AM2431: OSPI/QSPI Throughput of Writing Non-volatile memory

Hi Hideaki,

We have an app note that contains the benchmark information for the device: SitaraAM64x /AM243x BenchmarksCortex-R5 Memory Access Latency (Rev. B) (ti.com)

We only benchmark the read latency though, I am not sure if throughput is something we normally pec in this kind of document:

I'll reach out to more experts to comment on this

Best,

Daniel

Hi,

I'm not TI, but we've looked into something similar.

It really depends on the technology of their non-volatile memory.

The QSPI/OSPI interface itself is easily capable of this (1 MB in 100ms <=> 10 MB/s only requires ~10 MHz @ 8 bit or ~20 MHz @ 4 bit), but NOR flash (as found on the TMDS64EVM) is way too slow for that:

The S28HS512T from the EVM takes typically 570-680 µs (depending on sector size) to write a 512 byte page. 1 MB is 2048 x 512 byte pages, so 1.15s - 1.4s. Worst case values are 3x that. This is plus overhead for commands, addressing and data transfer, but that is usually negligible, because the interface is so much faster than the write speed.

They'd have to use other memory technology like FRAM or MRAM. These can typically sustain writing at the full QSPI/OSPI interface speed of 100-200 MHz x 4-8x bit [x2 for DDR]. Of course that comes at higher cost and smaller capacity.

We've successfully done something similar with an AM64x running Linux and FRAM. The driver needed some enhancements, because it is tailored to be able to read fast, not write fast, but that wasn't much of a problem. Not sure about the current state of the MCU+ SDK driver.

The AM24x hardware along with a suitable memory device can achieve that kind of throughput.

Regards,

Dominic

Hi Dominic,

Dominic Rath said:

The S28HS512T from the EVM takes typically 570-680 µs (depending on sector size) to write a 512 byte page. 1 MB is 2048 x 512 byte pages, so 1.15s - 1.4s. Worst case values are 3x that. This is plus overhead for commands, addressing and data transfer, but that is usually negligible, because the interface is so much faster than the write speed.

Is this data in context of TMDS64EVM or AM243x.

Hideaki Matsumoto said:

In their system, the write operation of 1MB to the memory has to be completed within 100 ms after triggered by interrupt.

Do you want us to check this possibility based of AM243x or TMDS64EVM ?

Regards,

Vaibhav

Hello Vaibhav,

sorry for the confusion. This is not my request, I was just trying to help with our experience based on the AM64x.

Hideaki asked for write throughput benchmark data on the AM2431.

Regards,

Dominic

Hi Hideaki,

I believe you have gone through: Benchmarks

If the above link is not sufficient enough, should I measure the throughput based off write operations in AM243x ?

Or should I go ahead with AM64x for calculations.

Regards,

Vaibhav