A Few Words about the Dalek Project¶
Dalek is an innovative cluster built around CPUs traditionally used in mini-PCs or laptops and from GPUs that can be found in gaming PCs (or iGPUs). The cluster comes with a wide selection of recent components that can be tested on a variety of algorithms. One of the main purpose of Dalek is to provide such a diversity of components at a moderate price. Indeed, components from the general public are much less expensive than server-class hardware. As a consequence, Dalek feats very well to software design and prototyping as it enables researchers to discover and to experience new hardware just after their release date.
The processors included in Dalek are x86 CPUs from Intel and AMD vendors (ARM CPUs may be considered in a future update). Some partitions focus on providing homogeneous multi-core CPUs while others expose heterogeneous SoCs with performance, efficient and low power efficient cores (in the Intel designation: p-cores, e-cores and LP e-cores respectively). Most of this heterogeneous SoCs come with a NPU accelerator for efficient inference of DNNs (like CNNs and LLMs). Typically these SoCs cannot be found in traditional computing centers, hence, it is very interesting to have them in an environment dedicated to reproducible measurements. The concerned architectures are generally optimized to minimize energy consumption while maintaining the best possible performance. They can be forerunners of the upcoming architectures in future computing centers.
Concerning the accelerators, Dalek tries to cover the main architectures of the major vendors, providing Nvidia, AMD and Intel GPUs. Nowadays, the professional GPUs are available at very high price (mainly because of the strong pressure on the AI market). Using gaming GPUs is a cheaper alternative that enables algorithms prototyping including for the deep learning.
The partitions are typically composed of four nodes of the same type. This is very small compared to what can be found in supercomputers or on large clusters. Nevertheless, it enables to start to implement distributed applications. The network is based on 2.5 GbE interfaces: it does not match the standard clusters that are generally equipped with fiber channel (infiniband) but it enables to work with low price motherboards and mini-PCs.
To summarize, Dalek is a unique cluster that enables the design and the optimization of future compute intensive applications. The later will need to take into account many different layers of optimization. Among them, there are dedicated instructions on a single core, multi-core parallelism with heterogeneous cores that sometimes even not share the same ISA, and multiple nodes that can exchange data on a external network. Of course, the same complexity can be transposed to the memory.
This project has been carried out with financial support from the French State, managed by the French "Agence Innovation Défense".
