155
Chinese algorithm claimed to boost Nvidia GPU performance by up to 800X for advanced science applications
(www.tomshardware.com)
this post was submitted on 04 Feb 2025
155 points (94.3% liked)
Technology
63082 readers
3659 users here now
This is a most excellent place for technology news and articles.
Our Rules
- Follow the lemmy.world rules.
- Only tech related content.
- Be excellent to each other!
- Mod approved content bots can post up to 10 articles per day.
- Threads asking for personal tech support may be deleted.
- Politics threads may be removed.
- No memes allowed as posts, OK to post as comments.
- Only approved bots from the list below, to ask if your bot can be added please contact us.
- Check for duplicates before posting, duplicates may be removed
- Accounts 7 days and younger will have their posts automatically removed.
Approved Bots
founded 2 years ago
MODERATORS
you are viewing a single comment's thread
view the rest of the comments
view the rest of the comments
Per the article, they just developed a faster algorithm for a specific type of material simulation.
You're underplaying it.
As per the article:
Yes, exactly… For another example, the DeepSeek team developed their own replacement to CUDA with PTX (Parallel Thread Execution) a lower-level assembly-like language that allows for more granular optimisations of GPU performance offering 10X efficiency improvement recently as GPU sanctions were levied on China. This innovative approach not only challenges the dominance of CUDA in the AI landscape but also opens up new possibilities for optimizing GPU performance in various applications and this is what missing not only from those relying on Nvidia but its competitors whether AMD or Apple that prefers to have its own proprietary solutions.
Nvidia developed PTX, DeepSeek leveraged it to do some load balancing work they couldn't do in CUDA. They still also use CUDA.
I am aware and hence the word optimisation (instead of simply pouring money by hardware).
Your original comment was wrong on several counts and this one... just says nothing. I'm not sure what you're trying to say even.
I guess some people are beyond help.
You are indeed.
Whatever you say, I'm not the one making wrong claims and then arguing completely different claims when it's pointed out lol
Well, the initial emphasis here was optimisation and the argument going on extension is anything other than the main emphasis here so… ¯_(ツ)_/¯
You literally said they developed it.
*Developed their own replacement using PTX to gain granular control over CUDA and hence the emphasis with optimisation as a user so all software, app, game and likewise developers take this practice into consideration again.
No they didn’t. NVIDIA created PTX
I write HPC and this really isn't that crazy. Its 800x vs a serial program ran on a CPU, it is only 100x a OpenMP (parallel CPU version).
While the gains are very impressive and show clear deep understanding of HPC programming, it is really nothing that out of the ordinary.
With proper memory optimisations and algorithm optimisations from a single thread to a dedicated GPU model a 100-1000x speed up is pretty normal.