Choosing a PC for High Speed Imaging with the Andor Zyla

We’ve been fortunate enough to get our hands on a 10-tap Andor Zyla camera. This is a new sCMOS camera that is capable of reading out its full field of view (5.5 megapixels) at 100 frames per second. At two bytes per pixel, that’s 11 MB per image, for a total data rate of 1.1 gigabytes per second. It turns out that acquiring and storing 1.1 GB/s is not so easy. A solid state drive (SSD) can deliver sustained write speeds of 520 MB/s, about half of what we need. The solution is to use multiple solid state drives, in RAID 0, so that we can write to them in parallel. Both Hamamatsu (for their Flash4.0) and Andor recommend four SSDs in RAID 0.  In principle, with four SSDs at 520 MB/s, we should be able to get a total data transfer rate of 2.08 GB/s, assuming nothing else is limiting.

However, getting this data transfer rate in practice is not so easy. We just bought such a setup to go with our new Zyla. It has an Intel RAID card and four 240GB Intel SSDs. The host PC has two quad-core Xeons with a 1600 MHz bus clock. In principle it seems that it ought to run somewhere close to 2 GB/s. However, it turns out that getting these speeds in practice is not so easy. Out of the box, it ran at about 500 MB/s. The problem turned out to be that we had not enabled write caching on the device in Windows; once we did that the write speed jumped to about 950 MB/s.  This is still lower than it seems like we should be able to get and not quite the 1100 MB/s we would need to continually stream data from that camera.  Unfortunately, figuring out how to adjust the raid settings for optimal performance is not so easy (for example) and there is a lack of real-world documentation on how to optimize RAID settings for maximum write throughput.  Partly this is because sustained write performance is a little bit of an unusual requirement – most applications care a lot more about random read/write performance. So it seems like for now I’ll be adjusting settings on the RAID controller to see what combination of settings gives the highest write throughput. If anyone reading this knows how to set up our RAID for maximum speed with out me blindly trying settings, please let me know.

It turns out that the PC hardware isn’t the only limiting factor in handling this much data. The Micro-manager team has been busy rewriting the Micro-manager core to minimize overhead and handle these high data rates in Micro-manager.  Right now we are able to run continuously with 15 msec exposure times, or 66.7 frames per second. Hopefully we’ll be able to get to the full 100 fps rate with a little tweaking. Not surprisingly, you can fill up hard drives pretty fast this way – at 66.7 fps, we’ll fill our entire RAID 0 drive in about 20 minutes.

Assuming we do get our RAID up to speeds above 1.1GB/s, I’ll post how we did it.

Cameras, Magnification and Field of View, Part 2

I had the pleasure last week of demoing a Yokogawa CSU-W1 spinning disk confocal – this is the new large field-of-view Yokogawa scanhead. Andor and Technical Instruments arranged a demo for us and paired the scanhead with an Andor Zyla camera. I’ll have more to say about the demo later – it’s a pretty cool confocal – but for now I want to focus on some field of view (FOV) issues it raised.

Continue reading

Camera sizes

Following up on the previous post, I thought it would be worthwhile compiling the pixel sizes and numbers of some existing cameras:

SensorDimensions (mm)Pixel Size (micron)MegapixelsExamples
Sony ICX2858.98 x 6.716.451.45Photometrics Coolsnap HQ2, Andor Clara, many others
Sony ICX6748.8 x 6.64.542.83Photometrics Coolsnap Myo, Raptor Photonics Kingfisher V
Sony ICX69412.49 x 9.994.546.05Raptor Photonics Kingfisher V
Fairchild sCMOS16.6 x Neo and Zyla, PCO Edge
Flash 4.013.3 x Flash 4.0

This is a very truncated listing of the cameras out there, and basically only includes cameras I’m directly familiar with.  I’m also excluding EMCCDs and only including cameras 1.4 MP or larger. The vendors mentioned here all many other camera models and there are a number of other vendors out there, some of which make pretty interesting cameras.

Artemis CCD, a company I came across while doing research for this post, appears to make some pretty nice CCDs. I’d be curious to know more about them.

For larger format CCDs there are a number of companies that target both the astronomy and the biology market. Apogee and Finger Lakes Imaging are two I’ve seen a lot. Many of their cameras use Truesense (formerly Kodak) sensors, which come in very large formats, but generally seem to have lower quantum efficiency and higher read noise.

Cameras, magnification, and field of view

For my inaugural post, I want to talk about something that I’ve been thinking about a lot recently – how to capture the maximum amount of information from your microscope.  A user came to me recently wanting to maximize the field of view he could acquire at high resolution from the microscope – he was doing an image based screen and wanted to maximize the number of cells he could capture in one field of view.  I immediately realized that our standard 1.4 megapixel, ICX285 based cameras weren’t going to cut it – this was a job for an sCMOS camera, or so I thought

Then I started thinking more about the problem. For his application, he didn’t need high resolution, so we were talking about imaging at 10 or 20x. When I started doing the math for the pixel size you need to acquire a diffraction-limited image from a 10x / 0.45 objective, I realized that our standard ICX285 cameras that are diffraction limited with a 100x / 1.4 oil lens aren’t diffraction limited for a 10x / 0.45 objective. Going from a 100x oil lens to a 10x air lens reduces the magnification by 10-fold, but the NA, and hence resolution, only drops by about 3-fold.  So you either need a 3X magnifier between your scope and your camera, or you need 3-fold smaller pixels.

Illustration of field of view

18 mm side port field of view, with inscribed and circumscribed cameras illustrated.

OK, so all the imaging we’ve done over the years with the 10x objective turns out not to be diffraction limited, and we need a camera with about 3 μm pixels if we want to be diffraction limited.  How many do we need? It turns out the side port of a Nikon Ti has a field of view of 18mm. The eyepieces and the bottom port have a bit larger field of view, 22mm, but since I’ve only ever seen one Ti with a bottom port, I’ll stick with the side port numbers.  If we want to truly maximize the field of view, will want a camera that’s 18 mm on a side. This will have black spaces in the corner, however, because the field of view is circular. If we want to have a camera that doesn’t have any black spaces, say, for tiled acquisition, we can inscribe a square camera in the 18 mm field of view. This gives a camera that’s 12.73 mm on a side, but we only capture 2/π = ~64% of the field of view. Continue reading