3

I'm currently running with Samsung EVO Micro SD card in my phone (GT-i9195 or serranoltexx) in Adoptable storage mode.

I'm looking for a faster card to improve performance. I was about to get Sandisk Extreme A2 card for my phone.

However, I found the official pages about A1 and A2 class by SD Association. According to the official documentation,

A1:

The Application Performance Class 1 (A1) was defined by SD Physical 5.1 specification.

A2:

The Application Performance Class 2 (A2) is defined by SD Physical 6.0 specification. It makes SD memory card much higher performance than A1 performance by using functions of Command Queuing and Cache.

Note that A2 uses extra tricks to achieve better performance, especially Command Queuing (A2 spec'd cards have minimum support of 2 for Command Queuing depth).

I'm looking for a card that has as low latency as possible for a single read. Should I get A1 spec'd card instead of A2 because even though A2 has higher random access "minimum" IOPS, it may only get it with QD32 queues? I'm looking for a good random 4K read QD1.

I wish storage manufacturers specified random 4K read QD1 performance. It seems that only storage with manufacturer published numbers is Intel Optane series where those numbers don't look too bad.

Update: more information:

Another document on the official site explains A2 features a bit more:

Command Queue:

The new CQ mechanism allows the SD memory card to accept several commands in a series (without their associated data) and execute them (with the data) whenever the memory card is ready. It contributes mainly to random read performance.

Note that this clearly uses "random read performance" as average IOPS, not the maximum time for any given read operation.

Cache function:

In order to overcome the relatively limited write speed operation of flash memory, the Cache function allows the card to accumulate the data accepted by the host in a high-speed memory (e.g., RAM[)] [...] Data loss is allowed if the data is not moved to non-volatile storage.

So... high write IOPS is allowed to be writing to card internal RAM and data loss is acceptable? The OS can already do exactly the same thing with higher performance!

Self-Maintenance:

Card-initiated operation allows the SD memory card to perform internal maintenance operations while the host bus is idle. When a new command is received, the card must ensure that new command is serviced without violating SD standard timing (Read Command, 100 milliseconds; Write Command, 250 milliseconds).

And if card has entered "card initiated self-maintenance" mode (which A2 card is allowed to enter any time no commands are in the queue) the effective read IOPS may go down to 10 and effective write IOPS may go down to 4.

I think A2 spec has way too huge tolerances for bad performance. Calling the above as "minimum 4K read: 4000 IOPS" and "minimum 4K write: 2000 IOPS" is a lie.

(I'm wondering if I should get Sandisk Extreme A1 card when those are still available on the market. Samsung EVO+ cards have nice performance in benchmarks, too, but if I've understood correctly those are meant for non-application usage and have worse wear leveling algorithm than Sandisk Extreme cards. The same thing with Sandisk Ultra cards - they don't last having the same sectors written repeatedly.)

1 Answer 1

6

TL;DR: Get an A1 card unless you know for sure that your workload greatly benefits from Command Queuing (CQ) and your host hardware supports CQ.

Long answer:

After looking around a lot it seems that the A2 specification could allow better performance but actual hardware implementations are slow enough that A1 is better choice nowadays (year 2019).

Thomas Kaiser has collected info about A1 and A2 rated SD cards and says

Application Performance Class 1 (A1)

This performance class requires at least 1500/500 read/write IOPS (IO operations per second) with a 4k blocksize (small data chunks) and at least 10 MB/s sustained sequential write performance. No special host requirements are needed, the card simply has to exceed the performance requirements on its own.

and

Application Performance Class 2 (A2)

A2 promises even better performance with 4000/2000 read/write IOPS minimum but there's a problem since as outlined by the SD Association A2 cards show "much higher performance than A1 performance by using functions of Command Queuing and Cache".

Cache and Command Queuing require host (driver) support since the host needs to activate those new features first. The cache feature on A2 rated cards makes use of volatile RAM on the card [...] involving the risk of data losses [...]

Another user wrote about real world performance between Sandisk Extreme A1 and A2 cards:

I did some searching online and found a few other SD card benchmark results where A2 cards performed worse than A1 cards, so my results seemed to be what others were finding as well.

[...] A2 cards not only require special hardware and/or software for maximum performance, they actually perform WORSE than A1 cards without it

[...] [Sandisk fine print about A2 cards]: Up to 160MB/s read speeds, engineered with proprietary technology to reach speeds beyond UHS-I 104MB/s, requires compatible devices capable of reaching such speeds

[...] As far as I can find, the only "compatible device" is the SanDisk Mobilemate USB 3.0 micro SD card reader

James A. Chambers has tested A1 and A2 cards with Raspberry Pi:

A2 is not supported by Raspberry Pi yet (or almost anything else really) so A1 is the important mark to hit.

So the current status seems to be that A1 is strictly about card hardware speed and A2 is about card + host hardware/firmware/software solution. The only bit about A2 which in fact could improve performance is Command Queuing (CQ) which allows card to process some requests slower and still allow higher average troughput even for small I/O operations. And if you care about worst case latency (QD1) then any queuing is never worth it so you do not even want this feature.

However, it seems that with current hardware (year 2019) of A2 cards the latency for any single request has been increased from A1 cards. I guess this is caused by too slow embedded processors inside the cards that try to do a bit too much. The firmware is probably optimized for benchmark throughput (the card can handle lots of requests in paraller but execute each request a bit slower) and as a result, the latency for any single request has gone up. As a result, if your host device doesn't support CQ feature of A2 cards, you get the worst of two worlds: the latency will go up and overall throughput will suffer.

In the end, if your host does not support A2 feature, get an A1 card. And if you care about QD1 performance, get an A1 card even if your host hardware/firmware + software does support all the A2 features.

According to latest benchmarks I would recommend either SanDisk Extreme A1 cards or A-Data Premier A1 cards. According to StorageReview benchmark, the A-Data Premier A1 card has highest random 4K performance: 2971 IOPS for 4K read, 1190 IOPS for 4K write. The SanDisk Extreme A1 cards is better known, though, so it may be a safer bet. The SanDisk Extreme A1 is already hard to find because SanDisk prefers to sell A2 variants now especially for bigger flash sizes.

I wouldn't be surprised if current crop of A2 cards were cheaper to manufacture than A1 cards because the manufacturers can use lesser spec'd internal processor to achieve the A2 minimum random access performance thanks to more buffering and queuing. I can only wish more benchmarking sites would focus on request latency over throughput.

4
  • Status: I've been using A-Data Premium A1 MicroSD card (64 GB) for a couple of months and it has worked well. The performance has been about twice the performance of Samsung EVO+ 32 GB (for random 4K the A-Data has more than twice the performance). I haven't encountered any media errors this far so this seems like a good card. Considering the performance, I'd assume that this card is a good choice for Raspberry Pi, too, but I haven't tested it myself. Oct 1, 2019 at 9:35
  • The exact card I have is AUSDX64GUICL10A1-RA1 from adata.com/en/orderinfo/451 Oct 1, 2019 at 9:41
  • Update 2020: I have purchased two of AUSDX64GUICL10A1-RA1 and one of the cards failed. I don't know if I were just unlucky or if there're quality problems. The another card is working perfectly. Jun 6, 2020 at 7:42
  • Update year 2023: only one card AUSDX64GUICL10A1-RA1 has failed this far. The another one is still in daily use and has worked perfectly. Jun 22, 2023 at 8:47

Not the answer you're looking for? Browse other questions tagged .