- AMD Megapod Design emphasizes the number of gross GPUs and interconnection efficiency
- Nvidia’s packaging method complicates direct comparisons with AMD approach
- Three racks will form the backbone of the megapod of AMD, and each calculation tray can combine verano processors with mi500 accelerators
AMD seems to prepare a new large -scale computer system for high performance computer science.
According to the analysis of the industry, the so-called “Megapod” should house a large number of accelerators, positioning it as a direct response to the established “superpod” of Nvidia.
The first signs of its conception indicate an emphasis on the raw GPU number and the effectiveness of interconnection, although the exact details remain speculative.
A system with 256 MI500 fleas
The reports suggest that AMD will organize the megapod on three separate racks.
The two external racks should contain 32 calculation trays each, while the central rack will contain 18 trays for networking switches.
In each calculation tray, a single Verano processor can be associated with four GPU MI500 instincts.
This provision would provide 32 CPUs and 128 GPU per rack, producing a total of 64 CPU and 256 GPU for the entire system.
The proposed design, sometimes called “UAL256”, reflects the intention to expand capacity so as to overcome competitors.
The comparison between the megapod of AMD and the Nvidia superpod is not simple.
On paper, the AMD system would offer 256 physical or logical GPU packages, while Kyber VR300 NVL576 from Nvidia only lists 144.
However, Nvidia’s architecture organizes four GPUs per package, which gives 576 GPU overall.
AMD’s dependence on single GPU packages means that its structure may seem more important in the number of physical packages, but not necessarily in the total density of the nucleus.
This complicates any attempt to declare a system as superior. For the moment, the Megapod seems more positioned as a counterweight than a leap forward.
The central rack of the projected system is designed to change infrastructure.
Analysts expect the new Vulcano network cards from AMD, derived from the Pensando line, are deployed here.
The use of these cards could determine if the megapod can provide enough bandwidth to fully use its gpu-lourse configuration.
Although the physical arrangement of racks and platforms seems simple, the latency and the rate of the network will play a role as critical as the gross GPU numbers.
The system is expected towards the end of 2027, by placing it on the same development horizon as other major HPC installations and large -scale data centers, including the German Herder SuperCalculator.
AMD has recognized the plans to combine the Verano processors, the MI500 accelerators and the Pensando Vulcano network cards, but the exact details of the Megapod remain unconcormed.
Via Computerbase (originally in German)
