草榴社区

How to Improve Throughput of AXI Bus

If a AXI Bus is being used for frequent bulk data transfer, achieving good throughput is very important. Throughput can be calculated by computing the sum of all data bytes (AxSIZE) in each beat (RVALID/BVALID) captured on the AXI interface during the observation window and dividing the sum by duration of observation window. A window showing low throughput generally doesn’t mean an issue, unless there was an expectation to move large amounts of data quickly. A few reasons for less throughput could be:

• Manager behaviour: Ideally the manager should assert AWVALID and WVALID on the same cycle. Also, the manager should be able be drive multiple beats by keeping WVALID high on consecutive cycles. If this is not the case, then the manager is restricting throughput for write transactions.
• Valid/ready handshaking: The best performance can be achieved if xREADY is always high on both the manager and subordinate side. However, real world DUTs must eventually de-assert xREADY when the internal pipeline is full. So, the manager/subordinate should ideally keep outstanding transactions within DUT pipelining limits to make sure there is no stalling.
• Request to response latency: The subordinate might take few cycles to respond to a write/read request. Peak performance is achieved when the response is on the next cycle to when the request was sampled by sub-ordinate. However, complex interconnect routing and memory access often takes a few cycles before response can be driven. (Figure 3)

How to Improve Transaction Performance of AXI Bus

Arm AMBA 3 AXI and Arm AMBA 4 AXI interconnects support outstanding transactions without any limitation, even allowing multiple outstanding transactions with the same ID. The ID (or few bits of it) is often used to route response from the subordinate to the correct manager with a unique ID. If the manager can issue multiple outstanding transactions, it should be done so only if the subordinate supports it as well, otherwise it will simply deassert xREADY signals and lead to a stall. Even if subordinate supports outstanding transactions, it can only do so for as long as its internal pipeline is not full. So, optimal performance can be obtained if the manager issues outstanding transactions equal to or less than the pipeline depth of the subordinate, which allows the interconnect to process multiple transactions without any serialization.

Smart Monitors for Arm AMBA AXI interface can allow a user to measure AXI Bus performance to optimise design for the desired performance before real silicon is taped out. For further debug into a window, AXI traffic needs to be analyzed to trace the transaction(s) responsible for a drop in performance. Finally, the design needs to be checked for possible reasons which lead to an observed deviation in the transaction.

How Smart Monitors for 草榴社区 ZeBu EP1 Help to Analyze AXI Bus Performance

Smart Monitors for Arm AMBA AXI interface are DPI based transactors, but they are passive components only used to capture bus traffic. A monitor can process protocol data either for functional verification or for performance analysis. For performance analysis, the monitor supports 3 modes (Figure 5) –

1. Python-bases batch visualizations
2. 草榴社区 Verdi? Performance Analyzer based performance visualization for verification engineers
3. 草榴社区 Platform Architect? virtual prototyping solution-based performance visualization for software engineers

Any one of these modes can be used as per requirement to analyze AXI bus performance.