RetroBSD

These cards are known to work, @ following SPI clock speed:
----------------------------------------------------------------------
1) NCP SD 256Mb - type 1, SPI 13.3MHz
2) Patriot SD 2Gb - type 2, SPI 13.3MHz
3) Wintec micro SD 2Gb - type 2, SPI 13.3MHZ
4) Transcend SDHC 4Gb - type 3, SPI 13.3MHz
5) Verbatim SD 2Gb - type 2, SPI 13.3MHz
6) Kingston micro SD 512Mb - class 1, SPI 20MHz
7) Kingston micro SDHC 4Gb - class 4, SPI 20MHz
8) Apacer micro SDHC 4Gb - class 4, SPI 20MHz
9) Sandisk micro SDHC 4Gb - class 4, SPI 20MHz

_________________
Pukao Hats Cleaning Services Ltd.

10) Kingston micro SD 512Mb - class 1, SPI 24MHz

_________________
Pukao Hats Cleaning Services Ltd.

I measured a write speed of several different SD cards, on kernel r269 and SPI 13.3 MHZ, using command:

# time dd bs=4k if=/dev/zero of=zzz count=1024

Results:
1) NCP SD 256 Mb - 440 kbytes/sec
2) Wintec microSD 2 Gb - 260 kbytes/sec
3) Sandisk SDHC 4 Gb - 198 kbytes/sec
4) Transcend SDHC 4Gb - 180 kbytes/sec
5) Patriot SD 2Gb - 145 kbytes/sec
6) Kingston microSDHC 4Gb - 55 kbytes/sec

_________________
--Serge

Interestingly, my Kingston microSDHC 4Gb (@13.3MHz spi, r269) shows:
# time dd bs=4k count=1024 if=/dev/zero of=zzz
20.1 real 0.0 user 19.7 sys
So when :
write_speed= (bs*count)/sys= 208kB/sec
p.

_________________
Pukao Hats Cleaning Services Ltd.

I changed sd driver to use write-multiple operations, and tested on several types of cards. Not committed yet - first need to fix virtualmips to support new SD commands.

_________________
--Serge

r287 (new sd driver) - Kingston uSDHC 512MB, SPI @20MHz:
# time dd bs=4k count=1024 if=/dev/zero of=yyy
15.6 real 0.0 user 15.1 sys
# time dd bs=8k if=yyy of=/dev/null
512+0 records in
512+0 records out
4.3 real 0.0 user 3.8 sys
#
this leads to write 271kB/sec (+31) and read 1078kB/sec (+398). Cool!

_________________
Pukao Hats Cleaning Services Ltd.

Hi Pito,
Try r288: it should be twice faster on write.
---Serge

_________________
--Serge

Hi Serge, r288, Kingston uSDHC 512MB, SPI @20MHz:
# time dd bs=1k count=4096 if=/dev/zero of=zzz
4096+0 records in
4096+0 records out
13.8 real 0.0 user 13.1 sys
# time dd bs=8k if=zzz of=/dev/null
512+0 records in
512+0 records out
4.6 real 0.0 user 3.8 sys
# time dd bs=4k count=1024 if=/dev/zero of=zzz
1024+0 records in
1024+0 records out
14.0 real 0.0 user 13.1 sys
# time dd bs=8k count=512 if=/dev/zero of=yyy
11.5 real 0.0 user 10.5 sys
# time dd bs=16k count=256 if=/dev/zero of=xxx
10.8 real 0.0 user 10.4 sys
# time dd bs=16k if=xxx of=/dev/null
256+0 records in
256+0 records out
4.6 real 0.0 user 3.7 sys
#

Max Write 394kB/sec, Max Read 1107kB/sec! Great improvement, thanks!
p.

_________________
Pukao Hats Cleaning Services Ltd.

New data, using 32-bit SPI mode (thanks Pito!).

_________________
--Serge

Hi Serge, I thought the 32bit spi sector xfer will have bigger impact, however there is an improvement seen:
Kingston, uSD, 512MB, SPI@20MHz
# time dd bs=4k count=1024 if=/dev/zero of=yyy
10.1 real 0.0 user 9.7 sys
# time dd bs=4k if=yyy of=/dev/null
1024+0 records in
1024+0 records out
3.5 real 0.0 user 3.0 sys
#
This leads to max write 422kB/sec and max read 1365kB/sec.

_________________
Pukao Hats Cleaning Services Ltd.

Hi Pito,
Try r291 - with read-multiple ops, reading speed improved by 50%!
--Serge

_________________
--Serge

Hi Serge, this is with r293 and with simple spi enhanced buffering, Kingston uSD 512MB @20MHz SPI, mind 8MB zzz:
# time dd bs=4k count=2048 if=/dev/zero of=zzz
19.8 real 0.0 user 18.7 sys
# time dd bs=8k if=zzz of=/dev/null
1024+0 records in
1024+0 records out
5.6 real 0.0 user 4.9 sys
#
Read 1672kB/sec, Write 438kB/sec.

_________________
Pukao Hats Cleaning Services Ltd.

Cool! What is the CPU frequency?

_________________
--Serge

default 80/80

_________________
Pukao Hats Cleaning Services Ltd.

Great: your SD card is very fast on read. Does SPI fifo really help?

_________________
--Serge

Yes, this single card is quite fast.. All others much slower.. The FIFO buffering helps a little (see my post in Drivers). Maybe the buffering can be pipelined somehow - so we do tx while we reading from buffer. This would require more complexity in the algorithm however.

_________________
Pukao Hats Cleaning Services Ltd.