SK-AM62: Memory Bandwidth may be too low

Can you explain a little more how you compiled and built the chromium browser on each platform?  The DDR4 configuration should be optimized for maximum frequency if you used the SDK, so i don't think the issue is with the memory.  Something must be unoptimized in the build that is restricting the browser performance.

Regards,

James

I knew that the web browser wasn't optimized enough, so I compared it in a software environment that I didn't build myself.

Well, I'd like to ask you the value of LMbench in the Perfomance Guide of the Processor SDK Linux Software Developer's Guide, which is what TI did.

I would like to use the benchmark results published in the Developer's Guide as the criteria for SoC selection, so please post the results in a properly optimized environment. Of course, I feel that the AM625x hasn't matured right after its release.

----

By the way, I don't know the details because the performance comparison of Chromium browser was done by the contractor. For the AM335x and AM572x environments, we used what was built into the pre-built boot SD environment of the processor SDK. In the AM625x environment, we matched the version of the AM572x Chromium browser for comparison with the AM572x. However, we changed to the 64-bit version. He tried to build the latest Chromium if he had to build it anyway, but the latest Chromium uses clang as the compiler, and there is a possibility that it does not fit the Arago Project environment very much.

AM335x​ Chromium 53.0.2785.143​ 32​bit
AM572x​ Chromium 75.0.3770.142​ 32​bit
AM62x​ Chromium 75.0.3770.142​ 64​bit

-----

The memory performance of the SK-AM62 is not good for the following reasons;

AM65xx EVM   DDR4-3200 8bit x 4chips in parallel -> 3200MT/s x 8bit x 4 = 102,400

SK-AM62         DDR4-3200 16bit x 1chip -> 3200MT/s x 16bit x 1 = 51,200

AM572x EVM   DDR3-1600 16bit x 2chips in parallel -> 1600MT/s x 16bit x 2 = 51,200

I suspected that even if the bandwidth was the same, the performance would not increase due to the overhead of bus switching. It has only 2GB of memory, so it may be better to use a 32-bit OS instead of a 64-bit OS. The time to prefetch 32bit instructions from DDR4 is half the time for 64bit instructions. 64-bit instructions are faster after being cached. Considering the impact of DDR4 effective bandwidth, caching is becoming more and more important. Even though it is a long-awaited 64-bit environment ... 

I will have to transfer your questions to our software experts to comment further.  

BTW, the AM572x operates DDR3 at a maximum of 1066MTs, not 1600MTs.  Also, AM62x and AM65x operate DDR4 at a maximum of 1600MTs, not 3200MTs

Regards,

James

I misread AM62x TRM. It was 1600Mbps instead of 1600MHz.

• Support of LPDDR4 memory type (up to 1600Mbps)

• Support of DDR4 memory type (up to 1600Mbps)

In other words, the upper limit is DDR4-1600, not DDR4-3200.

Even if DDR4-1600 is the same, AM65x has a bus width of 32 bits and AM62x has 16 bits, so it was natural that there was only half the memory bandwidth.

As for the performance of AM62x against AM65x, I understood that the bus width is half. So, If I think the AM62x has poor memory bandwidth, I can use the AM65x instead of the AM62x.

The approximate memory bandwidth may look like following:

AM65x 32bit DDR4-1600 : 1600x2=3200 : 4
AM57x 32bit DDR3-1066 : 1066x2=2133 : 2.66
AM62x 16bit DDR4-1600 : 1600x1=1600 : 2
AM335x 16bit DDR3-800 : 800x1= 800 : 1

So I think I'll make it Resolved, the high speed memory of AM572x is still a mystery, but the document link is broken, so now I can't confirm that the AM572x's values in the document were correct.

www.ti.com/.../PROCESSOR-SDK-AM57X