JVET-J0090 AHG5: Measurement result of memory bandwidth comparison with JEM and HM [R. Hashimoto, S. Mochizuki (Renesas)] [late]
This contribution was presented Thu 19 April 1120-1210.
This contribution shows a comparison between the memory bandwidth in JEM7.1 and that in HM16.16. At the 9th JVET meeting in Gwangju, JVET-I0033 showed the measurement results of memory bandwidth in JEM7.1. Based on comments about that analysis, this contribution shows additional results. This method can be used to check memory bandwidth in the test model for the video standardization work and be evaluated in comparison with HEVC.
The comparison was based on CfP anchor bitstreams.
The tool is capable of analysing the external memory bandwidth and cache memory bandwidth of a decoder, and is configurable for different cache configurations.
The results indicate that the average external memory bandwidth of the JEM is 6.4× higher in RA, and 4.0× higher in LD when compared to the HM (for the configuration B cache size of 64 kbytes). One effect observed was a tendency for the bandwidth (both external and cache) to be becoming higher in the JEM towards lower bit rates, whereas the HM shows a tendency for the bandwidth to increase towards higher bit rates.
It would require more analysis, with switching tools on or off, to get information which tools are requiring which amount of memory bandwidth.
The tool is only capable of computing the average decoder memory bandwidth for a given set of test sequences. For analysing worst cases, special bitstreams would be necessary.
It was asked how much is the decoder slowed down when using the tool. This was approximately 2x; it would be necessary to run the decoder again to measure the memory bandwidth.
It was commented that still the precise amount of memory bandwidth would be dependent on a given hardware platform. Nevertheless, the information that can be provided by the tool is very interesting to identify possible memory bandwidth problems of tools (and of an overall algorithm).
Decision (SW): Implement the tool in the test model software. It would be desirable to use it in the context of CEs starting from the next meeting.
Further development of the model (e.g. with other cache block sizes) is to be discussed in an AHG.
It was also commented that the bandwidth consumption of a tool is likely different when executing a “tool on” test (with the simple VTM not using other tools in parallel) and a “tool off” test (in BMS configuration). This aspect needs further study.