Intel, in collaboration with Dell and the University of Cambridge, has divulged details on the first phase of the Dawn supercomputer at the SC23 conference in Denver, USA. Constructed to drastically augment the UK's AI computational capacities, the Dawn supercomputer stands out with its significant infrastructure powered by Intel's components. Executives at Intel have announced the system will incorporate 512 of the latest 4th-generation Xeon Scalable processors and a robust array of 1,024 Datacenter GPU Max accelerators, known as Ponte Vecchio GPUs.
Each of the 256 liquid-cooled Dell PowerEdge XE9640 servers houses an impressive node with 1TB of DDR5 memory and 512GB of high bandwidth memory. Furthermore, the nodes are interconnected with Nvidia's top-tier Infiniband HDR200 interconnects, ensuring high-speed data transfer between components — a critical feature for complex AI computations.
Comparative Performance Benchmarks
Upon completion of its initial phase, Dawn has achieved a real-world performance of 19 petaFLOPS of FP64 (double-precision floating-point format) computation, placing it in the 41st slot amongst the world's foremost supercomputers. Yet, according to Intel, the peak performance of the machine could reach a staggering 53 petaFLOPS. This projection is based on claims that each Ponte Vecchio GPU can hit about 52 teraFLOPS at FP64, although there have been discrepancies in performance deliveries in the past.
In terms of AI capabilities, the Dawn supercomputer currently excels with its potential for lower-precision operations that are often employed in AI processes. Intel's data sheet indicates that the GPU Max 1550's can reach 832 teraFLOPS of Brain Float 16 (BF16) arithmetic, a benchmark particularly suitable for AI-related tasks. When expanded under the second phase to an anticipated tenfold increase, Dawn could carry out between 8.5 exaFLOPS of BF16 performance and as much as 17 exaOPS of INT8 operations, cementing its status as the UK's premier AI supercomputer.
Future Prospects and Global Standing
Should Intel and Dell successfully navigate and escalate the Dawn project to its envisioned magnitude, the UK will command a supercomputer of considerable power. The theoretical peak output after the second phase suggests a device capable of roughly 532 petaFLOPS at double precision, capable of rivaling and potentially surpassing current benchmarks set by leading systems such as Japan's Fugaku.
The full-scale Dawn system, targeting a ten times multiplicative increase in power from the first phase, demonstrates Intel's progress with Ponte Vecchio GPUs performance. Debate persists over whether the second phase will validate Intel's claims concerning the UK's AI supercomputer leadership. However, Intel's ongoing development of the Dawn supercomputer remains a testament to the continued advancements in the field of high-performance computing, positioning the UK at the frontier of AI innovation.
