[UPDATES: Camnostic just published a review of CFexpress cards as they perform in Panasonic S1R and S1H bodies. The take-away: those cameras have throughput limitations that makes CFexpress card choice less important, the best providing only 1/7th better throughput than the XQD cards those cameras supported natively.
A similar performance test was applied to the newly-firmware-updated Nikon Z7 II, which can be found here. The firmware for these older-design Nikons needs to be updated before the XQD port can do double duty as a CFexpress card port. As it is, the Nikon’s new application of the standard provides only about 1/3rd additional speed over the old XQD method.
We also recently published a similar review of SD cards here.]
– If you want the best price per GB while still being in the top 5 of performance, you’ll have the Delkin Power 512 GB at $0.98 per GB. The 2 TB version of that card costs only $0.50 per GB, about half as much. Alternately, you could purchase the Transcend 512 GB, which clears the cache more slowly, but is even cheaper at the 512 GB size. It costs a mere $0.64 per GB at one current retail deal ($328), although it lists at $450 normally.
– If you want the most number of shots in 30 seconds, you’ll have the Angelbird 660GB XT. But it’s best by only 1 picture out of 337 [Note: it originally showed an odd behavior of causing the Canon R5 startup to take 4 additional seconds when turned on, but a firmware update, and several Canon firmware updates have eliminated this issue.]
– The card that clears the buffer the fastest is the SanDisk Extreme Pro 512GB, clearing it more than half a second faster than the other cards of its capacity.
– The best all-around card is arguably the ProGrade Cobalt, which doesn’t win outright in any category, but ties for first in several of them.
– The card that takes the most shots before the buffer kicks in while shooting in mechanical shutter is the Delkin – but by just a smidge.
– The cards that perform best as hard drives hooked up to the computer are the Sony cards. They’ll offload images and video faster than the others.
Biologists call a flock of CFexpress cards a “kidney” of cards
So let’s get to the measurements….
It has to be said that the CFexpress format is a leap forward. The cards are all so fast that the difference between the top performer in clearing a full buffer and the card that came in second from last is just 1.5 seconds. The buffer is that on-board, super-fast memory that fills up, and then you’re just left with the raw speed of your memory card, chewing through what the buffer can upload to it. That the memory cards can eat up the full buffer so quickly now has implications for people thinking about which factors are most important. If clearing the buffer is the difference between 4.5 and 6 seconds, it might not be as big a priority as it use to be. In the old days (2019), the difference between cache clearing in one card to another could easily have been more than 15 seconds. When shooting SD cards on the Sony A9, it seemed you could do your laundry and have a little time left for a snack while the buffer unloaded to the memory card. With CFexpress, buffer clearing times are just less important. All that said, the SanDisk 512GB card is tops at clearing the cache, as the chart to the right shows. (The Lexar card was eliminated from consideration because, while speedy, it failed. More on that later.) This is the time it takes to completely move the contents of the cache to the card. In the R5, this appears to be about 2 GB of data, or about 40 shots. (The cache will appear to be about 60 shots deep because fast cards will allow you to fill up about half again as much data during the time it is uploading to the memory card. This number – reported in the viewfinder for those who have it set so – will vary mostly based on the ISO setting, as that will affect the size of the files.)
Since this comparison review of CFexpress cards was published, we’ve added data from three additional cards. The Mark II of the Delkin card is a new firmware version. The Transcend 512 GB card was added (and much liked) and the Lexar card was added (and, later, subtracted as it failed and the company could not replace it).
This is Walt the blue-winged teal, our test subject. We have 378,000 shots of him.
A critical factor – perhaps the most important for sports and wildlife shooters – is the number of shots you can rip before you get the unpredictable stutter of a full cache. In mechanical shutter at 12 FPS, the R5 will give you between 10 and 15 seconds of glorious, uninterrupted shooting. This is the sum of the buffer and the number of shots a card can manage to ingest while the buffer is filling. The Delkin consistently took the lead here, aside from the Lexar, which initially beat it out, but eventually l slowed upon its failure after a day of testing.
Of interest here is the fact that Sandisk’s smaller card sizes – which sport very similar performance metrics on the label – perform terribly relative to the 512 GB card. Unfortunately, this seems to be the norm. Sony’s smaller sizes also perform poorly in the other tests relative to the 512 GB, although they are keeping up with their bigger sibling in this test of initial speed capacity.
An early version of our cache depth test showed much more variation among the cards, but we discovered that running tests very quickly, one after the other, would increase a card’s temperature, and this write performance would be throttled. It appears from talking to reps at multiple manufacturers that this is something that will be true for all CFexpress cards. It also appears to be the reason you don’t see people brandishing minimum write speeds, because it is very temperature dependent. It also may be because an “effective minimum write capacity” number that indicated a speed below 40 degrees celsius would be embarrassingly low – likely around 550 MB/sec. for the best cards, which wouldn’t look cool emblazoned on a card label.
The good news is that there are very few people who will use their cards more intensively than we did for these tests, and the cards very rarely got hot enough to affect performance. After shooting about 3,500 stills in a row, formatting the card after about every 300, we saw some performance degradation that may have been heat related. This is likely an area that video shooters will be more concerned about. We did keep track of temperatures after this, monitoring how much the temperatures rose between each run of 30 seconds of shooting. It was so consistent among the cards, it’s not worth reporting other than to say that we’d see a roughly 2 degree celsius increase per run, and simply plugging 5 numbers into a computer between runs and formatting the card would provide enough time for the camera’s card slot to lose 1 degree. We never saw odd performance changes below 50 degrees C, and stills shooters are unlikely to see those temps. Even after several thousand shots in a few minutes, the cards were typically measured to be only in the mid-40s. Stills photographers won’t need to worry, but video people may find some cards lower writing capacity below the bitrate required for 8k shooting once a certain heat level is reached. More likely, the R5 will stop recording due to its own heat detection firmware routines. Video users of the R5 should be using an external recorder for high bitrate clips longer than a few minutes. That also has the benefit of removing recording time limits. The Ninja V external recorder today costs roughly the same as one of the 512 GB cards in this review.
Another alternative to using CFexpress cards and putting that heat internally in the camera can be found with the just-released ZITAY XFexpress to SSD Converter, reviewed by Matthew Allard over at Newsshooter.
Shots Per Half-Minute
Here is what happens when you set a camera on electronic shutter (20 frames per second) and mash that trigger for half a minute. The scale of this chart starts at 250 images, which exaggerates the differences between the cards. All the cards perform decently, even the pokey SanDisk ones of 256GB and below.
Here is where you see the smaller cards in the Sony system also start to falter, even though their labels claim equivalent performance with the 512 GB version.
It should be said that each of these cards have their different shooting rhythm styles. The Sony 512 has a pleasing, consistent rhythm to the shots, shooting roughly equal amounts on and off after filling the buffer. The Lexar was a big more syncopated, perhaps less predictable, before it failed. The Angelbird and ProGrade cards create more tension, with a spattery non-pattern. You never really know in which tenth of a second the shutter will start up again, or for how long.
The affordable Delkin surprises here by coming in 8th after winning the buffer test that showed how many uninterrupted frames it could take before the buffer caused skipping. It appears to be a sprinter of a card, perhaps with firmware that is more cautious on heat throttling, but that is just speculation. The Transcend card, too, sprinted just as fast, and wound up with the poorest performance in the remaining seconds after the buffer was filled.
The chart above shows the performance of the cards in mechanical (first curtain) shutter mode. This is favored by many with the R5 because its maximum of 12 frames per second is a reasonable rate when you’re not trying to capture the fastest movements. It’s also not very loud for a mechanical shutter. The R5 forces electronic shutter users into the 20 fps rate, which often is a drain of time later when culling pictures. By and large, the cards keep their relative rankings. The scale of this chart changes to a minimum of 200 frames, so the two charts of how-many-frames-in-30-seconds aren’t as different from one another as might initially seem.
Below, is a table of the claimed maximum write and read speeds of the cards. As discussed earlier, dynamic heat reactions prevent there being a real minimum write speed figure, but even so, these maximum speeds show the manufacturers being pretty generous with themselves. These numbers are an industry convention that doesn’t have a great deal to do with actual card performance. To round out the table, we added a couple additional rows with information just slightly more useful than the marketing claims.
The “Spooling Up” Gotcha
When looking at online forum discussions for Canon EOS R5 performance, we noticed that several people seemed to think that the R5 had a terribly long initial startup time, making it inconvenient for action and wildlife photographers. Many of us other R5 owners were puzzled by this, as the spooling-up time seemed to be impressively short. Over the course of jockeying CFexpress cards through the camera for a few days, we may have discovered why some are affected by this and not others. The camera “polls” the cards upon startup, and some can cause the camera to spend an extra 4 seconds during that time. In our tests, one in particular showed this: the Angelbird 660 XT GB. This isn’t as crippling as it may sound, as this polling happens only in two cases; during the turning on phase, and right after a card formatting. So if a user leaves the camera on, merely waking up a viewfinder won’t trigger the extra delay. But turning your camera off after each shot will.
[Update: Recent R5 and R6 firmware updates have changed this behavior so that the first time you start up with one of those cards, it will do the 4-second delay, but after that, turning the camera off and on produces only the normal, very short delay. There remains a 2-second delay before the viewfinder interface elements are superimposed on the sensor image. This means you can start shooting almost immediately upon turning the camera on, but your focus point, etc., will show up only after perhaps 2-3 seconds. This has been observed with an XT card, but reliable reports indicate that the firmware did not fix the issue for other variants.]
What about Video?
Video people really need a binary answer of yes or no… Does this card write the desired video format and not skip frames? With all of the cards in the first tier of performance, they can write 4k video until they are full; or more likely, until the R5 gives them an overheating timeout. Whether they do higher-bit video has not been tested, aside from the Sony 512, the Delkin 512s, the ProGrade Cobalt card, which all appeared to work fine. The Transcend card shot 10 minute of 4K HQ video, but could only manage 10 seconds of 8K. A future review will suss out what lengths you might expect from each card from each video format.
Of course, if you’re using the R5, and you’re at all serious about video, you’re recording to an external recorder, and the memory card is precisely the thing you’ve removed and put aside. There is some thought that some CFexpress cards may generate more heat than others. At the request of a forum dweller over on FredMiranda.com, we measured before and after temperatures during a test run of all of the cards, making sure the camera’s slot had cooled to at least a common maximum temperature (45 C). Keeping the temperature to that or lower, we saw no degradation in the data with higher temps.
Price mattered to start, of course, but when you figure out that the performance on these cards is pretty consistently good, price takes on an even greater importance as a differentiator. The table below shows the breakdown of prices as of mid-September 2020 at B&H. Since our first tests we added the Transcend 512 GB ($450 officially, but available for $328 elsewhere) and the Lexar 512 GB ($600) before we removed it from our consideration due to both its failure and the company’s inability to provide replacement service for US customers.
|Angelbird XT 660||Delkin 512||Prograde Cobalt 325||Prograde Gold 256||Sony 512|
|Sony 256||Sony 128||SanDisk 512||Sandisk 256||SanDisk 128|
CFexpress Card Reading/Writing
- There is some correlation between drive performance and actual performance in terms of throughput when used in a camera. The chart here shows a comparison of the drive performance (orange) versus the number of frames that can be blasted through the card in 30 seconds on an R5. These numbers are presented as indices, not actual performance metrics.
- Potential performance as a drive is much higher than actual delivered performance in a camera – by roughly a factor of 2. Even more interesting, that drive performance is only about half as fast as the claimed maximum performance. This means that the R5 is currently getting about a quarter of the performance promised by the maximum write figures.
- The correlation between drive performance and camera performance is inconsistent for some cards. This may have to do with firmware differences. Some smaller-than-512 GB cards seem to have similar performance between functioning as a drive on a computer and functioning as a memory card in the camera, where the others show a much better performance when used as a drive.
In conducting the tests for the cards hooked up to a computer, we saw some very strange results, eventually figuring out that major differences would be introduced not only by using different card readers, but also even which port we chose to hook it up to on the computer, and also what sort of cord was used. These results retested, standardized on a Sonnet dual CFexpress card reader using a Thunderbolt 3 port and cord.
Our experience was that on a USB port, the Sonnet was slower than the other two, but on the Thunderbolt port, the Sonnet was fastest yet. Combine that with different cords providing different bandwidth, and some computers having different levels of USB support on different ports, this is rife for confusion and points out that people can compare their own bandwidth results to ours to see if perhaps they could get significantly better download rates by tinkering with those factors.
[Since this review was first published, we’ve come out with a more comprehensive review of card readers.]
In the course of comparing cards, we discovered two genuine bugs. In talking these over with the manufacturers, we were surprised to see them being very open about this, with them asking specific questions to help narrow down the problem. The two manufactures, Delkin and Angelbird were not apprised of the fact we were reviewing the cards and were treating us as just another customer. When Delkin received word we had a problem with the card being read through a third party card reader, we got a call from the person who ran service at their California headquarters. When Angelbird was told of a very similar problem, they immediately sent one of their own card readers 1-day mail from Austria for free. In both cases, we had working cards in cameras within 36 hours. There are three important take-aways from this:
- The CFexpress standard, as interpreted by camera manufacturers, is still being established in implementation, so we can expect issues to crop up as new cameras are introduced and firmware needs to be adapted. There is no amount of testing that would solve the issue of a third party camera maker launching a camera with a slightly different protocol interpretation a year later, so all manufacturers are likely as vulnerable to this problem as all the others.
- The good news is that the card producers, with one exception, appear to be standing behind their cards both with a surprising service level, and also an attitude of collegial cooperation, using the customer community to help track down issues. In both cases, we had great follow-up communication to make sure we were shooting without issue. The best sales asset Delkin has is a guy named Mark, and for Angelbird it’s a fellow named Fabian. After all of this testing is over, I will personally be buying one more large card, and it’s likely going to be a Delkin or an Angelbird, partly due to performance/price, but also because they have given me the greatest confidence about their commitment to standing behind the cards.
- The problems noted above can be solved with the use of either the Delkin or the Angelbird card readers, which each work fine with all 10 cards tested.
We had a more concerning experience with Lexar. The card was super fast for a day of testing, but then started behaving erratically. For a time it seemed to have two performance levels, a fast and a slow one, and would roughly alternate between those states. We were able to speak to a support person in California and established that it was indeed acting inappropriately. The service rep suggested we return it to the retailer, however, as Lexar returns tend to be slow. We’d find out soon why. After hanging up, we got an email from Lexar indicating that its Chinese parent company, Longsys, is under US restrictions that prevent the firm from delivering replacement products to end users.
Indeed, the restrictions were covered by a PetaPixel story in 2018. But while the CFIUS restrictions put in place in August 2018 appeared to have caused the issues raised in that story, Longsys had since represented to investors, in presentations still available on the internet, that it had put all of its regulatory issues behind it. In fact, consulting the series of US Treasury Department databases on listed entities fails to pull up Longsys as a current entity in poor standing. Inquiries to Treasury were not returned at the time of the publishing of this review.
Lexar responded to our questions by email, stating, “The restrictions do not allow us to send physical product directly to our end user customers. We are still allowed to provide stock to our authorized resellers/retailers, however.”
We will continue to look into the details of Lexar’s service capacity, but in the meantime it is likely safest to use a retailer that accepts returns, and to perhaps avoid Lexar products until the issues are cleared up
All shots were taken as RAW files at 1000/th of a second. Two Canon R5 cameras were used to check consistency of results. Both mechanical and electronic shutter were used in a significant number of tests, just to ensure that there wasn’t a relative difference in performance among the cards (there wasn’t), with one shooting 12 frames per second (FPS) and the the other 20 FPS.
Two sets of data were taken at 250 ISO and at 3200 ISO to see if there were relative performance differences with the different sized files (there weren’t any observable). Some of the data averages are rounded to the nearest ten so we don’t mistakenly imply that our precision is better than it actually is.
Relative to one another, the cards in the same class as one another showed only about 10 percent variance in performance. There certainly is shown a different set of classes, though, when looking at the 128 GB and 256 GB cards versus the higher capacity models. Those smaller cards get a roughly 20 percent performance hit on average. So, with everything performing at a very, very high level, and little variation among them, factors like price and service weigh heavily. My own assessment is that Delkin wins here because the 512 GB card reviewed is less than $1 per GB, and they even sell one with four times the capacity for $0.50 per GB, which is less than half the cost per GB for most of the other brands’ largest cards. Their firmware issue was solved just this week with a new version. The card’s biggest downside is a longer full cache clearing (by 1.5 seconds) and a lower number of frames on very, very long shutter presses (>15 seconds). But it won the category for the number of shots taken in mechanical shutter before it stutters. That’s pretty representative. You can get the best card for the first 15 seconds of shooting, but that same card will perform in the lower half of the cards for the second 15 seconds of shooting. This sort of tradeoff is apparent with all the cards and across the different specs measured.
But, really, you can’t go wrong with that one, or the Angelbird XT 660GB, or the Prograde Cobalt 325GB, or the Sony Tough 512 GB, or the SanDisk Pro Extreme 512 GB, or – now – the Lexar Professional 512 GB. I’d just stay away from the sizes below those. Some of those larger cards will be a bit better in one area or another. If money isn’t an object, pick your most limiting factor and look at the relevant chart above to pick the winner. But for everyone else, just go get a big Delkin. And, while you’re at it, get one of those little Angelbird CFexpress readers, so you’ll never worry about which reader can read what card.
Even the people coming from the Sony A9 bodies or the Canon 1DX Mark III will find the throughput of the Canon R5 a bit mind-bending when using CFexpress cards. The SanDisk 512 card is pushing 560 MB per second to clear the cache. That’s 10 shots per second of 45 megapixel RAW files. That said, even the fastest performing card isn’t close to the claims on the labels. It is about 1/4 the capacity of the CFexpress standard (~2TB) and about 1/3 the claimed maximum write capability (~1.5TB). The implication is that better is still to come. A card functioning at the bandwidth claimed on these cards’ labels wouldn’t even need a cache. The card would suck in the data as quickly as the camera can produce it. I expect that will happen in a camera generation or two. We’ll be watching and reviewing.
Finally, we include a “too long; didn’t read” section below to show all of the tests across all of the original 10 cards. This may save you some reading, but it’ll certainly cause some squinting. We’ve titled this cityscape of a chart “Downtown CFexpressville.”
Special thanks to a few people who helped push this review along. Two Canon R5 shooters who wish to remain anonymous lent multiple CFexpress cards. LensRentals.com was able to provide a couple of the brands on short notice. Jeff, over at cameramemoryspeed.com gave great advice on figuring out methodology. Forum dwellers over at FredMiranda.com also helped figure out what tests would be most useful and helped inspire the effort in the first place.
– 30 –