and that has become a de facto standard in the ecosystem through the CXL Consortium. CXL solves the key challenges of scaling, faster access to data, and more efficient interconnect, as it offers low latency, high bandwidth, and a coherent interface. CXL is already implemented in systems using 4th generation Intel? Xeon? Scalable Processors and Intel Agilex? 7 FPGA, enabling a cache and memory coherent discrete accelerated solution.
• And for programming these modular, heterogenous and more and more often disaggregated systems, Intel has contributed a standards-based solution, oneAPI, an open, multi-architecture, multi-vendor programming environment. OneAPI is an open industry specification, enabling freedom in hardware choice. It delivers performance, productivity, and portability. It is standards-based (SYCL) and community driven. What is new is that oneAPI enables to write a source code once, and run it across architectures and across vendors.

I would be delighted to discuss this further with you if you wish to. I’m interested in your feedback on the subject.

#IamIntel #intel #Forumteratec