STM32H5 Slower than expected

[F14V]

Describe the bug
I've been running the WhetStone Benchmark from the Examples repo on different boards to compare the performance of the MCUs but I got unexpected results. The following results have been obtained with the Generic variants of each MCU and no optimization (O0):

• H503 @ 250 MHz: ~38 MIPS
• G474 @ 150 MHz: ~40 MIPS
• F411 @ 100 MHz: ~30 MIPS

The H5 running at a higher frequency is slightly slower than the G4 running at a lower frequency.

To make sure I'm not doing something wrong, I've run the benchmark using the CubeIDE on all chips, with all 3 Cache configs for the H5 (disabled, 1way, 2way). The results for F4/G4 are similar with Arduino, together with H5 with disabled cache. When enabling the cache, the score is way better:

• H503 @ 250 MHz with 2 way Cache enabled: ~91 MIPS

This shows an improvement of almost 3x compared to the F4 @ 100MHz, as expected .

To Reproduce
Run the WhetStone benchmark from the Examples repo on H5 and F4/G4 MCUs and compare the results, taking into account the CPU frequency.

Expected behavior
I expected the H5 to outperform the G4/F4 only taking into account the frequency, let alone the M33 improvements over M4.

Desktop:

• OS: Windows
• Arduino IDE version: 2.3.2
• STM32 core version: main branch (d7019d1)
• Tools menu settings: O0, Newlib Nano + printf, USB CDC, UART disabled
• Upload method: DFU

Board:

• Name: Generic H503CBUx (on WeAct STM32H503CoreBoard)

Additional context
I know that support for H5 is still relatively new, but the family is supposed to be high performance so it would be a shame to have it underperform.
I have been looking through the core code and I have seen the Cache being enabled for Cortex-M33 MCUs, but there were also some cache defines missing from the default config header. I don't know what is happening, am I missing something?

[fpistm]

Hi @F14V
the core provides basic support of each family and be as much as possible Arduino compatible.
For several mcu we could certainly increase performance but could not fit with Arduino API.
About the benchmak, it was provided "as is" at a certain time so I cannot guarantee that the results are relevant.

If you think you can increase performance, feel free to contribute. Cache support is enabled by default but can bring some issue and could probably be improved (ex: MPU definintion) but hard to deploy/maintain due to the generic support.