Sd video card: Choosing the Right Memory Card for Video Recording

Choosing the Right Memory Card for Video Recording

Every digital camera uses memory cards for storing your videos. Choosing the right one, however, can sometimes be overwhelming. After all, it is an important decision. This straightforward three-step guide is designed to help you find the best memory card for your specific use case.

1. Compatibility. Compatibility. Compatibility.

Memory cards come in various shapes and sizes, and different cameras use different memory cards. Hence, before all else, you have to know what type of memory cards are compatible with your camera. You can usually find this information in the user’s manual or on the manufacturer’s website.

Additionally, you can check our free compatibility charts to find the right card for your camera. These charts cover all the popular cameras and their compatible cards. Besides, we’ve also clearly marked which cards are recommended for best performance and highest resolution video recording.

What if my camera supports multiple cards?

Some higher-end DSLR and mirrorless cameras support multiple cards, making the decision a bit more complicated. For example, the Sony a7S III supports SDXC and CFexpress Type-A cards. In situations like this, it is generally recommended to pick the type with faster write speed, which, in this case, is Cfexpress Type-A. Your camera usually supports this faster card for a good reason (more about that below).

SD, SDHC, or SDXC?

Since SD cards are backward compatible, your camera’s manual may indicate it supports all of them: SD (Secure Digital), SDHC (high capacity), and SDXC (extended capacity).

We strongly recommend serious videographers avoid cards that are labeled SD or SDHC. SD cards have a maximum capacity of 2GB, which is too low even for today’s minimum video standards. SDHC cards do have a higher capacity, but they use the FAT32 file system and will split your longer movies into 4GB chunks because of that. Meanwhile, SDXC cards that utilize the exFAT file system have no such limitations.

At the same time, it’s important to remember that SD cards are not forward compatible. Meaning, SDXC compatible cameras can use SDHC cards but not the other way around. So don’t throw all the SD cards into one pot because they have the same form factor. Instead, make sure your camera actually supports SDXC cards before you buy one.

Similarly, CFast cards are relatively easy to mistake for CF cards. Keep in mind that those two are entirely different, despite being physically interchangeable. They have the same size and shape but completely different connectors.

Once you’ve taken note of what type of cards you are able to pick from, you’ll want to think about what kind of videos you will be shooting. Different tasks have different demands for speed and capacity.

2. Consider speed before capacity and minimum sequential write speed before maximum write speed.

It is natural to think that capacity is the first concern whenever choosing a memory card. In reality, it should almost always be the speed. Specifically, write speed.

Write speed determines how fast data can be saved on the memory card. This is especially critical when dealing with higher-resolution videos. If the memory card is too slow and unable to adequately handle the incoming data during recording, it may start dropping frames or store the video in lower than expected quality. In worst-case scenarios, your camera may stop recording altogether.

Minimum Sequential Write Speed

If not stated otherwise by the manufacturer, the write speed marked on a memory card usually does not indicate the minimum sequential write speed but the maximum a card can reach. Even though paramount, you should not be basing your purchase decision solely on the maximum write speed.

Recording videos means you’ll be continuously writing data on the card. For that reason, a memory card must foremost be reliable and capable of sustaining a certain minimum write speed instead of peaking in astronomical heights every now and then. An SDXC card with a maximum write speed of 200MB/s is completely useless for shooting 4K videos when it’s unable to sustain a minimum write speed of 30MB/s.

How to determine the minimum sequential write speed of SD, SDHC, or SDXC cards?

The highest minimum write speed any SD card can sustain is determined by the highest Speed Class marking on the card. Symbols of the Speed Class, UHS Speed Class, and Video Speed Class (with a corresponding number) indicate a specific speed at which your camera or any other device can write data on the card consistently. The higher the rating, the more data you can write to the card in the same amount of time.

As seen in the picture above, the so-called original Speed Class is denoted with a number inside the symbol “C.” The UHS (Ultra-High Speed) Speed Class is characterized by the number inside the symbol “U.” And the Video Speed Class is marked on the card with a letter “V” followed by a number. The letter “V” stands for video, and the number indicates the minimum sustained write speed in MB/s. Meaning that the minimum write speed of the V90 card will never drop below 90 MB/s.

One SDXC card can be certified in every speed class. It is so because there is a lot of overlap between the classifications. For example, Speed Class rating C10, Ultra-High Speed Class rating U1, and Video Speed Class rating V10 each refer to a memory card that has a minimum sequential write speed of 10MB/s. Similarly, a V60 card with a minimum sequential write speed of 60MB/s is pretty much automatically handed the Speed Class C10 (10MB/s) and UHS Speed Class U3 (30MB/s) rating as well.

For that reason, whenever looking for an SDXC memory card for video recording, you can pretty much just focus on the Video Speed Class and ignore the rest. The following chart helps to navigate the speed classes easier.

How to determine the minimum sequential write speed of CFexpress and CFast memory cards?

CFexpress or CFast memory cards have the Video Performance Guarantee (VPG) speed. It’s shown by the number inside a clapper icon. The principle is similar to the Video Speed Class markings on SDXC cards. The number indicates the minimum sequential write speed in MB/s.

If a CFexpress or CFast memory cards card does not have the VPG marking, turn to the manufacturer for more information.

How do you choose the memory card with the right speed for your specific use case?

For shooting full HD or 4K videos with lower bitrates and lower frame rates, a V30 memory card might be enough. If you’re shooting 4K video at higher bitrates and higher frame rates (60-240fps), you will need a faster V60 card. For the best performance and resolutions above 4K, choose the SDXC UHS-II cards with a V90 rating.

As mentioned before, if you are one of the lucky ones whose camera supports the CFexpress Type A cards alongside the SDXC cards, go with these. They not only ensure you get the maximum out of your camera, but they are also the most future-proof option, as more and more camera manufacturers adopt the format.

When to consider memory card read speeds?

Read speed determines how quickly you can retrieve recorded videos from the memory card. It does not affect the video quality nor can it hinder your camera’s performance, but it does affect the workflow.

A faster read speed decreases your wait time by allowing you to transfer data from the card to your computer faster. That means you can get to editing and sharing your work a lot quicker. Faster read speeds are critical for projects with short deadlines requiring quick turnovers.

Cards that have faster write speeds usually also have faster read speeds.

3. Last but not least – make sure the card has enough capacity so that you don’t run out of space.

Needless to say, every memory card can record only a certain amount of data. This amount is shown on the front of the card in gigabytes (GB) or terabytes (TB). As you probably know, the larger the card’s capacity, the more media it can hold.

Video bitrate, frame rate, length, and format each affect the size of your video files. For instance, a single minute of 1080p video may take only around 130MB of space. At the same time, one minute of ProRes 4K video at 880 Mbits/s will consume roughly 5.3GB. That works out to 318GB every hour. A lot, to say the least. You definitely want to use higher-capacity cards if you are recording at such high data rates.

However, even a lower bitrate 4K video can easily consume north of 45GB per hour. The same goes for higher resolution 1080p files. A 1080p video at 120fps and 4K at 60fps will generate almost identical file sizes when shot with the matching bitrate. So consider at least 128GB cards unless you are shooting way below an hour, and there’s absolutely no risk of running out of space.

In conclusion:

When it comes to stocking up memory cards for video recording, it is critical to choose wisely. There are close to quadrillion memory cards out there and not all of them can meet the demands of modern cameras and your workflow.

Always follow the guidelines given by the manufacturer of your camera and always consider speed before the capacity. Keep in mind that slower cards that are unable to sustain certain write speeds can cause recordings to fail and are highly likely to create bottlenecks in your workflow. There’s nothing more annoying than having to reshoot scenes because your memory card can’t keep up with you or your camera and keeps constantly failing.

Professional-Grade Memory Cards

If the safety of your data, uninterrupted workflow, and creative freedom are all mission-critical to you, consider ProGrade Digital memory cards.

We manufacture a wide range of memory cards and accompanying accessories, all created specifically for professional cinematographers and photographers. Plus, all our memory cards with the “R” Refresh Pro logo on the label are compatible with our one-of-a-kind Refresh Pro™ software, enabling you to check your card’s health and refresh it back to factory-new condition. Even after years of use.

Still undecided or need more help?

Our friendly staff is just one click away and always ready to answer your questions and assist you with choosing the best memory cards for your specific use case.

Shop memory cards>

Contact us>

Speed Class | SD Association

SD Standard for Video Recording

There are wide discrepancies in memory access speed depending on the SD memory card manufacturer and brand. Varying speeds make it difficult to make out which card can surely record streaming contents.Recording video require a constant minimum write speed to avoid ‘frame drop’ during recording for a smooth playback. The SD Association has defined various Speed Class standards to answer a demand for advanced video quality recording. Speed Class symbols indicated to host and card products help users decide the best combination for reliable recording (no frame drop). There are three kinds of speed indications:

Speed Class*, UHS Speed Class** and Video Speed Class*** symbols with a number indicate minimum writing speed. This is mainly useful for camcorders, video recorders and other devices with video recording capabilities. Regarding bus mode, it is necessary to use a bus mode fast enough that does not affect memory write speed. C10 is used in High Speed mode or faster, U1 and U3 are used in SDR50/DDR50 or faster, and V60 and V90 are used in UHS-II mode or faster.

Video Speed Class is defined to answer a demand for high resolution and high quality 4K8K video recording and it also has an important feature to support next generation flash memory such as 3D NAND. Furthermore,as it covers speed of HD(2K) video, it is possible to integrate into Video Speed Class from now on.

*
The Speed Classes defined by the SD Association are Class 2, 4, 6 and 10. Class 10 can be applied to High Speed Bus IF product family.
**
The UHS Speed Classes defined by the SD Association are UHS Speed Class 1 (U1) and UHS Speed Class 3 (U3). U1 and U3 can be applied to UHS Bus IF product family (UHS-I, UHS-II &UHS-III).
***
The Video Speed Classes defined by the SD Association are V6, 10,30,60 and 90. V6 and V10 can be applied to High Speed and UHS Bus IF product family. V30 can be applied to UHS Bus IF product family. V60 and V90 can be applied to UHS-II / UHS-III product family.

SD Speed Class

Video Format

Best Combination between Speed Class Host and Card

Speed Class supported host can indicate Speed Class symbol somewhere on the product, package or manual.Consumers can find the best card for a host via Speed Class symbol match; choose the same or higher class symbol card than class symbol of the host indicated.

For example, if your host device requires a Speed Class 4 SD memory card, you can use Speed Class 4,6 or 10 SD memory cards. If your host device requires a UHS Speed Class 1 SD memory card, you can use UHS Speed Class 1 or 3 SD memory cards. Video Speed Class is also the same. Note that expected write speed will not be available by a combination of different class symbols between host and card such as Host Class 10 and Card U1, Host U1 and Card V10, etc. even those are indicated to the same 10MB/sec write speed.

Fragmentation and Speed

By repeating deletion and write of files, data area is gradually fragmented and it influences write speed.Generally, write speed to a fragmented area is slower than sequential write speed due to flash memory characteristics. In an era when memory capacity is not large enough, fragmented write needed to be considered. However, high capacity memory card is available at this time, Speed Class write is defined to perform sequential writes to a completely un-fragmented area (called “Free AU”). It makes Speed Class controls of host easy. On the other hand, even unused memory exists in total, there is a possibility that host cannot perform Speed Class recording. In that case, data arrangement to reduce fragmented area or move data to anther storage to re-format the card will be required. Video Speed Class supports “Suspend/Resume” function that can stop and retrieve sequential write. By using the function, it is possible to improve memory usage ratio considerably.

Matrox Millennium G200 SD Video Card

Description

Matrox launched the Millennium G200 SD in 1998. This is a G200 architecture desktop card based on 350 nm manufacturing process and primarily aimed at gamers. It has 8 MB of SDR memory at 0.11 GHz, and together with a 64-bit interface, this creates a throughput of 896.0 MB/s.

In terms of compatibility, this is a single-slot card connected via the PCI interface. The length of the reference version is 150 mm.

We don’t have test results for the Millennium G200 SD.

General

Information about the type (for desktops or laptops) and architecture of the Matrox Millennium G200 SD, as well as when sales started and cost at that time.

Performance rating not included
Architecture 90 022 G200 (1998−1999)
GPU Eclipse
Type Desktop
Release date 1998 (25 years ago) 9002 0

Features

Matrox Millennium G200 SD’s general performance parameters such as number of shaders, GPU core clock, manufacturing process, texturing and calculation speed. They indirectly speak of Matrox Millennium G200 SD’s performance, but for precise assessment you have to consider its benchmark and gaming test results.

900 18

Core clock 84MHz of 2610 (Radeon RX 6500 XT)
Number of transistors 10 million out of 14400 (GeForce GTX 1080 SLI (mobile) )
Process 350 nm of 4 (GeForce RTX 4080)
Texturing speed 0.08 of 969.9 (h200 SXM5 96 GB)

Compatibility and dimensions

Information on Matrox Millennium G200 SD compatibility with other computer components. Useful for example when choosing the configuration of a future computer or to upgrade an existing one. For desktop video cards, these are the interface and connection bus (compatibility with the motherboard), the physical dimensions of the video card (compatibility with the motherboard and case), additional power connectors (compatibility with the power supply).

interface PCI
Thickness 1 slot nutrition no

RAM

Parameters of memory installed on Matrox Millennium G200 SD – type, size, bus, frequency and bandwidth. For video cards built into the processor that do not have their own memory, a shared part of the RAM is used.

Memory type SDR
Maximum memory 8 MB of 128 (Radeon Instinct MI250X)
Memory bus width 64 bit of 8192 (Radeon Instinct MI250X)
Memory frequency 112MHz 90 022 of 22400 (GeForce RTX 4080)
Memory bandwidth 896. 0 MB/s of 3276 (Aldebaran)

Video outputs

Types and number of video connectors present on the Matrox Millennium G200 SD. As a rule, this section is relevant only for desktop reference video cards, since for laptop ones the availability of certain video outputs depends on the laptop model.

Video connectors 1x DVI, 1x VGA

API support

APIs supported by Matrox Millennium G200 SD, including their versions. DirectX 5.0 OpenGL no of 4.6 (GeForce GTX 1080 (mobile)) OpenCL no Vulkan N/A

Benchmarks

These are the results of Matrox Millennium G200 SD rendering performance tests in non-gaming benchmarks. The overall score is set from 0 to 100, where 100 corresponds to the fastest video card at the moment.


We don’t have test results for the Millennium G200 SD.


Recommended processors

According to our statistics, these processors are most often used with the Matrox Millennium G200 SD.


atom
D410

100%

User rating

Here you can see the rating of the video card by users, as well as put your own rating.


Tips and comments

Here you can ask a question about the Matrox Millennium G200 SD, agree or disagree with our judgements, or report an error or mismatch.


Please enable JavaScript to view the comments powered by Disqus.

Making a generic build of Stable Diffusion web UI for AMD/NVIDIA

This content was written by a website visitor and has been rewarded.

Foreword

Usually, neural networks are associated with NVIDIA video cards, at least because people prefer to work with CUDA, and the company itself is rapidly promoting all sorts of interesting things, but AMD video cards are also capable of working with neural networks, and even there are DirectML version of the Stable Diffusion web UI.

But there is a problem, the software that works with AMD video cards is, to put it mildly, very bad, and I already had an unpleasant experience with such software:

However, I didn’t get a graphics card from AMD, but will that stop me? Of course not, I noticed that one person has an AMD graphics card, and at the same time he already had a working version of SD WebUI DML…

So I got what I needed, yes, this is an archive of the SD WebUI DML application that worked with the AMD Radeon RX 6600 video card:

The files are unpacked, and of course they are absolutely useless, because Stable Diffusion web UI is in fact “Linuxoid” software, which means that it will not work if you just transfer the application, the swamp of dependencies will not allow …

I I can’t just pick up and run this pile of files, even if I had an AMD video card. But it’s okay, I know what to do with this rubbish:

Why did I call this bunch of files junk? It’s simple, this is rubbish drowning in a swamp of dependencies, and I’m just going to make a working application out of this rubbish:

It could work for another person as much as necessary, but it won’t even start for me, probably my hands are crooked? No, everything is much simpler, the developers of such software have crooked hands.

How dependencies destroy the Linux ecosystem and not only

Well, okay, let it be on the conscience of developers (if they have any conscience at all), who release software unsuitable for normal launch and use.

This whole fuss with commands, dependencies and other obscurantism that can happen in the process is simply terrible, especially considering that all this is dependent on the Internet and needs to be done anew on each operating system separately, because it is not portable:

Assembly

Where to start? Perhaps from the dependencies, you need to cut off all the dependencies that are possible, and in order to facilitate the work, I will take the developments from my previous build and simply adapt it to the new build:0005

That’s one of the reasons why it’s impossible to just transfer SD WebUI to another PC or to another OS, this is the “pyvenv.cfg” file, but it’s okay, in my build this misunderstanding was fixed with a simple BAT file.

By the way, I have Windows 7 loaded now, I should load Windows 10, because the DirectML curve does not work in the supposedly “outdated” operating system…

models.

Fortunately, there were no additional problems, the dependencies stuck normally, except that the warnings are due to the fact that I do not have an AMD video card: with DirectML versions of SD WebUI…

A bit of magic with launch options and it worked even with NVIDIA video card: , this time I took files that already worked with AMD video cards, i.e. I did not build the application from scratch according to the instructions from the repositories.

Although not without problems, it speaks of a lack of memory… at the first generation, everything seemed to work out, but it didn’t work out again, swearing at the lack of memory.

In general, I started to pick the settings, it still fills the memory of the video card, but now I was able to generate several pictures without any problems.

Although memory allocation problems still haunt me. What I also noticed is the speed of generation, the CUDA build of SD WebUI runs at about ~1.6 iterations per second with similar settings, but the DirectML version pulls only ~1.1 iterations per second.

However, CUDA does not work with AMD video cards, so it’s better than nothing at all:

Even the “–lowvram” parameter does not completely solve the problem of lack of memory, this is how it was generated, and at the very end of the work it was interrupted.. But it’s worth recognizing that without “–lowvram” at a size of 1024×1024 it broke off at the very beginning of the generation.

I experimented, tried everything, but still I couldn’t master 1024×1024, everything always broke off, if not in the middle of the process, then at the very end:

And then I realized that the problem is not that I have a video card from NVIDIA, but in DirectML itself it’s crooked, because for a person with an 8 GB RX 6600 it generally generates adequately only in the size of 384×384, then what I have achieved relatively normal generation of 512×512 can already be said to be progress, but I still aim at 1024×1024 . ..

Moreover, it is important to note that even people with a Radeon 7900 XTX have a memory shortage problem:

I can generate 512×512 images on my 8 GB GTX 1070 without any problems:

Okay, I’ll do better with the design of BAT files, apparently the memory problem cannot be solved by simple methods.

In general, I redid the BAT files, of course I check that it works, I generate an image with the processor, yes, for a long time, 10 seconds per iteration with the R7 2700X, but it always works stably:

Then I add several extensions, and I have serious fears about the “multidiffusion” extension…

However, I can’t say that everything is terrible, the Tiled Diffusion extension has worked and it has borne fruit, though you still need to know how to use this extension, otherwise the results are very strange.

Maybe I don’t know how to work with the Tiled Diffusion extension, but even I was able to generate a 1024×1024 image with only 8 GB of video memory, however, if I used the CUDA version of SD WebUI, then without “crutches” I could generate such a size , but now we are dealing with DirectML. ..

In any case, the extension can obviously be useful, so I definitely leave it in the kit:

What’s next? Of course, to check the possibilities, my 8 GB GTX 1070 was enough to generate an image with a size of 768×512, but 1024×768 was no longer enough, and it is important to note that the lack of memory occurs at the VAE stage, i.e. the image was generated, but the VAE runs out of memory.

Just to confirm the problem when working with VAE, I turned it off completely:

But in the end the problem was not solved … The maximum image size that I managed to generate is 768×768, the first without VAE, the second with VAE:

And now a fun fact, the lack of memory occurs at the very last stage when the image is displayed in the viewport, and if you force the preview size to “full”, then the error will manifest itself from the very first iteration . ..

In fact, I can generate a 1024×768 picture, yes, but when it is displayed in the preview, there is a lack of memory and the result is reset, and when the preview size is set to full, then the lack of memory occurs immediately, and the complete disabling of the preview is not solves the problem, because the final result will be formed anyway, and there is an error.

However, by playing around with the launch options, I managed to generate an image of 1024×768…

Sometimes, of course, generation can break with an out of memory error, but in general, images of 1024×768 in size are quite possible to generate now, and in fact at the very beginning it was a problem to generate a 512×512 image, so I didn’t waste my time:

9 0020

Unfortunately, you can forget about generating in 1024×1024, at least using the DirectML version of SD WebUI with an 8 GB video card … But in 1024×768 it is quite possible to generate, although sometimes you can lose the result due to lack of memory:

In any case, this is already an acceptable result for the build, with a size of 768×512, all 20 images out of 20 were generated without errors, and a few dozen more “behind the scenes”:

then the routine of the following plan:

Next, you need to go through the functionality of the application, of course I already copied all the models so that it works without Internet dependencies, but suddenly I missed something, so yes, I double-check everything, except for LDSR of course, because it causes problems breaking interface:

One problem, of course, was found, SwinIR_4x tried to work on the processor instead of the video card, but could not, because the processor can not work with Float-Point 16 (FP16), this of course can be solved by switching SD WebUI to Full-Precision (FP32) mode , but this will negatively affect memory consumption, for the sake of one algorithm I see it as inappropriate. ..

But with the training of Hypernetwork and Embedding, everything is not very good, however, I already heard somewhere that the DirectML version of the Stable Diffusion web UI has problems with this, even running in FP32 mode I could not do anything, swearing at mismatched data types disappeared, but there was a curse about lack of memory…

In general, training networks in the DirectML version of SD WebUI will not work.

This completes the checks, it remains to work on the ReadMe file, and send the assembly to a person with an AMD video card for re-checking, after which it will be possible to finish both the article and the assembly.

And yet, there will be no support for Windows 7 this time, because the DirectML dependency curve does not work in Windows 7, yes, this is not CUDA from the “terrible and bad” NVIDIA… If everything was simple with CUDA, then with DirectML is much more complicated, I just don’t have any desire to try to modify this bike with square wheels.

It’s time to archive the assembly, and then I thought, because some people do not need the models in the kit, so why not make several options to choose from? The first option “All in One” (all in one), weighs more than 4 gigabytes, but is ready to work immediately after unpacking, and the second option is to do without models, the user will need to download the models himself.

Yes, it looks great, for everyone who needs a standalone build, there is “All in One” a little more than ~4 gigabytes in size, downloaded, unpacked and ran. But for those who do not care about complete autonomy, there is the option “NO_MODELS”, the name speaks for itself, I removed the models from the assembly, from which the archive with the assembly weighs only ~ 1 gigabyte:

In general, the assembly worked without problems and with AMD Radeon RX 6600 graphics card, as far as 8 GB of memory was enough, of course, I said not to immediately put 80 passes on the image at 1024×1024 because it would fall at the end, but the person did not listen and in the morning caught a memory shortage error, and the lowvram option did not help . ..

However, in the mode of operation without a video card (Start WebUI CPU), you can generate images even as large as 2048×2048.

And now a few words about system requirements, why 64 GB of RAM?

In fact, it will be able to work with less system memory, but the application takes ~15 GB of RAM at the first generation in the size of 512×512 with standard settings…

But in addition to SD WebUI DML, you still need to keep the browser running, which also takes up RAM, and the operating system is not a fluff, so the system requirements are 64 GB of RAM, you could of course specify 32 GB, but I’m not sure enough is there such a volume without a crutch called “paging file” if you use all the available functionality …

And for the sake of small edits in the ReadMe file, I really had to repack the archives, but this is how truly standalone software works, someone needs to put everything together, so that the user does not have to dance over the repositories collecting the same piece by piece.

In any case, it’s time to upload the assembly to GitHub, with mobile Internet this is certainly an unpleasant task, but what can I do, I have no other options besides mobile Internet at the moment, for like 5 months …

Conclusion

As you can see, making an initially “Linuxoid” application suitable for normal launch and use is not such a difficult task …

Just break the dependencies by placing them in the same folder with the application, make a crutch for venv, configure everything, test, arrange so that users understand what’s what (the rule of least surprise), pack and upload to the repository, here the assembly is ready!

Exactly, I almost forgot, because you need to specify a link to the repository where you can get the assembly for free, without SMS and registrations:

Build repository: ( https://github.com/Shedou/Neuro ).

Well, for Linux fanatics who have facts stuck in their throats, there is always a “blacklist” button, use it to your health and do not thank for the discovery:

There is such an approach to development “the worse, the better”, but practice shows, in reality it is usually worse than better, at the same time, I took up the revision of the initially crooked software, which means that the approach partly worked.