Hacker News Reader

Comment by degenerate | Hacker News Reader

It doesn't really capture 10 trillion frames per second. The laser pulses at a regular interval, and the combination of the regular pulsing + multiple pictures taken at different points along those pulses makes for 10 trillion distinct frames captured within the timeframe of 1 second. There is no actual video captured.

I get that being a tech journalist is difficult, you have to juggle between the tech and the layman. But after writing a headline, read it back to yourself and ask: will this put the wrong idea in people's minds? If the answer is yes, rewrite it... even if it sounds less cool.

View on HN

I don't quite understand how your original point refutes the headline. They are saying that it captures frames at a rate that, measured in seconds, would be ten trillion frames per second.

The headline seems entirely accurate.

How long can they maintain that frame rate for?

How is that relevant?

Just trying to understand.

Sorry, I thought you were challenging the notion that it's 10 trillion frames per second because it's not a full second (someone else did that elsethread)

Apologies for my tone

The article mentions 25 frames in 400ns...

From what I understand, it is a camera, so as long as you can record data fast enough. I didn't read any limitation of the current setup across time (there is one across space though). So far, it is only 25 frames captured, it is fundamental research, not yet a product people can get value of.

Note that what we want to observe with those cams are very short, transient phenomenons. When I was doing my internship at a particle accelerator called GANIL, we were only recording 0.5 second, which already represented close to 1TB worth of raw data. It takes months to interpret and analyze results.

EDIT: typo

That's incredible. How long did it take to write 0.5s of data to disk? I'm guessing there's no way to sustain this as you'd be so far behind after only a single second. I'm pretty sure we can still only store a few gigs per second. Please correct me if I'm wrong. Very interesting though!

The best way to think of this is it might take 100 seconds to 'record' those 10 trillion frames that occur in 1 second.

That doesn't seem to make sense, but imagine this. You want to shoot 100 frames of the first millisecond of an airsoft pellet leaving a gun, but you have a camera that only shoots around 2 frames per second.

Your airsoft gun shoots 1 ball exactly (1ns accuracy) every second, exactly the same velocity and direction.

You have your camera, that only captures 2 frames every second, but this camera has an insane shutter speed, 1 microsecond, and has a shutter with that you can time to the gun exactly.

You can also delay the release of the shutter by 1 microsecond increments.

So, you start by taking 1 picture, 10 microseconds after you shoot your pellet. Then in the next second, 20 microseconds, you do this 100 times. You stitch this all together, and you have a video in super slow motion of an airsoft pellet leaving a gun. It just happens to be 100 different airsoft pellets.

I agree that the article isn't very clear on this, but I believe you're describing the previous work.

The new innovation here actually records the frames right after each other of one single event:

> The first time it was used, the ultrafast camera broke new ground by capturing the temporal focusing of a single femtosecond laser pulse in real time (Fig. 2). This process was recorded in 25 frames taken at an interval of 400 femtoseconds and detailed the light pulse’s shape, intensity, and angle of inclination.

I'm more curious about how they can store it. even if each frame could somehow be encoded as a single byte, you're looking at ~10TB/s, which afaik exceeds even the bandwidth between L1 cache and the execution core of a modern cpu.

It may take 100s of seconds to capture the data.

It does not capture events as they occur. You can't do a 10Tfps recording of a lightning strike with it for example because every lightning is different. It only allows you to synthesize an averaged video of repeatable events.

Did you read TFA? It seems that you are describing prior techniques for ultrafast image capture, whereas this technique is for capturing single, non-repeatable events.

> Using current imaging techniques, measurements taken with ultrashort laser pulses must be repeated many times, which is appropriate for some types of inert samples, but impossible for other more fragile ones. [...] The first time it was used, the ultrafast camera broke new ground by capturing the temporal focusing of a single femtosecond laser pulse in real time (Fig. 2). This process was recorded in 25 frames taken at an interval of 400 femtoseconds and detailed the light pulse's shape, intensity, and angle of inclination.

The technique of making "ultrafast" movies of repeating phenomena by imaging "one pixel" at a time and then assembling the raster-scan of all the pixels has been used for a long time.

The technique described in the paper is different from that....

    > Thus far, established ultrafast imaging
    > techniques either struggle to reach the
    > desired exposure time or require repeatable
    > measurements. We have developed 
    > single-shot 10-trillion-frame-per-second 
    > compressed ultrafast photography (T-CUP),
    > which passively captures dynamic events with
    > 100-fs frame intervals in a single camera exposure.

I feel like many readers of news carry this misunderstanding, so just to clarify: journalists very often do not get to write their own headlines. Many times they aren't even consulted on the headline. They may be just as surprised and dismayed as a reader/critic.

It is frequently editors who write the headline, and it might even be editors who weren't involved in the reporting. Editors are always writing headlines with a balance of different motives and goals. Strict accuracy isn't always on the top of that list.

In this case, the ambiguous statement is repeated 3 times in the article.

That is not the case, from the article that you criticized:

Obviously they're only recording extremely small timeframes with this setup, but it is indeed real time.

Note that a rate such as "x frames per second" doesn't mean it needs to be able to operate for the full second. That would be like saying you can sprint at "12 miles per hour" means you can run all of 12 miles in an hour.

Some people can...

That's irrelevant to the point.

The fact that "statement A does not imply statement B" itself does not mean "statement A implies statement B is false".

Just because the phrase "X can run at 12 miles per hour" does not imply that "X can run 12 miles in an hour", it does not mean the ummonk was claiming that no one can run 12 miles in an hour.

Watching a 1 second video shot at 10 trillion fps would make for a very boring lifetime or 15. A 240Hz display would take ~1300 years to display all 10,000,000,000,000 frames.

"20 mph" (slower than the fastest 100m sprint, but faster than the fastest half marathon) would be more clear of an example.

"The first time it was used ... the process was recorded in 25 frames taken at an interval of 400 femtoseconds"

1 / (400 femtoseconds) = 2 500 000 000 000 hertz

So definitely in the "trillion fps" for a very short period of time with multiple frames.

What is confusing is why they call it 10THz when it appears to be (on average 400/25=) 16fS/frame or closer to 60THz.

It's almost like they are underselling it?

They mentioned that in some of the use cases, the thing they're imaging is so fragile that it can only tolerate one laser pulse. So I don't think you're correct.

There's even an easy analogy for this! Putting strobe lights on fast, repetitively moving machines allows you to see their motion slowed down.

EDIT: I'm responding here to the parent poster's claim, not to the original article. For more information on the technique used in the article, see this other article[0] about "CUP" vs this article about "T-CUP". In CUP they use a random 2D sampling of a scene projected onto a streak camera.

[0] https://www.osa-opn.org/home/articles/volume_26/december_201...

That's not what they're doing. They're seeing the actual motion of something that doesn't necessarily move repetitively.

Agreed. We work with traditional high speed cameras (up to 1Mfps) and I thought, wow, someone is going several orders of magnitude faster??

Turns out the research is super cool and the paper is well written and interesting but it's not what I thought.

Reply on news.ycombinator.com