For the first time in the history of computing, an accelerator-based architecture is powering the world’s fastest machine. That machine is the Tianhe-1A, a bespoke machine (custom one off) that is not commercially available (Tianhe translates literally into “sky river” or the Milky Way.)
SGI today introduced SGI Prism™ XL, a purpose-built insanely-scalable accelerator platform, based on standard components and open software, leveraging a new architecture named SGI® STIX™.
The difference between Prism XL and Tianhe-1 is that Prism XL can scale to 20 times that of Tianhe-1 and is a commercially available product, built on standard technologies, using open software.
Up to 20X the scale. Open. Standards Based. Commercially Available.
It is simple arithmetic. A 125 cabinet deployment of Prism XL could potentially replace the entire TOP 100 fastest computers in the world.
The Double Precision scale for Prism XL looks like this:
| 1 Cabinet | .2 PFLOPs |
| 5 Cabinets | 1 PFLOPs |
| 50 Cabinets | 10 PFLOPs |
| 100 Cabinets | 20 PFLOPs |
| 200 Cabinets | 40 PFLOPs |
Today’s announcement also completes the SGI architectures:
- Scale-out: with Rackable and Altix ICE
- Scale-up: with Altix UV
- Accelerators: with SGI Prism XL
This is a significant step towards ExaScale and moves the pendulum for technical computing.
SGI Prism XL is an open, standards-based accelerator platform for large-scale computation, supporting single and double-wide GPU cards from nVidia and AMD/ATI, as well as Tilera Tile-based processing cards – both leveraging the full bandwidth of PCIe interconnect and on-board dual QDR Infiniband or GigE (10Ge or 40Ge) for cluster integration. Each STIX can also have local direct attached storage (SATA and SSD).
Here is a picture of a STIX:
SGI Prism XL is a long-life platform and will easily adapt as technology advances. As more PCIe accelerator-based technologies come to market, plug them in. As PCIe Gen 3 comes to market, PRISM XL will support it. As GPGPU’s gain smarter interconnects, we will support that. As probability processors come to market, plug them into Prism XL.
Another great example is that Prism XL will support the Knights Ferry software development platform for Intel® Many Integrated Core Architecture.
Furthermore, given our cluster interconnect of GigE and Infiniband, many topologies will be possible: from fat-trees to tori to hypercubes.
Also, our new architecture and packaging were designed with extreme density and scale in mind. To achieve density we needed to go “beyond the blade” – the basic building block is not a chassis, not a blade, not a tray – but a “stick” – or SGI STIX whose dimensions are 5 x 3 x 37 inches. For example, an x86 PFLOP takes roughly 100 cabinets today, but with our new architecture, it takes only 5 cabinets and reduces the facilities footprint from a city block down to a small car.
As you need to scale, simply add a STIX at a time.
We see strong advantages for GPU accelerators in oil & gas for seismic analysis, image and signal processing within defense & intelligence, rendering and image processing within media & entertainment, as well as bio-informatics and physics.
I would like to highlight that accelerators are not just GPUs. For example, Tilera is NOT a GPU; it is a tile-based processor (think fast integer arithmetic). Tiles are low power and ideal for encrypt/decrypt, image and signal processing, deep packet inspection, format conversion and certain cloud-based content delivery. SGI Prism XL supports Tilera. We also support Violin and Fusion IO. Prism XL also supports high-speed I/O cards for high-speed data ingest and data acquisition. Below is a picture of what a fully populated cabinet would look like:
Equally important to the hardware is the software environment.Prism XL supports the SGI run-time and programming tools for massively parallel-environments that we support on our scale-out and scale-up lines. This includes standard LINUX, CUDA and OpenCL compiler tools, SGI Management Suite for provisioning to production, as well as job scheduling packages.
SGI Prism XL general availability is scheduled for December 2010.
Our customers now have 3 choices of technical computing architectures:
| Scale-out: | Rackable and Altix ICE |
| Scale-up: | Altix UV |
| Accelerators: | Prism XL |
This is the Prism and the Pendulum.
Comments