#programming
1 messages · Page 347 of 1
shadow why’d you drop lilac from your name
i dont think i need a crazy amount of data
Not reasoning but these were a couple local runs I did
god that gradient norm
increase weight decay
also i only have the hardware for small models (8gb or under pretty much) and dont wanna pay for apis so idk if that would work well
good question
that was a 3 epoch runs with a minimum learn rate
oh what
fair enough so prob overfitting
the one spike was a boundary but otherwise within norms for mistral 24b
I was literally thinking like an hour ago "what happened to shadow they just disappeared"
i used to scam togetherai by using the 30$ of credits i had racked up (that also didnt ever go down for some reason) but now u need to pay like 5$ to access the platform at all
lmfao
i dont wanna give any company money if i dont have to
mind you that I was only hitting a train/gradnorm of 1 at over step 800 which was 36 hours in
36h
i hate how llms take so long to train
i trained the 1.7B on the distil on dgx spark and it took 24h for only 15% of the data
that was a 24B model. One group I work with is currently messing with 123B and Command_A
god
for me doing 1 epoch on my chat dataset with a 7b it takes 3 hours
for lora tho
identity crisis? 
my sister said my username was cringe
lmao
my dataset was something like 28mb of raw text
your sister is cringe 
yeah she is also lol

hi VamManSaele
mine is 12.5mb
mine is uhhhh
.parquet?
parquet
apache parquet ye 
mf got that wooden flooring file exstenssion
open source version of googles sstable
i just use sharegpt format in jsonl cuz unsloth supports it really well
can hf dataset load jsonl natively
idk
ok genuine question do data science/ml people ever just use sqlite
it’s so much better 
im too much of a noob to answer, but i dont even know what that is lol
parquet is columnar storage
better for analysis/whatever
true didn’t think about that 
parquet with duckdb tho 🤤
just use clickhouse instance

ye also can't really modify parquet files AFAIK
need something like DuckDB 
@olive sable 
user@nixos ~/c/vrrr [SIGINT]> wc -l triangle.c
131 triangle.c
ye
is the print always in the renderthread?
i thought we moved that to a different thread

131 line of triangle
not need engine
vulkan needs at least 1K lines

user@nixos ~/c/vrrr> ./result
Gtk-Message: 20:49:16.885: Failed to load module "colorreload-gtk-module"
AMD Ryzen 5 7600 6-Core Processor (RADV RAPHAEL_MENDOCINO)
FPS: 42241
FPS: 43228
FPS: 43883
FPS: 43324
real fps number should be closer to 43k for me
discover 
sam synchronisation is incorrect
while (1) {
vkQueueSubmit(queue, 1, &(VkSubmitInfo){ .sType = VK_STRUCTURE_TYPE_SUBMIT_INFO, .commandBufferCount = 1, .pCommandBuffers = &cb }, 0);
vkQueuePresentKHR(queue, &(VkPresentInfoKHR){ .sType = VK_STRUCTURE_TYPE_PRESENT_INFO_KHR, .swapchainCount = 1, .pSwapchains = &swc, .pImageIndices = &(uint32_t){0} });
vkQueueWaitIdle(queue);
double currentTime = glfwGetTime();
frameCount++;
if (currentTime - lastTime >= 1.0) {
printf("FPS: %d\n", frameCount);
frameCount = 0;
lastTime = currentTime;
}
}
see here, if i comment out vkQueueWaitIdle(queue); it go to 106-108k fps
vulkan 🥀
if i have it uncommented, i remain at 43k
because waiting for frame to be rendered
rather than just spamming frames
we're not waiting for the frame to be rendered tho
wait
triangle 2
vkQueueWaitIdle wait until eevrythiogn is done, what we're doing is keeping the queue full so it always has work
ye obviously not work for sam, just use gpu semaphore instead 
// triangle.frag
#version 450 core
layout(location = 0) out vec4 outColor;
void main() { outColor = vec4(1.0, 0.0, 0.0, 1.0); }
// triangle.vert
#version 450 core
const vec4 positions[3] = vec4[3](
vec4( 0.0, -0.5, 0.0, 1.0),
vec4( 0.5, 0.5, 0.0, 1.0),
vec4(-0.5, 0.5, 0.0, 1.0)
);
void main() { gl_Position = positions[gl_VertexIndex]; }
triangle shader
we are
just listen
once the queue gets full
vkQueueSubmit starts blocking
so we should be good
depends on drivers tho since this is technically undefined by vulkan and left to the drivers to solve
it might aswell jsut be deleting drawcalls that dont fit in the queue for all we know
it is true that the fps number largely just shows cpu throughput tho
i think the assumption was that with 1 triangle we shouldn't be gpu bound anyways, and i still believe that is largely true.
when we made that decision we were still at 13K fps


i give up
i did type that 

just som?
ye

Sam trpongle (why did I use brainpower to merge triangle and typo?)
by the way
fps with glfw3 is higher than sdl3
for whenever you do your next rewrite

i was using sdl cuz of it being webgl capable
but for vulkan i might switch to glfw then
actually send the file, i want to see
ok
dont laugh though
here's the default.nix to build it with
{ pkgs ? import <nixpkgs> {} }: pkgs.gccStdenv.mkDerivation rec {
pname = "vrrr";
version = "0.1.0";
src = ./.;
nativeBuildInputs = with pkgs; [ shaderc ];
buildInputs = with pkgs; [ vulkan-loader vulkan-headers glfw ];
buildPhase = ''
glslc -O $args triangle.vert -o vert.spv
glslc -O $args triangle.frag -o frag.spv
cc triangle.c \
-Ofast -march=native -mtune=native \
-lglfw -lvulkan -fuse-ld=gold \
-flto -fomit-frame-pointer -s -fno-plt \
-o $out
'';
env.NIX_ENFORCE_NO_NATIVE = 0;
}
you have the shaders already

i might look at this wednesday afternoon
im too mentally drained to do vulkan rn

all
well yes, but that isn't very descriptive you see
VkPhysicalDevice pdev;
vkEnumeratePhysicalDevices(inst, &(uint32_t){1}, &pdev);
VkPhysicalDeviceProperties props;
vkGetPhysicalDeviceProperties(pdev, &props);
printf("%s\n", props.deviceName);
float prio = 1.0f;
VkDevice dev;
vkCreateDevice(pdev, &(VkDeviceCreateInfo){
.sType = VK_STRUCTURE_TYPE_DEVICE_CREATE_INFO, .queueCreateInfoCount = 1,
.pQueueCreateInfos = &(VkDeviceQueueCreateInfo){ .sType = VK_STRUCTURE_TYPE_DEVICE_QUEUE_CREATE_INFO, .queueCount = 1, .pQueuePriorities = &prio },
.enabledExtensionCount = 1, .ppEnabledExtensionNames = (const char*[]){ "VK_KHR_swapchain" }
}, NULL, &dev);
VkQueue queue;
vkGetDeviceQueue(dev, 0, 0, &queue);
more descriptive
want more?
is that what you removed or?
okay
VkSwapchainKHR swc;
vkCreateSwapchainKHR(dev, &(VkSwapchainCreateInfoKHR){
.sType = VK_STRUCTURE_TYPE_SWAPCHAIN_CREATE_INFO_KHR, .surface = surf,
.surface = surf, .minImageCount = 1, .imageFormat = VK_FORMAT_B8G8R8A8_UNORM,
.imageColorSpace = VK_COLOR_SPACE_SRGB_NONLINEAR_KHR, .imageExtent = { WIDTH, HEIGHT },
.imageArrayLayers = 1, .imageUsage = VK_IMAGE_USAGE_COLOR_ATTACHMENT_BIT,
.imageSharingMode = VK_SHARING_MODE_EXCLUSIVE, .preTransform = VK_SURFACE_TRANSFORM_IDENTITY_BIT_KHR, .compositeAlpha = VK_COMPOSITE_ALPHA_OPAQUE_BIT_KHR,
.presentMode = VK_PRESENT_MODE_SHARED_CONTINUOUS_REFRESH_KHR, .clipped = VK_TRUE
}, NULL, &swc);
VkImage img;
vkGetSwapchainImagesKHR(dev, swc, &(uint32_t){1}, &img);
here's the
swapchain
vulkan is so verbose
explod
konii write
no wonder low line count
despair
i tried but i couldnt stand the line length
when this was all one line
ping me again wendesday afternoon for this and i might have some semblance of what is happening
.pColorBlendState = &(VkPipelineColorBlendStateCreateInfo){ .sType = VK_STRUCTURE_TYPE_PIPELINE_COLOR_BLEND_STATE_CREATE_INFO, .attachmentCount = 1, .pAttachments = &(VkPipelineColorBlendAttachmentState){ .colorWriteMask = 0xF } },
:mlntcandy:
it not thaaaat long
#define GLFW_INCLUDE_VULKAN
#include <GLFW/glfw3.h>
#include <stdio.h>
const uint8_t vert_spv[] = {
#embed "vert.spv"
};
const uint8_t frag_spv[] = {
#embed "frag.spv"
};
#define WIDTH 300
#define HEIGHT 200
void main() {
glfwInit();
glfwWindowHint(GLFW_CLIENT_API, GLFW_NO_API);
GLFWwindow* w = glfwCreateWindow(WIDTH, HEIGHT, "vk", NULL, NULL);
uint32_t ext_count;
const char** ext = glfwGetRequiredInstanceExtensions(&ext_count);
VkInstance inst;
vkCreateInstance(&(VkInstanceCreateInfo){
.sType = VK_STRUCTURE_TYPE_INSTANCE_CREATE_INFO,
.enabledExtensionCount = ext_count, .ppEnabledExtensionNames = ext,
}, NULL, &inst);
VkSurfaceKHR surf;
glfwCreateWindowSurface(inst, w, NULL, &surf);
...
no no come back
look how easy simple vulkan is
the beginning of the file is misleading
im a broken washed up man breaking down over his blender file
christmas tree
my status as the vulkan guy has been revoked


trangl
there's something wrong with summary in this channel
it have brain damage


puding
it learning
elvspeak

microwave
you cant stop me

Question about memory and data types.
As a C# dev, I don't really give too much thought to how much memory my apps use, so I'll use ints and doubles everywhere. Then this makes me think of how many modern apps are RAM hogs and I wonder if this is all a death-by-thousand-cuts thing
Does it actually matter these days in a memory managed language? I could reduce my footprint a bit, but I may only be saving pennies-worth if I'm building off of other libraries or frameworks or webviews
Use pool
It depends on your data structure. but if you have an estimation on memory consumption then you can work around that.
For instance, if you are saving 1 million double but you can store them in int. Then you just reduced your memory consumption by half.
But, it always depends on much data are you working with. If you are dealing with 100 double. Well, you probably don't have to worry about anything
And if you are using object to store intermediate data, then you probably should use pool to reuse object. You will save on GC and increase on performance.
Do you have any articles or links where I can learn more about this? The extent of my knowledge of pooling is seeing a tourism poster
Actually, now that I think about it, I've already used that concept without realizing it...
Virtual UI lists that only load visible-on-screen items (+ buffer) are a pool of UI objects
Thank you, I'll go read this
Im throwing an entire DB into ram abd can be stupid lazy about ram usage
It always depends on your application and input data and performance requirement.
small data -> more fits into cache -> higher performance 
ymmv
Mind you that won't save you from bad memory access patterns
Related to that, I've recently watched videos on struct-of-arrays vs array-of-structs, and it all sounds good, but it's always through the lens of lower level languages, so makes me wonder if I should care as much in a managed language
Though I guess like you said, if there's less data to process, better chance of cache hits (probably no matter the language)
Got any examples handy?
If you can't come up with a good data structure, then use dictionary/hashmap!
Think about big O
Ö
Nothing specifically but random = bad (this includes hash tables)
Is that where RAM gets its name? Keep the relevant "deterministic" stuff in cache, let RAM be the larger free-for-all?
Been a long time since I've looked into it, isn't it only for time complexity?
time complexity = performance
When you're google
there is memory complexity and performance complexity right?
or how are they called correctly
Time and space yes
oh right space
oh right that too
I should go back and refresh on big O tbh, been too long
O(1) , <=O(n) and else!
I have questions and nobody in #general-network is answering me so I will ask here despite not being super related
So does Ellie often do electronics build streams or is it just for special occasions like the Neuro dog and Neuro car
I think she does tech streams not very often
she streamed building it?
There was a whole thing where she never revealed she was an engineer until the vedal raid IIRC
Really
it was streamed yeah
at least some parts of it IIRC
At the moment, not much as over 60+ hour work weeks drain a bit of energy
Google DeepMind 🤝 @BostonDynamics
Our new research partnership will bring together our advancements in Gemini Robotics’s foundational capabilities to their new Atlas® humanoids. 🦾
Find out more → https://t.co/Z4fL9ixjW3
60+ hour? that's crazy
Makes sense
Only stream she does is the balding streams
I really admire her streaming as a hobby, I work less than 40hr/week programming and feel like I can't even do laundry without great effort once I'm home

I dont work and i cant do laundry 
is ellie employed in japan
No, usa iirc
Maybe in weekends too?
@sick owl dont let subathon distract us from the fact that the holidays are over 
oooo
Gemma 4?
That'd be nice
last balding streams revealed that she had another offer for more cash, so I pray it works out 
what is a "balding stream"?
3.0 flash lite perhaps
Chill group stream with a few other faces. Mini, Shoomni, and Chuchie (forgive my misspellings)

the last few times when cerber raided they joked about shaving the dawg
one day
I make it a point to watch those streams personally, sometimes it gets clipped from staz for neuro stuff but it is just fun.
Neuro-robot
I'd take a flat 3d print over the circle screen sure
nvm
spotted

Does anyone have experience running vLLM server on WSL?
Does this mean us regular devs aint getting access to the Gemini VLA any time soon though 
I'm running vllm serve [my_model] on WSL. Engine initializes correctly.
I can see VLLM::EngineCore process in top and the model is taking the right amount of VRAM
But when I curl, it always fails. I'm suspecting the server isn't alive
so far nothing I found online solves this 
that makes sense
i've never had issues with ollama over wsl
must be some weird WSL setting
yeah, just tried ollama, it let hit the WSL endpoint without any issues from a non-WSL terminal
weasel
in WSL settings, check if you have enabled localhost forwarding enabled
When i have problems with things connecting to things i try disabling all firewalls
Only when everything is local or course but i wonder what would i risk if you do that while connected to the internet
what can happen?
surely there is no terminal access server running in my computer? I didnt launch anything like that
as an aside, anyone had to reset their twitch 2fa? I tried to reauth a device and 2fa code kept failing and I had to use the sms backup option
Genuinely interested, don't you need some things running with open ports for these ports to work?
A port must lead somewhere
And if there is nothing its useless
i saw some weird stuff happening with payout info being updated without creators knowledge so I'm wondering if their auth got compromised somehow, or a an app/connection was
could be for downloading models or for listening for the API it exposes if it's meant for something else to connect to
shouldn't disable all firewalls, should just add permission to the app which requires it
i think my api server just never started 
usually there should be a line of log from uvicorn
Yes if its wsl2, You need to set it to a static IP and create a port forwarding map to windows
Something like this:
netsh interface portproxy add v4tov4 listenport=xxxx listenaddress=0.0.0.0 connectport=xxxx connectaddress=$wslIp
I just realised they gave atlas wide hips and totally different limb joints now
In general the new model looks a bit cheaper, safer, sturdier and more suited for mass production
At the cost of losing style points
Like the left side head looks better but the one on the right is definitely gonna be more durable and practical for real world use
Ah yeah as I suspected, its a more production ready model optimised for cost and simplicity/durability
Oh like actually, its for sale wtf
We’re excited to announce the product version of our Atlas® robot. This enterprise-grade humanoid robot offers impressive strength and range of motion, precise manipulation, and intelligent adaptability—designed to power the new industrial revolution.
Learn more about Atlas: https://bosdyn.co/4jrAuRG
To learn more about our journey with A...
huh the right one seem less durable imo
though right one seem more dynamic
but i feel like in terms of tension stress its worse mechanically
though depending on implementation it does seem cheaper.
though the hand on the right one do seem better designed
Excuse me, but the Chinese humanoid robots are on a completely different level... especially when it comes to customization.
Extremely ugly, those robots up there.
"we've got humanoid robots at home"
wrrr
wrrr
plays fart-echo sound through built-in speaker
?
don't get me wrong, it would be funny. But there are so many better things to use
I propose dark souls music
of course only activated once it notices you
if it doesn't twist it's neck 180 degrees to look at you while playing dark souls boss music, is there even a point to the tech?
it's like a level in a puzzle game with time travel.
the one on the left looks like a well designed body with simple hands added on as an afterthought
the one on the right looks like well designed hands with a simple body added on as an afterthought


Based
ert
we can make next gen triangle have semaphores and fences if you want
its more upto spec
but it is a lot slower
Been playing alot with Vulkan but I might also try WebGPU at some point given they are both really popular atm
Its also more widely supported
you mean Metal?
Vulkan has MoltenVK too
hardware support should be basically identical for Vulkan and WebGPU
Still willing to give it a look as I heard there is less boiler
ye there's a little bit less IIRC
I feel like both are good contenders for replacing the current standards
definitely
though arguably both are already starting to be outdated for modern GPUs 
at least the Web will finally have compute shaders once WebGPU is available everywhere
Once NVIDIA implements their new production decrease we won't have to worry too much about that for a while
JK
It never died it just got repurposed for at home computing
RGB gaming DIMMs, famously always deployed inside fully opaque servers
badly made meme 
ye but they don't know that
so they use the boring ones with the PCB exposed
Thats why they keep buying more ram not knowing the RGB lights would speed up their current

wrr
wrrr
wrrrr
@olive sable
wrrr
ye wrrr

Gets rid of the speed in RAM



surprisingly the latency is not that bad
something like 10x worse than DDR5 IIRC? not horrible
tons of shared memory
Real cloud computing then
can just access the same memory -> no need to directly synchronize between hosts 
Sounds good in theory
become banana 🍌
insufficent color in pfp 
me not have banana in pfp
anyone can become banana 
you have color almost same as banana in pfp
pfp can become banana too 
but ye not necessary

hi ylhfecaa
i shall 🍌

ive cursed myself to white 
maybe try switching to white and back?
banana in spirit 


Conductor we have a problem
mto frend 
you probably can with the right printer 
will maybe even store an entire bit
Lmao
no me want 2 color 
Fully mechanical ram
ground control to major mito
we have a two coloured


mto make white-banana gradient


My steam is bugged but ive kept it forever
mto away 
user@nixos ~/c/vrrr> wc -l triangle.c
244 triangle.c
bwaa
no longer horizontal monster
also i added frames in flight

banana
#define FRAMES_IN_FLIGHT 3
#if FRAMES_IN_FLIGHT == 1
#error "broked"
#endif

unless c 

so me fault with alignas
alignas(uint32_t) const uint8_t vertCode[] = {
#embed "vert.spv"
};
alignas(uint32_t) const uint8_t fragCode[] = {
#embed "frag.spv"
};
it was supposed to be type
gcc just let me do integer
because gcc
babna nadog
fetching path input 'path:/nix/store/as8ksm9hafq3x3cvml95i5s092903kvk-source'
this derivation will be built:
/nix/store/4drhrzak7301izsgi6yjqjzzhl69bmf7-vrrr-0.1.0.drv
building '/nix/store/4drhrzak7301izsgi6yjqjzzhl69bmf7-vrrr-0.1.0.drv'...
Running phase: unpackPhase
unpacking source archive /nix/store/7svs6vgb1mk87dz3biclfnlzbq9yrbpq-vrrr
source root is vrrr
Running phase: patchPhase
Running phase: updateAutotoolsGnuConfigScriptsPhase
Running phase: configurePhase
no configure script, doing nothing
Running phase: buildPhase
Running phase: installPhase
no Makefile or custom installPhase, doing nothing
Running phase: fixupPhase
shrinking RPATHs of ELF executables and libraries in /nix/store/m97ap218gn2xpc4m9spa8fs8k6rz3m68-vrrr-0.1.0
shrinking /nix/store/m97ap218gn2xpc4m9spa8fs8k6rz3m68-vrrr-0.1.0
patchelf: getting info about '/nix/store/m97ap218gn2xpc4m9spa8fs8k6rz3m68-vrrr-0.1.0/lib/libglfw.so.3': Not a directory
checking for references to /build/ in /nix/store/m97ap218gn2xpc4m9spa8fs8k6rz3m68-vrrr-0.1.0...
patching script interpreter paths in /nix/store/m97ap218gn2xpc4m9spa8fs8k6rz3m68-vrrr-0.1.0
/nix/store/m97ap218gn2xpc4m9spa8fs8k6rz3m68-vrrr-0.1.0
source root is vrrr
neuro not noted
72302
70253
69890
70545
71034
71136
frog nog
i thought removing that synchronisation would put me up to 100k fps
only up to 72k
i guess there's something else blocking
Are u using dynamic rendering
probably vkAcquireNextImageKHR
Or render passes still
render passes
Give dynamic rendering a try at some point
Its the new standard and it's less of a pain
Yes
The only thing holding it back for a while was mobile support but its far better now
U dont need render passes or frame buffer objects anymore
ハクセキレイ(セキ助とモミ子)の雛の一日。
父さんご用達の居心地の良いベランダで一日を過ごすおチビさん😊
※雛の入ってる巣は巣立った巣では無くセキ助が去年巣立った古巣です。
#Wagtail
#ハクセキレイ
wagtail
ye
wtf the alarm application on windows just started working again for no reason???
Windows has an alarm application? 
ye
this shit stops working and removes itself from the taskbar one day for no reason so i thought it was gone
idk who uses the alarm app
and now, a week or two later, i get a notification from it again
i am on 2 days of uptime right now and that is a daily alarm that went off
WHAT CHANGED
windows is so weird about things like this
the alarm went on a christmas holiday 
i am so confused
ain't no way
right
"I need to wake up at 7am today I better setup my alarm on my Windows 11™️ Alarm app whilst keeping my computer up and running all night and hope that Windows Update doesn't restart the computer and automatically stop the alarm from working because I'm not logged in"
if i want to get a reminder for something while i'm at my computer it's convenient
I just opened the Clock app for the first time
it had to update 
egg yellow 
banana 
more convenient than getting up and suffocating the life out of one of the buttons on the alarm clock to get it to go from 6am to 7pm or whatever
how
a traditional AA battery powered clock that doesn't connect to the internet and just rings the bell super loudly
do you have a physical one
i've had a physical alarm clock for the last 15 years or so, yes
based
it's got automatic daylight savings rollovers somehow this thing has been great
t if one day everything will be heavily optimized for avx-512 and it will be a sluggish hell without it, know that it will have been you who set the chain reaction in motion for my demise

sturdy and doesn't care about updates
the devil can control this with its power
they fear win11 tho
and I can fix it
i will sleep in peace knowing i have brought a terrible fate to yet another person

i still can't believe they took down alarms for 2 weeks just to rename it to clock
what the fuck are they doing over there
clocks apparently
they are explaining to their manager how they're agilely transforming the business through disruptive copilot technique integration into their mindset workflow
like 6 hours a day
lmao
clocks
⏲️
smashing this one with a comically large hammer must feel otherworldly
my condolences
in this market? 
before the market gets even worse
or we get invaded by big bear country
probably both
it's pre built
i built it on a build site separately, was 35% more expensive on the build site
so it's like a neutral deal
yeah only decent deals i've seen atm are all prebuilts kek
cuz they ordered the ram before the high probs
i was looking yesterday actually at the ram i've got in my pc
tripled in value since i bought it in 2024
unironically great investment
I picked up a vfd based nixie style clock thats great. Reminds me of vha clocks but way larger and brighter
if it doesn't normalize i'll be like
haha all you peasants waited for too long
and if it does normalize i'll be like
yay i can buy another kit

I really dont think neuro and evil have think instruction sets and it makes buckshot a slog
i think its supposed to have less cpu overhead for complex stuff
cpu saturate anyway
renderpass is more efficient for tile-based mobile rendering
which is why i initialy stuck with it
im sure there is a specific combination that is faster than we have it rn, but it would get to a point of being very platform specific
Chinese bots look prettier/appear more impressive but are on the whole less practical than what Boston Dynamics and Figure are doing
Generally speaking anyway
Unitree are really cool though
oh iirc its faster when recording the command buffer, but we dont re-record the command buffer at all since triangle
so for this it doesnt have any effect ye
Mainly for bringing costs down ridiculously low
VK_KHR_dynamic_rendering_local_read fixes this issue supposedly
But most are impractical for production, whereas the Boston Dynamics bots are being deployed en masse into Hyundai factories and Figure bots already proved their worth in BMW fabs
I guess I haven't been keeping up, I didn't know BD managed to land a big contract like that
oh it was literally yesterday
Yeah the cheaper non prototype Atlas is moving into mass production
so they're finally commercializing a humanoid model
Yup, though the full supply for next year is exclusively contracted for Hyundai and Deepmind
was Tesla actually a serious contender?
oh, right, hyundai owns boston dynamics
so it's not quite landing a major contract "organically"
china tech is actually very ahead
partly because their tech industries really geared toward thing for consumer
which is good because research fund is used for thing that improve their life
Ah, the Infamous 3rd Generation of the Intel GMA Series (GMA 900/950/3100/3150). "DirectX 9" Support (Read: Partial, Realistically the most you're getting is DirectX 8.1), OpenGL 2.1 Support (Read: Only on linux, and at the point that i'm writing this, its "support" so deteriorated that you might as well use llvmpipe instead - Realistically Speaking, you're Getting GL 1.4... if you want to use Hardware Acceleration.)
That said, However... if you're doing it just for the sake of a retro gaming PC? most DX 8.1 games should run okay on it. it should be usable enough for a Early Windows XP Gaming Rig. The VGA Output of these iGPUs is reportedly pretty clean. Throw it to a VGA 5:4/4:3 CRT/LCD Display and it should look pretty good.
though games after 2005 on those iGPUs are very much a "Graphics Decelerator"
more rambles about random pieces of technology whenever i feel like doing them
i actually did try running minecraft on that hardware. on Linux - i could actually get hardware acceleration up to 1.14.4 on those iGPUs. 1.15 refused to boot, however.
i forgot the uppermost limit on windows, soz
I'm convinved my PC is haunted, I started trying to look at logs and then I get 2 back to back crashes during it
With like less than an hour between the crashes
Wasn't it already haunted months ago
Ask for tech priest
Well now it's extra haunted
Ghost demand ram
And obviously the logs don't actually have anything useful
what did he mean by "she can change herself"
@olive sable
the 76 line monstrosity
im done
never touching vulkan again
bye
i completed vulkan
oh sorry it actually wordwrapped
did you notice

vulkan 2 when

Well, I rebuilt myself a little PC with the parts I had left after my burglary on December 23rd. The worst part is the GPU and all my lost work... 3 months of work down the drain... sniff
anyway
76 lines, 95k fps
i broke synchronisation a little bit but
it doesnt matter
i get 105k fps when i remove sync entirely though
thats with 3 frames in flight though, at 300x200
At worst, you add a resynchronization loop every X seconds
im only using 1 semaphore to track the frames in flight

i should be using at least two
but
cost lines
and fps
i never even looked at the code, i have exams and dont really have the energy for vulkan
39727 fps with llvmpipe
at 300x200
at 7x6, its 61941 fps
i have to clean up this code to make it presentable to send
but
later
I saw a funny post on Threads
Related to ASI
But I think the original poster is likely out of their mind
So did Vedal killed Neuro and replaced her core with fake one paid commercial ai?Or he just asked about personality because of something else that's deeply concerns me. Guestion to you guys maybe you know some insider info?
it's later already
man i need to refresh myself on DSA
been doing game dev for so many years now that ive lost some of the deeper comprehensions i used to have for areas such as ML and some other lower level things
WHAT THE HELL WHAT THIS IS?!
I think I've found that post
It's just AI nonsense and people with severe delusions
Weirdness
reminds me of the github repo of the "universe simulator" like 50 random files of python files generated by chatgpt that "simulate the universe 1:1"
"quantum error correction" points to it being silliness.
LoL
It's says to dd if=/dev/random of=/dev/sda bs=4M
ye youshould
didn't work 
For once, there is a staz clip that released on this. Confirming that isn't the case but also beginning a new wave of misinfomation from people who can't understand the explination.
Man, LLVMpipe is fast.
39k fps! Woah!
also audio nerds
NVIDIA GPUs be damned
if I want a cheap set of speakers with the sole intention of being better than my monitor (screen) speakers, where should I look
The bar isn't that high for what I know, so almost anything dedicated. (not audio nerd)
actual answer is anywhere
i doubt it gets worse than built in display speakers 
ok, will look at the bottle opener on my desk
is that the one you were flashing to replace the audio
it plays audio use that 
its a single triangle, id hope its fast
If I recall correctly, doesn't the current triangle code seperate sending data to the vulkan shader off main thread or am I vastly misremebering?
it doesnt send data to the shader at all
the benchmark one at least
if you mean the production one. i think its split up ye
Makes sense
I think I have made progress with my little alam setup so that it now accepts user-configurable stuff such as...
Nevermind I just remembered a lot is still missing HAYIaaa
I have no mouth and I must scream
wrrr
hello leader

why am i leading another cult
there are no boundaries of banana
another 

Tackling this as a CPU issue, it feels like the best way for triangle would to be offloading generating the frame to elsewhere and let the presenting thread persent. Issue of course being random concept called race condition. 
nah
concert
without cpu synchronisation you dont get an accurate fps number
Are you speaking about just using a software raster engine?
thats only partialy true since it depends on driver behaviour on if it blocks at the drawcall when the queue is full, but it should
no
Isn't there a way to measure the flipping of framebuffers?
if it blocks at the submitting to the queue it will be synced
provided you are the one doing the flipping, yes
Like... If a buffer flips, then that's 1 frame
unfortunately for performance we dont do that
anyway, vulkan isnt necessarily double buffered
What else could be doing the flipping?
the gpu 
But like... Isn't that done through a command?
if you want to wait for the gpu to stop doing stuff, you can do a semaphore or use vkQueueWaitEmpty or something
yes but there's a whole motherboard and pcie bus between the cpu and gpu

we are on the cpu and the flipping happens in gpu land. to accurately get fps we need t monitor the gpu which actually costs perofrmance which is why we have foregone the syncing
Yeah.
That bus is fast though
also, the cpu can queue commands a lot faster than
the bus
or the gpu can render, typically
no not really
connect them with copper wire
if you measure the command queue you will get a hugely inflated number
if you wait with proper synchronisation you will get a number below the real value
there's no perfect way of measuring fps
You only have to flip a buffer like 60 times, at max.
The PC bus is plenty fast, with very little latency.
You guys don't know what you are talking about, and it shows.
brother
limit it to 24 eyes can't see more
there is a difference between what gets rendered and what gets diapalyed
Of course buffers flip faster than that in practice
Brother, do you need to see the repo and perhaps try it yourself
when you look at your minecraft fps you arent looking at 600 and going "ye my monitor runs at 600hz" 
also, vulkan doesnt do "buffer flipping"
I've actually worked with GPU buffers.
whatever
Do me a favor at put aside your ego and accept this isn't a normal case
bro needs to work with manners next lmao

this discussion is pointless.
anyways for our purposes we need the fps of how many frames get rendered. otherwise i can just say im at 180fps the whole time
https://github.com/samvanmaele/vulkan/tree/main @severe path throw your optics over here if you really want to judge, might as well have code to judge
im being used as an example
Konii is in the middle of a rewrite 
ex-sam-ple
[that code is not up to vulkan spec cuz its made to get the fps number as high as possible]
there are three metrics for fps
- how many times a second you queue the full pipeline (on the cpu)
- how many times a second you render the full pipeline (on the gpu) (alternatively, you could call it how often the gpu presents)
- how many times the compositor decides to update the image displayed on your screen (and the various driver overheads)

only useful ones are the first two
well, third one is useful if you care about what it looks like 
thankfully its a red triangle

not necessarily, the compositor can decide to present frames as often as they arrive if it wants to
whether or not the display actually renders all of those frames is different
"renders"
dawg
Vukan still hides some low level stuff from the user, such as mode setting (which is done by the compositor).
SDL is hiding the communication and setup taking place between your program, and the compositor.
how is modesetting even remotely relevant here 
and sdl is just the vessel for the window creation
Fine, but that is the domain of the problem we are solving. The point of the challenge is to use it :mug:
it doesnt actually play any part in, well, anything to do with rendering except handing off to the compositor
The compositor has to modeset its own frame buffer, before your triangle can be displayed on top of it.



You guys should go and figure this out yourselves.
How would an application display a triangle on Wayland?
Question for Sam's repo. Anyone have the nix shell for building it? I lost it in a drive wipe
this is either godly ragebait or you just have a hole in your head

i do 
{ pkgs ? import <nixpkgs> {} }: pkgs.gccStdenv.mkDerivation rec {
name = ":small_red_triangle:";
nativeBuildInputs = with pkgs; [
lld
clang
];
buildInputs = with pkgs; [
glm
sdl3
vulkan-loader
vulkan-headers
gamemode
];
env.LD_LIBRARY_PATH = pkgs.lib.makeLibraryPath buildInputs;
env.NIX_ENFORCE_NO_NATIVE = 0;
}
Because sam is on nix we should upstream it 
No, I'm being serious.
Vulkan renders a triangle. That has to be sent to the compositor, somehow.
How?
Once you can answer that, you will at least know enough to understand somewhat what I'm trying to say.
What does dmabuf do?
add flake
nix run github:samvanmaele/vulkan
I already know the answer.
I'm guiding you
No one needs guiding here
this topic is being revisited from over 5 months ago
when every person in chat was pitching in for results
I tried my hand at the code as well, didn't do much but it was fun
Vulkan isn't magic... You are treating it like it is magic. Your lack of understanding is what's holding you back.

Shove my words into an LLM checker, then. See what it says.
we all know vulkan isnt magic, ive been writing shaders for years
Then try your own hand at making something that runs faster, providided it is in the rules of using vulkan
I dare you.
if vulkan was magic it would be seriously less frustrating
i double dare you to shut up and take a walk before you make yourself sound more stupid than you already do 
cpu does the drawcall
cpu does present call
driver hands the rendered image to the compositor
compositro decides what to do with it
compositor makes the final image that needs to be sent to the screen
image gets displayed on screen
what frames exactly get sent are dependent on the vulkan settings and modestting, but that has no influence on how many frames actually got rendered
Clearly the "master of gpu" can use the underlying parts of vulkan to beat the current rewrite under the contest rules. @severe path I'm not taking you seriously until you take a stab at it.
there's so much vulkan boilerplate around that even a vibecoder could write a vulkan triangle
probably
definitely
But what do you call a rendered frame, exactly?
Where in this process does something become fully rendered?
well, you see, i call them rendered frames, or in other words, jsut frames

But like... when is the frame finished rendering?
wasn't that the OSX port by Vani actually? I recall it being vibecoded to work on Mac
Is it when a buffer is flipped? Is it when a semaphore is triggered?
"buffer flip"
when the render pipeline is finsihed
Wait until the meta of not using a GPU at all is revealed
i present, the render pipeline.
our lord and saviour in the holy year of 2026
hello smart swarm people im a bit doopid and idk waht im doing but im tryna learn to code a discord bot right but uhh idk waht i did wrong but the packages wont install and idk waht to do

language and package name
How do you determine when the render pipeline is finished?
I wonder if you can get a faster frame if you use mesh shaders instead
right so im using python and pycharm
Your unfortantly picked a bad time 
you can, depends on usecase tho
im using the discord.py and python-dotenv
oh
jetbrains superiority 


year of robotics










