Rendering Engineer at @disneyanimation working on Disney’s Hyperion Renderer and more. Previously @pixar, @dreamworks, @cornell, @penn. Views here are my own.
To new followers of mine: sorry to disappoint, but I mostly just post about nerdy graphics stuff. Sometimes I promote stuff from my day job at Disney Animation. Recently I posted some takes on AI, but I'm stopping that since I don't have much useful to add to the discourse.
@jasondfarris If it has the ConnectX-7 link that the DGX Spark has, that would certainly be interesting!
I think you can only cluster two DGX Spark units together though.
Microsoft announced a NVIDIA RTX Spark based dev box, and I love that they went for a really out-there design instead of just aping the Mac Studio.
The question comes down to price again though. $2K is an instantly interesting Mac Studio rival. $4K and up is dead on arrival.
@Lndsp12 Yeah, the memory bandwidth is tiny compared to Apple Silicon, and both the CPU and GPU are slower than M5 Max. That's why it can't be more expensive than Mac Studio if it wants a chance; lower specs are okay if the price is correspondingly lower too.
There's a lot of Surface hardware right now that is in a vacuum cool and interesting hardware, but once the price is factored in, it's basically instantly dead on arrival compared with recent MacBooks, Dell XPS, etc.
For example: the problem with Nvidia's DGX Spark box isn't the chip; the chip is very cool! The problem is that they're priced at around $5k starting, and for that price you can get a M5 Max MacBook Pro that has a faster CPU and GPU, same amount of RAM, and also it's a laptop.
I'm surprised at how much people have already drawn conclusions about Nvidia's RTX Spark chips without either benchmarks or pricing info. So much of if it's good or bad depends on pricing! Like, sub-$3K is probably a home run, above $5K is dead on arrival, but nobody knows yet!
@planetzomax Yeah. It's really bad on some older patents where they don't give IORs and instead just give references to like, available glass at the time which may not match what's available now. In some cases I think Bill Claff wound up just re-solving for the IORs in Zemax from scratch.
@planetzomax ...and sometimes (often) even if you can get the prescriptions that were calculated in Zemax, they still won't reproduce the real lens exactly, because the real lens has a bunch of hand tuning and adjustments made in the grind / final assembly.
@planetzomax I learned the hard way that a lot of the most interesting anamorphics and cine lenses by companies like Panavision and Arri and others aren't even patented, so there's no data out there at all. They're all kept as trade secrets instead.
@planetzomax Cool stuff. Yeah I mostly use Optical Bench as well. I talked with Bill Claff (he runs Optical Bench) a few years ago; he puts a lot of effort into trying to fill in holes / missing stuff from patents in Optical Bench, but it's hard.
@planetzomax Something odd might be going on with your polynomial optics implementation if it's slower than tracing lens elements. I've usually found it to be at least an order of magnitude faster. But I'm not sure if OSL changes things here; mine is all C++ for both approaches.