As our world continues to be more connected, people and machines have an increasing hunger for content rich processing, higher speeds and shorter response times. Prodrive Technologies’ Nucleus HPC and datacenter architecture shows how the growing demand for energy efficient processing can be effectively addressed.
Nucleus is built around the core principles of locality, dependability and energy efficiency. It is unique – based on its gravitation towards energy efficiency, low latency, high density, processing heterogeneity, high level of scalability and best of all, application versatility.
Compute and switch functions are closely coupled and physically densely packed, similar to protons and neutrons on the atomic level. In the same vein, the flexible balance between the two enables a wide variety of applications.
The electrons that revolve around Nucleus’ core, are today’s and tomorrow’s high performance, large-scale applications. For instance, applications that are compute-, bandwidth- or latency-bound and do not fit on a single server.
Putting it all together, Nucleus can scale up from an efficient 1U server with integrated, low latency switching to multiple 42U racks. Applications include mobile edge computing, radar processing, data analytics, high frequency trading and more.
The hardware building blocks for Nucleus were designed especially for high reliability and low energy consumption. This allows for increased performance density compared to what is currently in the market. Processing options can be selected from the latest Intel Xeon Broadwell-DE/EP processors, Xilinx 7-series and UltraScale FPGAs and many more – all tightly interconnected and optimized for multiprocessing.
Low latency, high bandwidth and lossless RapidIO connectivity is used within the rack, allowing existing applications to instantly reap the associated benefits. Non-deterministic delays, cache misses and costly processor cycles for data transport are now part of the past – RapidIO is a lossless hardware terminated protocol supporting remote direct memory access (RDMA). Its strong and growing ecosystem, support for IP networking and roadmap to 50Gbps and 100Gbps bandwidth assure for rapid application deployment, provides future growth and protects your investment.
Multiple 10/40/100 Gigabit Ethernet connections allow for rack-to-rack or rack-to-spine connectivity, including bridging to the RapidIO endpoints inside the rack.
Nucleus programming model
Processing blocks are programmed using ISO C, C++ and Fortran. OpenCL extensions are available for offloading and acceleration.
Networking inside Nucleus is facilitated by Berkeley sockets over RapidIO. This enables existing sockets or MPI based applications to be quickly ported to and evaluated on Nucleus.
Prodrive Technologies’ Meshworks suite allows for hot spot identification and other network diagnostics.
Nucleus’ state-of-the-art hardware
- Up to 64 Intel® Xeon® cores / 128 threads per rack unit
- Up to 4 sockets per rack unit
- Based on Intel®’s 14nm “Broadwell” architecture
- Heterogeneous compute, including FPGA, DSP and others
- Ethernet outside and RapidIO inside the rack
- Multiple 10G / 40G / 100G Ethernet rack-to-rack
- 240Gbps low-latency RapidIO 2.1 switching per rack unit
- Roadmap to >1Tbps per rack unit with RapidIO 3.1
- Scalability from single rack unit up to multiple 42U racks
- Scalable from 300 watts per rack unit to 22 kilowatts within rack
- Integrated IPMI v2.0 based hardware monitoring
- Standards & compliancy
- NEBS, ETSI, SEMI
- RoHS, WEEE
Nucleus’ standards-based programming model
- ISO C/C++/Fortran, POSIX
- IP sockets over RapidIO, MPI
- Linux operating system
- OpenMP, OpenCL