In my personal opinion as well "specialized enough" chips are exactly the right way to go - though disclaimer, also At Nvidia, working on spec dec/external drafts :)
SIMD is for vector OPs, which you can use to compute matrices, but typically modern GPUs use specialized matmul cores - I.E tensor cores - that are another order of magnitude faster in terms of FLOPs.
Right, but most ANN ops aren't inherently matrix multiplications - that's just an efficient way of batching up a bunch of vector dot products - so single instruction, multiple data. What makes matmul viable for use here is because you want to do the same operation in parallel multiple times.
Color science/color management seems like a complete dark art; super interesting post. The rest of the blog is full of cool posts about film production in general.
It’s also a field chock full of misinformation. A couple of years ago I made a short film and when I was looking to color grade it I found many apparently well respected resources that offered outright contradictory information on what should be straight up technical topics. It was very surprising to me.
I’m not sure I really buy the original article here. The reason I shot in log is because it recorded a wider color gamut than otherwise and skipped in camera processing. Doesn’t mean I didn’t try to nail the exposure in the first place, just that I had more latitude to make adjustments in post without as much image quality loss.
Sure but either you create unrealistic scene where nothing is clipped but it lack a depth or you need wide gamut monitor. So technically, unless your content is created to cinema or HDR streaming service, log is pointless and you need properly lit scene with standard profile.
I know how to properly expose a shot. My camera shows me a histogram and I shoot with natural night so I don’t control the lighting, it’s impossible to create an unrealistic scene.
They mean that you can create a scene which has much greater dynamic range than your target display medium, which means that you either have to give it a very flat rendering or clip the highlights or shadows (or both). If you aren’t displaying the results on a high dynamic range display, then all you really need is an exposure aligned with your creative goals. Whether or not you record in log is largely moot.
Yeah I know what they’re saying but it’s easy to look at the histogram and also configure your camera to show you when it’s outside an SDR gamut, which I do. I’m sorry if that wasn’t clear.
You should have goal before shooting first. If not, then you need accept that your audience will never see scene in 14 EV stops and grade either to clip shadows or lights. Or... Just properly lit scene.
If you're paying actors and an entire film crew for their very expensive time, then you'd probably better have a clear idea of the visual result that you want to achieve beforehand.
In the analogue photography world there was/is hype around Cinestill, which is cinefilm rolled into 35mm cassettes to get the 'cinematic look'. I found out that most of the 'cinematic look' was just accurate skin tones (besides lighting and composition), and I can make any film scan have that movie still vibe.
In my opinion, a better & more comprehensive look at a similar idea for all state space attention types is this paper from AWS: https://arxiv.org/abs/2605.08301
They seem combinable to me, i.e. use the layer-selection from that paper and initialize using the Hedgehog-linearization as in this paper, then distill to smooth over inaccuracies of the approximation.
I mean, I can't remember what I had for dinner 6 months ago, but if I snuck into a presentation and anyone famous was announced as the surprise guest, I think I would remember that! I would assume most people do not recall everything with this much detail, just the (ha) memorable things.
It has! Hypernetwok/weight generation has been a thing for a long time - in fact my Master's thesis back in 2018 was based around PRNG weight generation for large percentages of the network. (https://open.library.ubc.ca/media/stream/pdf/24/1.0371928/3)
Why is this not overhyped in my opinion? Nothing has worked this well, ever.
Microsoft basically does the same with internal Azure services? When I was there our budget/usage was based on and consumed at the consumer price for GPUs and other compute. What happened behind the scenes really I have no knowledge.
Not true. Azure has a very different internal pricing. You can often get a glimpse of it if you have a.. particularly enterprising TAM. What happens sometimes (Like now when compute/GPUs are in high demand) is finance department would calculate the opportunity cost on allocating those resources internally vs selling them externally. The calculation isn't even always straight forward. What is the opportunity cost of using compute for internal CI or company operations vs selling them externally? It's not a straightforward equation and you would have to adjust the scale often depending on how clogged your internal processes get.
The extremely discounted internal cost promote waste, so there is always a room for improving efficiency, but that too has a cost.
It's not racist to acknowledge that countries have cultures that are different from other cultures and this results in different personalities on average. Every culture has its pros and cons.
With things like https://developer.nvidia.com/blog/co-designing-ai-models-usi... you can specialize one chip for decode/drafting, and the other for prefill and draft verification.
reply