Ryan Flaherty

round 2

Previously Thekla/Google.

[Ryan Flaherty @rflaherty71]

Assassin's Creed Valhalla - 170 GB, hour+ long download on Steam at 120 Mbit/s

With Via you are in game in <5 min with <5 gigs of data on disk.

Ubisoft's Anvil engine tech seems pretty good to me, good loading code and no fps stutters due to streaming.

[Ryan Flaherty @rflaherty71]

@AgileJebrim @dubya_brian @BadalK99277 No one actually uses the macos part, they all just open a web browser and sometimes ssh to a Linux machine

[Ryan Flaherty @rflaherty71]

Looks like I was a couple years too early when I tried joining. Leadership and priorities must have changed a lot.

[Elon Musk @elonmusk]

SpaceX has almost finished writing V1.0 of an in-house AI training stack in C that exact-maps to 220k GB300s with 800G NICs, making heavy use of pipeline parallelism and getting as close to bare metal as possible.

The potential speed improvement vs JAX for large training runs is over an order of magnitude.

[Ryan Flaherty @rflaherty71]

@SheriefFYI Your math assumes synchronous one at a time compute, 100k isn't that big of a number on multicore+simd / the GPU.

[Ryan Flaherty @rflaherty71]

@MarnerJoshua Yes and you are the reason we are running software from the 90's

[Ryan Flaherty @rflaherty71]

You literally can't fix the code. The company has to be willing to cut and start over for a new release version, which none of them are. They'd prefer to waste their engineer's time on extremely expensive incremental tiny nudges on the existing system.

[Kelly Sommers @kellabyte]

I have to be honest, this doesn’t bring confidence to Windows users that your solution to making key Windows interactions faster by pre-boosting the CPU clock on key interactions the user expects to be snappy.

Can we fix the code for fucks sake? God damn MSFT what is going on?

[Ryan Flaherty @rflaherty71]

@Sam75614812 Yes I am the source

[Ryan Flaherty @rflaherty71]

It is unbelievably difficult to make changes to the systems at big tech companies. It is at least 10x harder than just writing the thing from scratch yourself.

[siddharth @buildwithsid]

amazon has like 100k employees, why can't one of them improve the main website wtf

it still gives stone age vibes 💀

[Ryan Flaherty @rflaherty71]

alternatively someone could have just fixed the horrible language 20 years ago

[Olex (Solo gamedev Diablo-like) @OlexGameDev]

What is AI good at saving you time? One example is writing boring helpers like this.

[Ryan Flaherty @rflaherty71]

it's theft I mean tax day, they took all my money

[Ryan Flaherty @rflaherty71]

what if instead of a slow path and a fast path we just made the slow path fast

[Ryan Flaherty @rflaherty71]

@przekursor bootstrapped my own kernel in qemu, so the code here is running against the qemu virtualized nvme device

[Ryan Flaherty @rflaherty71]

got a successful round trip through nvme (I may have started writing an operating system...)

[Ryan Flaherty @rflaherty71]

@bmattlife just c++, I stay away from most of the language though

[Ryan Flaherty @rflaherty71]

@cmuratori I keep expecting to hit some wall but every time I go to implement something I just find it to be shockingly smooth and easy to do (I'll stay away from USB and the GPU for now..)

[Ryan Flaherty @rflaherty71]

finally got around to that fil and casey podcast on fil-c, fil is kinda awesome

[Ryan Flaherty @rflaherty71]

@lettuce_isgood OOOE is on the backend?

[Ryan Flaherty @rflaherty71]

This was before my time but I would guess the problem wasn't compliers supporting it, that is easy. The problem was no back compat for already compiled x86 apps. That is kinda what we have today, x86 front ends on itanium backends.

[Ryan Flaherty @rflaherty71]

@DonielOediv @lettuce_isgood fired?

[Ryan Flaherty @rflaherty71]

@lettuce_isgood what are your credentials, anon?

[Ryan Flaherty @rflaherty71]

@graykevinb I think you forgot to read @maxliani's bio before replying

[Ryan Flaherty @rflaherty71]

Twitter is too funny

[Kevin Gray @graykevinb]

You need to add accumulation and temporal reprojection.

For CPUs optimizing the scene for cache hits also matters a lot. Naively using if statements or switch will cause massive slowdowns compared to using goto statements.

On a simple scene this should get something that's noisy but shockingly clear.

If you notice it looks a bit muddy and blurry with DLSS. So its not actually solving the issue.

The image attached is a path tracer with temporal reprojection and running on WebGPU. There's a bit more noise while moving, but its smooth and very usable and instantly converged when you stop.

And this is running on old intel iris xe at 1440p.

The noise problem oddly gets worse if you decrease resolution. So a light denoiser is honestly the best solution.

Imagine how good this would run with a lightweight denoiser.

Like you gotta understand. I wasn't even trying. I just had gemini 3 Pro make this in an evening.

If I had used vulkan, written it myself it would be even faster.

And I bet I could do a lot more optimizations.

or heck get a better GPU.