Not a word on compute or interconnect speeds. All this really says is they stuck some HBM on a chip.
It can be rather difficult, mostly due to the occlusion calculations having to be conservative (must count visible things as visible, allowed to count invisible as visible, or things pop) and latency (must account for…
Just write a 10 TB hard disk full of plaintext mantras and let that bad boy spin at its usual ~5k rpm for a cost effective 50 PB of mantras per minute. Or go MAAS and write a few them into S3.
I'd say `x * a + (x * x) * b + (x * x * x) * c` is likely faster (subject to the compiler being reasonable) than `((((x * c) * x + b) * x) + a) * x` because it has a shorter longest instruction dependency chain. Add/Mul…
Did you explore microbatch sizes below 100? Curious about how far this can be pushed and what happens when approaching the limit microbatch size of 1.
Baseline time to grok something looks to be around 1000x normal training time so make that $20k per attempt. Probably takes a while too. Their headline number (50x faster than baseline, $400) looks pretty doable if you…
For a bit of reference, a full frame of Crysis (benchmark scene) was around 300k vertices or triangles (memory is fuzzy), so 3-10 log piles depending on which way my memory is off and how bad the vertex/triangle ratio…
The framing here is a little unfair, its DLSS that does not support AMD hardware, not the other way around. This is clearly a deliberate decision by Nvidia. Meanwhile AMD/FSR supports both AMD and Nvidia hardware.
Sounds like using the wrong metric, the upscaled image should be compared against the original full resolution one, not the downscaled one. Obviously you can't know what the full resolution one looks like when actually…
That list is only really useful for major styles that are well known/tagged, early SD versions knew a list of artists _much_ longer and more relevant than that. (some of the styles on that list are not related to 2d…
The thing is, most styles don't have names, only artists using them. If you remove artist names from the training data, you lose the ability to describe those styles, including as combinations (so you can't even do…
You have chroma and luma the wrong way around, chroma is color and luma is light intensity. Chroma is still the thing being subsampled. I've never really noticed effects from chroma subsampling so my guess would be the…
So I took a look at the actual shader, and it really should not require anything powerful (decade old, worst-you-can-get hardware might struggle). A much more likely explanation (that the gp is alluding to but didn't…
Not quite ELI5 and there's a few partially overlapping answers around already but here goes. The key part is the attention mechanism, as title of the paper may have spoiled. It works moreless like this: - Start with an…
If consciousness is binary like you state, we can hypothetically remove parts of it until we reach some arbitrarily small part after the removal of which the remainder is no longer conscious. This seems like a very weak…
The problem here is the definition of real, which is ultimately a matter of perspective. What is simulated rain to an external observer is real rain to one that only sees the simulation from within. In this sense real…
Based on some quick* testing using z-order instead of going row by row ends up taking 1-10% less space. Most images were on the lower end but I didn't run into any taking more space. * Swapping to z-order requires about…
Having a few false negatives in the control group shouldn't really matter since that would only make the groups more similar and hide any differences that exist rather than hallucinating ones that don't.
We have pieces of code at (gaming related) work that have very tight, and carefully thought out, error bounds (think 4 ULP) and can certainly break if float rounding doesn't work as expected. I'd expect a physics engine…
I don't have a good article on this but I'd assume the rocket delivers the satellites to an orbit slightly below the target one. The satellites can then move to the final orbit on their own (small transfers take much…
Collision detection and response are the hard parts and they are poorly suited for GPU implementation. Some specific cases (fluids, particles, maybe cloth) that don't require these in full work quite nicely and are used…
Vulkan doesnt really expose tile based rendering. In the demo the scene is split into tiles and since the geometry is static they can prepare commands for gpu execution once for each of these tiles. This cannot be done…
Not a word on compute or interconnect speeds. All this really says is they stuck some HBM on a chip.
It can be rather difficult, mostly due to the occlusion calculations having to be conservative (must count visible things as visible, allowed to count invisible as visible, or things pop) and latency (must account for…
Just write a 10 TB hard disk full of plaintext mantras and let that bad boy spin at its usual ~5k rpm for a cost effective 50 PB of mantras per minute. Or go MAAS and write a few them into S3.
I'd say `x * a + (x * x) * b + (x * x * x) * c` is likely faster (subject to the compiler being reasonable) than `((((x * c) * x + b) * x) + a) * x` because it has a shorter longest instruction dependency chain. Add/Mul…
Did you explore microbatch sizes below 100? Curious about how far this can be pushed and what happens when approaching the limit microbatch size of 1.
Baseline time to grok something looks to be around 1000x normal training time so make that $20k per attempt. Probably takes a while too. Their headline number (50x faster than baseline, $400) looks pretty doable if you…
For a bit of reference, a full frame of Crysis (benchmark scene) was around 300k vertices or triangles (memory is fuzzy), so 3-10 log piles depending on which way my memory is off and how bad the vertex/triangle ratio…
The framing here is a little unfair, its DLSS that does not support AMD hardware, not the other way around. This is clearly a deliberate decision by Nvidia. Meanwhile AMD/FSR supports both AMD and Nvidia hardware.
Sounds like using the wrong metric, the upscaled image should be compared against the original full resolution one, not the downscaled one. Obviously you can't know what the full resolution one looks like when actually…
That list is only really useful for major styles that are well known/tagged, early SD versions knew a list of artists _much_ longer and more relevant than that. (some of the styles on that list are not related to 2d…
The thing is, most styles don't have names, only artists using them. If you remove artist names from the training data, you lose the ability to describe those styles, including as combinations (so you can't even do…
You have chroma and luma the wrong way around, chroma is color and luma is light intensity. Chroma is still the thing being subsampled. I've never really noticed effects from chroma subsampling so my guess would be the…
So I took a look at the actual shader, and it really should not require anything powerful (decade old, worst-you-can-get hardware might struggle). A much more likely explanation (that the gp is alluding to but didn't…
Not quite ELI5 and there's a few partially overlapping answers around already but here goes. The key part is the attention mechanism, as title of the paper may have spoiled. It works moreless like this: - Start with an…
If consciousness is binary like you state, we can hypothetically remove parts of it until we reach some arbitrarily small part after the removal of which the remainder is no longer conscious. This seems like a very weak…
The problem here is the definition of real, which is ultimately a matter of perspective. What is simulated rain to an external observer is real rain to one that only sees the simulation from within. In this sense real…
Based on some quick* testing using z-order instead of going row by row ends up taking 1-10% less space. Most images were on the lower end but I didn't run into any taking more space. * Swapping to z-order requires about…
Having a few false negatives in the control group shouldn't really matter since that would only make the groups more similar and hide any differences that exist rather than hallucinating ones that don't.
We have pieces of code at (gaming related) work that have very tight, and carefully thought out, error bounds (think 4 ULP) and can certainly break if float rounding doesn't work as expected. I'd expect a physics engine…
I don't have a good article on this but I'd assume the rocket delivers the satellites to an orbit slightly below the target one. The satellites can then move to the final orbit on their own (small transfers take much…
Collision detection and response are the hard parts and they are poorly suited for GPU implementation. Some specific cases (fluids, particles, maybe cloth) that don't require these in full work quite nicely and are used…
Vulkan doesnt really expose tile based rendering. In the demo the scene is split into tiles and since the geometry is static they can prepare commands for gpu execution once for each of these tiles. This cannot be done…