In `libnvidia-nvvm.so` the string `cutlass` appears right after `Memory Dependence Analysis` and `memdep`. Perhaps it acts as an optimization attribute of some sort, where the compiler is allowed to make assumptions about the kernel's behavior that are not valid in general?
The link is long dead and the Wayback machine doesn’t have a copy.
But in 2001 ATI was caught applying optimizations to Quake 3 when someone realized if you renamed the executable from “quake” to “quack” the score dropped a ton. It was a big scandal.
I know that’s common now but that wasn’t a thing that was done at the time.
Was it a scandal at the time? My understanding of how per-game card-driver optimizations work today is:
1. AAAA Game Studio shits out another unoptimized clunker
2. nvidia considers it a reputational risk if games run at 30 FPS on a 5090
3. They go in, look at the perverse ways the game misuses rendering primitives, and then hacks shit in to make whatever bad things they're doing less bad.
As a gamer, this seems fine to me and i generally blame the AAAA devs for being bad at their jobs or AAAA studio leads for being ok shipping unoptimized messes.
As a software developer, it almost certainly has a bad effect on the ecosystem long term. "Hacks shit in" is the very definition of technical debt, and that has a cost that someone, somewhere is going to have to pay in some form.
I can't reply to the person that replied to you, so
> You’re looking as a dev, but the reality is that a consumer cannot see technical debt.
The consumer can't _see_ technical debt, but they sure as heck can be impacted by it.
- Technical debt means the code base is harder to work with later. So fixes/enhancements take longer to make it into the code (and sometimes never can)
- This particular type of technical debt means the code by the game developers sets precedent, and the next developer may us it as an example. So the amount of code incorrectly using the api grows faster over time
You’re looking as a dev, but the reality is that a consumer cannot see technical debt. If the studio churns out a game, the vendor sprinkles on some optimizations, people play it and move on, then the tech debt just vaporizes into the void. It’s not real at that point.
Except that if a developer has that kind of market pull, nVidida will gladly help those devs with getting it right. They are excellent at maintaining developer relations.
In at least one past version of Windows (circa 1990s), if you tried to replace the default web browser of IE with another choice you were given an Open File dialog window to choose the executable.
Funny quirk, though: that particular window wouldn't show files named firefox.exe. It would accept that as typed input, if you were at the correct folder, but the file listing omitted that particular file.
Maybe it was mozilla.exe; it was a long time ago. But that was the discovery that pushed me off IE forever.
I vaguely remember that being the start of the browser prompts to set your current browser as the default. It was so hard to just configure that they had to build a way to set it within the browser.
You saw that again in more modern times when Microsoft removed support for the APIs they provided to set browser defaults, forcing browser makers to write step by step instructions on what to click to set the default browser.
I believe they walked that back, but it left such a bad taste that I switched my installation of Windows from default mode to EU mode in order to avoid it. And come to think of it, I haven’t used my windows machine for much outside of AI in about 6 months.
But Microsoft is not alone in these sort of defaults games - every OS or browser maker, Apple, Google, Firefox, wants to create moats so they can more easily monetize your usage of a product. I never thought I’d prefer the business model of free to play games, where they just outright ask you for money and have to keep finding new ways to entertain instead of relying on hard to change defaults and selling your data.
There are bugs that certain games rely on and features that some don’t use. I’m currently trying to optimize a library out of spite. (I want it to work better than the competitor that caused me a lot of problems on a recent project). The amount of conditional logic around what is essentially a function to increment a value is breathtaking.
A simple example would be that the function glGetString(GL_EXTENSIONS) crashes the original Quake engine and many licensees, because it's expecting no more than a 256 character string.
The driver looks to see if a known old game is calling it, and if it's one known to crash, it returns no more than 256 characters, and likely also puts all the _old_ extensions that the game is likely to know and react to in the string.
There are also all sorts of games that called APIs in a particular order or set particular options, because they represented a "fast path" at the time, and now they don't, but if you're that program, then yes they do.
Ultimately, this clutter is what let do the development of the Vulcan API, to stop games second-guessing graphics APIs which themselves second-guess the games.
To avoid doxxing myself: In a deep call stack it’s possible to end up sanitizing inputs multiple times and in different ways.
A frequent example I’ve encountered is web frameworks that have to keep checking for escaped text because they didn’t write it in horizontal layers where you know for sure that all inputs have been scrubbed when they reach this function but not that one. So the same functions get called with data that comes from your team and from customers. Reuse is tricky.
I was pretty young at the time, but I recall the market for graphics being a lot wider open at the time Quake was released. Remember 3dfx? They produced the Voodoo series of graphics cards. They're barely a distant memory now.
Quake was also the standard for a game that was willing to fully exploit the hardware of the time.
That's strange because the cutlass docs explicitly does NOT mention fp8 support. So it looks like it can be used nevertheless with fp8 by using the name hack.
Seems this is likely due to ongoing work on FP8 support on nvidia/cutlass. From my reading, the alternative code path was likely added recently for testing by external contributors to the cutlass project, and other involved parties. (Rather than attempting to distribute custom packaged internal builds of cuda.)
I have small experience with compilers and llvm but youd be shocked how many things rely on names and parsing names
If you have hundreds of passes that are complex and rely on various "contracts" like type names or some shit, then really crazy things like this can happen unintentionally and not maliciously
Some names are standardized items, like memcpy. Matching those is ok, nothing sneaky going on there. Matching something vendor-specific in a general-purpose API is different story.
Ooh, I remember this, but actually the thing is older than it.
First, nVidia and ATI used executable names for detecting games, then they started to add heuristics.
If you think they stopped the practice, you're very mistaken. Every AMD and nVidia driver has game and app specific fixes and optimizations.
nVidia cheated in 3D Mark that way, so they patched/changed their benchmark to prevent it. Also, again nVidia, patched their drivers so some of the more expensive but visually invisible calls like scene flushes in a particular game is batched (e.g. do all 50 flushes at the 50th call) to prevent the game becoming a slide show on expensive hardware.
This is also why AMDs and Intel's open source drivers under Linux a success, because they are vanilla drivers written from scratch per spec, and if your code calls OpenGL/Vulkan to spec, then you're golden.
Even some companies cross compile AMD's Linux drivers for windows on embedded systems since they're free from useless optimizations from them.
Interestingly, most benchmark controversies back in the day are now expected behaviour, i.e. game-specific optimizations with no (well, in this age of upscalers and other lossy optimization techniques, probably even somewhat) visible image degradation. A gaming-specific driver with no game-specific improvements in its changelog would be considered strange, and it very much works with executable detection.
Back in the day, there was still the argument that drivers should not optimize for benchmarks even when visually identical, because it wouldn't show the hardware's real world potential. Kinda cute from today's perspective. :)
But of course there were the obvious cases...
The Quack3 lowering filtering quality as shown above, of course (at least that one was put into the driver as a togglable setting later on).
But the most cheeky one has to be nVidia's 3dmark03 "optimizations", where they blatantly put static clip planes into the scenes so that everything outside the predefined camera path from the benchmark sequence would simply be cut from the scene early (which e.g. fully broke the freelook patched into 3dmark and would generally break any interactive application)
I think that was the first case (to go public), but I remember reading about this in game magazines a couple times after this, for both ATI and nvidia.
Even in the middle of that turmoil, we managed to compile some code with Intel's ICC and make it go faster on AMD Opterons, breaking Intel's own numbers.
When my colleague said that they managed to go faster than intel with icc with some hand tuned parameters, I remember answering "youdidwat?".
Runtime of the compiled code. The ostensible intent is so that new processors can use new features like SIMD, while offering a fallback for older ones. In practice, they’re detecting an Intel processor, not just the specific feature.
Intel's quest to move from "trusted by default / the reference" to "check for scam" is getting worse every release. And it's 100% self inflicted. How weird.
This tweet appears to be taking the original material out of context to misrepresent it:
> Rewrite the attention kernel to be persistent. This gives better performance at low-contexts. However, fp16 at large context has suffered a bit due to a ptxas instruction scheduling issue in the softmax partition. fp8 is ~100 tflops faster when the kernel name has "cutlass" in it.
The charitable reading is that, on certain kernels, using fp8 rather than fp16 values gives better performance. (Although I can't even see how the numbers relate to a "~100 tflops faster" claim in any respect, nor does it even list any kernel names or suggest a control kernel!) But this is being presented as if someone has uncovered evidence of cheating on benchmarks.
The tweet is quoting from the first message in the "conversation" on the PR. There are 93 commits in the PR and GitHub doesn't even default to that tab. I looked at the obvious text and drew the conclusion that was obvious to me.
I already had to deal with Twitter and a link shortening service just to get to GitHub and then it still only pointed to the facing page of a 93-commit PR.
In `libnvidia-nvvm.so` the string `cutlass` appears right after `Memory Dependence Analysis` and `memdep`. Perhaps it acts as an optimization attribute of some sort, where the compiler is allowed to make assumptions about the kernel's behavior that are not valid in general?
yes, that is a very usual way (known practices) of vendors applying specific optimizations for known things.
It is also part of the benchmarks game they play against each other.
The link is long dead and the Wayback machine doesn’t have a copy.
But in 2001 ATI was caught applying optimizations to Quake 3 when someone realized if you renamed the executable from “quake” to “quack” the score dropped a ton. It was a big scandal.
I know that’s common now but that wasn’t a thing that was done at the time.
Was it a scandal at the time? My understanding of how per-game card-driver optimizations work today is:
1. AAAA Game Studio shits out another unoptimized clunker
2. nvidia considers it a reputational risk if games run at 30 FPS on a 5090
3. They go in, look at the perverse ways the game misuses rendering primitives, and then hacks shit in to make whatever bad things they're doing less bad.
As a gamer, this seems fine to me and i generally blame the AAAA devs for being bad at their jobs or AAAA studio leads for being ok shipping unoptimized messes.
> As a gamer, this seems fine to me
As a software developer, it almost certainly has a bad effect on the ecosystem long term. "Hacks shit in" is the very definition of technical debt, and that has a cost that someone, somewhere is going to have to pay in some form.
I can't reply to the person that replied to you, so
> You’re looking as a dev, but the reality is that a consumer cannot see technical debt.
The consumer can't _see_ technical debt, but they sure as heck can be impacted by it.
- Technical debt means the code base is harder to work with later. So fixes/enhancements take longer to make it into the code (and sometimes never can)
- This particular type of technical debt means the code by the game developers sets precedent, and the next developer may us it as an example. So the amount of code incorrectly using the api grows faster over time
For some reason HN sometimes hides the reply button on leaf comments. I think this only happens for very new comments.
You can click the timestamp ("X minutes ago") to view the comment without context, and reply from there.
>the next developer may us it as an example
These hacks are game specific, so another developer wouldn't get them.
You’re looking as a dev, but the reality is that a consumer cannot see technical debt. If the studio churns out a game, the vendor sprinkles on some optimizations, people play it and move on, then the tech debt just vaporizes into the void. It’s not real at that point.
I believe the driver silently swapped the textures to lower quality ones that looked worse but gave a performance boost.
Except that if a developer has that kind of market pull, nVidida will gladly help those devs with getting it right. They are excellent at maintaining developer relations.
it rendered in lower quality, IIRC lower textures / much more aggressive mipmapping and/or LOD
In at least one past version of Windows (circa 1990s), if you tried to replace the default web browser of IE with another choice you were given an Open File dialog window to choose the executable.
Funny quirk, though: that particular window wouldn't show files named firefox.exe. It would accept that as typed input, if you were at the correct folder, but the file listing omitted that particular file.
Maybe it was mozilla.exe; it was a long time ago. But that was the discovery that pushed me off IE forever.
I vaguely remember that being the start of the browser prompts to set your current browser as the default. It was so hard to just configure that they had to build a way to set it within the browser.
You saw that again in more modern times when Microsoft removed support for the APIs they provided to set browser defaults, forcing browser makers to write step by step instructions on what to click to set the default browser.
I believe they walked that back, but it left such a bad taste that I switched my installation of Windows from default mode to EU mode in order to avoid it. And come to think of it, I haven’t used my windows machine for much outside of AI in about 6 months.
But Microsoft is not alone in these sort of defaults games - every OS or browser maker, Apple, Google, Firefox, wants to create moats so they can more easily monetize your usage of a product. I never thought I’d prefer the business model of free to play games, where they just outright ask you for money and have to keep finding new ways to entertain instead of relying on hard to change defaults and selling your data.
There are bugs that certain games rely on and features that some don’t use. I’m currently trying to optimize a library out of spite. (I want it to work better than the competitor that caused me a lot of problems on a recent project). The amount of conditional logic around what is essentially a function to increment a value is breathtaking.
Do you have any kind of example you're able to share? I don't mean to take your IP but I want to see this breathtaking vista.
A simple example would be that the function glGetString(GL_EXTENSIONS) crashes the original Quake engine and many licensees, because it's expecting no more than a 256 character string.
The driver looks to see if a known old game is calling it, and if it's one known to crash, it returns no more than 256 characters, and likely also puts all the _old_ extensions that the game is likely to know and react to in the string.
There are also all sorts of games that called APIs in a particular order or set particular options, because they represented a "fast path" at the time, and now they don't, but if you're that program, then yes they do.
Ultimately, this clutter is what let do the development of the Vulcan API, to stop games second-guessing graphics APIs which themselves second-guess the games.
To avoid doxxing myself: In a deep call stack it’s possible to end up sanitizing inputs multiple times and in different ways.
A frequent example I’ve encountered is web frameworks that have to keep checking for escaped text because they didn’t write it in horizontal layers where you know for sure that all inputs have been scrubbed when they reach this function but not that one. So the same functions get called with data that comes from your team and from customers. Reuse is tricky.
It’s really strange for established companies to waste their credibility on games like that…
Never underestimate how much human ego will control actions.
I was pretty young at the time, but I recall the market for graphics being a lot wider open at the time Quake was released. Remember 3dfx? They produced the Voodoo series of graphics cards. They're barely a distant memory now.
Quake was also the standard for a game that was willing to fully exploit the hardware of the time.
Thats very likely imo
actual link: https://github.com/triton-lang/triton/pull/7298
Thank you, perhaps the parent can be edited to use this URL instead
The Volkswagon emissions testing model
[dead]
So, what is Cutlass, can someone explain whether checking for kernel names makes sense here or is a form of cheating?
https://docs.nvidia.com/cutlass/index.html
Github version: https://github.com/NVIDIA/cutlass
I wonder if we search the comments if we can find something referencing this.
That's strange because the cutlass docs explicitly does NOT mention fp8 support. So it looks like it can be used nevertheless with fp8 by using the name hack.
It supports e5m2 and e4m3 right in the doc linked.
Seems this is likely due to ongoing work on FP8 support on nvidia/cutlass. From my reading, the alternative code path was likely added recently for testing by external contributors to the cutlass project, and other involved parties. (Rather than attempting to distribute custom packaged internal builds of cuda.)
This ticket is a good starting place to see the chain of issues around the ongoing work: https://github.com/NVIDIA/cutlass/pull/2037
I have small experience with compilers and llvm but youd be shocked how many things rely on names and parsing names
If you have hundreds of passes that are complex and rely on various "contracts" like type names or some shit, then really crazy things like this can happen unintentionally and not maliciously
Web-developers are well aware of this too. Sincerely, Mozilla/5.0 (X11; Linux x86_64; rv:139.0) Gecko/20100101 Firefox/139.0
Funny we send a browser wars tombstone in every request!
Let's have a moment of silence for Gecko/20100101
Some names are standardized items, like memcpy. Matching those is ok, nothing sneaky going on there. Matching something vendor-specific in a general-purpose API is different story.
Why would i be shocked that a name is informative. Like... are you surprised that wrought iron is wrought? Or cast iron is made from a cast?
Dog piles are often neither composed of dogs, nor actual piles.
Names can be both informative, and misdirecting, at the same time.
GenuineIntel moment.
Or maybe Quack III: Arena. https://m.slashdot.org/story/21054
Ooh, I remember this, but actually the thing is older than it.
First, nVidia and ATI used executable names for detecting games, then they started to add heuristics.
If you think they stopped the practice, you're very mistaken. Every AMD and nVidia driver has game and app specific fixes and optimizations.
nVidia cheated in 3D Mark that way, so they patched/changed their benchmark to prevent it. Also, again nVidia, patched their drivers so some of the more expensive but visually invisible calls like scene flushes in a particular game is batched (e.g. do all 50 flushes at the 50th call) to prevent the game becoming a slide show on expensive hardware.
This is also why AMDs and Intel's open source drivers under Linux a success, because they are vanilla drivers written from scratch per spec, and if your code calls OpenGL/Vulkan to spec, then you're golden.
Even some companies cross compile AMD's Linux drivers for windows on embedded systems since they're free from useless optimizations from them.
Aah, that brings back memories...
Interestingly, most benchmark controversies back in the day are now expected behaviour, i.e. game-specific optimizations with no (well, in this age of upscalers and other lossy optimization techniques, probably even somewhat) visible image degradation. A gaming-specific driver with no game-specific improvements in its changelog would be considered strange, and it very much works with executable detection.
Back in the day, there was still the argument that drivers should not optimize for benchmarks even when visually identical, because it wouldn't show the hardware's real world potential. Kinda cute from today's perspective. :)
But of course there were the obvious cases...
The Quack3 lowering filtering quality as shown above, of course (at least that one was put into the driver as a togglable setting later on).
But the most cheeky one has to be nVidia's 3dmark03 "optimizations", where they blatantly put static clip planes into the scenes so that everything outside the predefined camera path from the benchmark sequence would simply be cut from the scene early (which e.g. fully broke the freelook patched into 3dmark and would generally break any interactive application)
You beat me to it. Grrr...
Just kidding, nice to see another person who remembers these things. Want some root beer?
Now I want a Quake shooter but with ducks.
Not ducks, but chickens, was very popular in Germany back in the day: https://en.wikipedia.org/wiki/Crazy_Chicken
Oh wow, that was a blast from the past. The Moorhuhn craze!
Many people, including me, didn't have an internet connection back in the day. The Sneakernet went into overdrive so get everyone a copy!
A Duck Hunt, if you will…
I think that was the first case (to go public), but I remember reading about this in game magazines a couple times after this, for both ATI and nvidia.
I'm interested in that story, what are you referring to with "GenuineIntel"?
Intel's C++ compiler is known to add branches in its generated code checking if the CPU is "GenuineIntel" and if not use a worse routine: https://en.wikipedia.org/wiki/Intel_C%2B%2B_Compiler#Support....
Also MKL:
https://danieldk.eu/Intel-MKL-on-AMD-Zen
Even in the middle of that turmoil, we managed to compile some code with Intel's ICC and make it go faster on AMD Opterons, breaking Intel's own numbers.
When my colleague said that they managed to go faster than intel with icc with some hand tuned parameters, I remember answering "youdidwat?".
Good times.
Is this for the runtime of the compiled code or for the compiling machine? Do they generate slow code if the compiler is running on non-intel?
Runtime of the compiled code. The ostensible intent is so that new processors can use new features like SIMD, while offering a fallback for older ones. In practice, they’re detecting an Intel processor, not just the specific feature.
the runtime. patching cpuid makes the code go faster
For the compiled code. Its output deliberately runs slower on non-Intel CPUs.
seems like you don't understand complex hardware.
And what’s the downside of using that kernel name? It can’t just be that it’s faster and nothing else. Unless they included lots of sleep(x) calls.
There might be optimizations that are only safe for the code that this was an intender for.
I wish people either learned how to use git or just wholesale stopped using it.
is 100 tflops a lot?
It's like 5-10% here
Correct, this is the actual headline too. 100 tflops sure seems like it'd be more than that, but here we are.
If the headline was "FB8 is ~7% faster when kernel name has 'cutlass' in it...", it wouldn't seem sensational.
According to Terminator 3 Skynet used a mere 60 TFLOPS
How much is that in jiggawatts per parsec?
5060 ti +~15%
yea
Let's hope for Nvidia this is an innocent optimization only valid for internal kernels that cannot be applied in general.
In which case checking for a string inside arbitrary name is sloppy (a bug).
[dead]
Intel's quest to move from "trusted by default / the reference" to "check for scam" is getting worse every release. And it's 100% self inflicted. How weird.
In my understanding of the PR, it rather seems that it is NVidia is the company that is cheating. :-)
NVIDIA-inflicted in this case.
This tweet appears to be taking the original material out of context to misrepresent it:
> Rewrite the attention kernel to be persistent. This gives better performance at low-contexts. However, fp16 at large context has suffered a bit due to a ptxas instruction scheduling issue in the softmax partition. fp8 is ~100 tflops faster when the kernel name has "cutlass" in it.
The charitable reading is that, on certain kernels, using fp8 rather than fp16 values gives better performance. (Although I can't even see how the numbers relate to a "~100 tflops faster" claim in any respect, nor does it even list any kernel names or suggest a control kernel!) But this is being presented as if someone has uncovered evidence of cheating on benchmarks.
No, that sentence is separate from the rest. Take a look at the pull request:
The tweet is quoting from the first message in the "conversation" on the PR. There are 93 commits in the PR and GitHub doesn't even default to that tab. I looked at the obvious text and drew the conclusion that was obvious to me.
I think you're the one doing that to the tweet, actually.
What are you talking about? When I view the tweet, the only text I see is:
> > fp8 is 100 tflops faster when the kernel name has "cutlass" in it
> kms
https://github.com/triton-lang/triton/pull/7298/commits/a5e2...
It's literally in the code.
I already had to deal with Twitter and a link shortening service just to get to GitHub and then it still only pointed to the facing page of a 93-commit PR.