The longterm goal for AMD is not to convince the retail market to use its APUs, but the lucrative server and HPC industry. Penguin Computing has revealed that it has started building the first HPC cluster based on an APU from AMD.
AMD launched the APU Llano with its four Husky cores earlier this summer, which is built upon the Phenom II series. Llano comes with an integrated memory controller, PCI Express and a GPU with 400 Radeon cores for gaming and parallel computing.
The A series, Llano, targets retail and there is no server version, but the goal for AMD is to use APUs for servers and HPC (High Performance Computing). Penguin Computing has built the world’s first HPC cluster with an APU from AMD.
“The unique AMD APU processor design combines multi-core x86 processing, memory controllers, a PCI-E interface and massively parallel GPU computing on a single piece of silicon.”
The cluster uses 104 servers (Altus 2A00) that each support 64 GB DDR3-1600 and DDR-1866 memory. Since Llano is intended for retail it lacks ECC, but that doesn’t have to be a deal-breaker, but the perhaps most important is that Llano lacks any kind of bus for communicating directly with other APUs, like AMD’s processors that have multiple HyperTransport buses. This has been solved by connecting the servers with 40 Gb/s QDR Infiniband.
Penguin Computing’s HPC cluster has a theoretic performance of 59.6 TeraFLOPS. This implies the server is built with 104 AMD A8-3850 that has 573 GigaFLOPS each. Penguin Computing says it is easier to make use of both CPU and GPU in APUs since they share the memory controller, which offers great advantages over other solutions with combined processor and graphics card.
The new cluster will be used to explore the use of OpenCL for heterogen programming between processor and graphics circuit in APUs. It also says that AMD’s APU architecture has potential to become the main ingredient in future “exascale” systems, with ExaFLOPS performance*. Exascale is something both AMD and Intel is striving for and hopes to reach, before or during 2020.
*1 ExaFLOPS = 1 000 PetaFLOPS = 1 000 000 TeraFLOPS = 1 000 000 000 GigaFLOPS
Penguin Computings Altus 2A00 servers will be available from today for curious customers. When AMD server APUs will become available we don’t know, but we do know that AMD has great ambitions for such a product. It looks like 2013 or 2014 will be the year it happens.
