Array
(
    [content] => 
    [params] => Array
        (
            [0] => /forum/index.php?threads/how-china%E2%80%99s-new-ai-model-deepseek-is-threatening-u-s-dominance.21956/
        )

    [addOns] => Array
        (
            [DL6/MLTP] => 13
            [Hampel/TimeZoneDebug] => 1000070
            [SV/ChangePostDate] => 2010200
            [SemiWiki/Newsletter] => 1000010
            [SemiWiki/WPMenu] => 1000010
            [SemiWiki/XPressExtend] => 1000010
            [ThemeHouse/XLink] => 1000970
            [ThemeHouse/XPress] => 1010570
            [XF] => 2021770
            [XFI] => 1050270
        )

    [wordpress] => /var/www/html
)

How China’s New AI Model DeepSeek Is Threatening U.S. Dominance

XYang2023

Well-known member
It contains a 30-minute interview with Perplexity CEO Aravind Srinivas, which I found to be very informative.

 
You can test model on Your computer. Download and install Ollama , then enter ollama run deepseek-r1 into command line. It should download and run 7B model.

You can also test different sizes or different model from library. https://ollama.com/library/deepseek-r1

I think this model is pretty good but size is still limiting factor at least for personal use (locally). I spend 30 minutes forcing it to fix one function (unsuccessfully). Copilot (o1) fixed same issue instantly (literally just "fixt it" prompt). But again, it is more issue with size and 600B+ is probably better...
 
Ollama distributes quantized versions. With quantisation, perplexity increases. But as you said, it could be due to model sizes:

1737815625604.png


14B @ 4bit -> 12GB graphic memory requirement

.

Intel B580 should work for the 14B model size.

32B @ 4bit -> 24GB graphic memory requirement

 
I really hope Intel improves its efforts to promote their GPUs. I have many ideas they could consider, but I don't work for Intel, and they should strive to do better in this area.
You might consider trying posting your ideas on the intel.com community forum for graphics products, which includes the B580. Perhaps someone from Intel will see the value in them, and contact you.

 
You might consider trying posting your ideas on the intel.com community forum for graphics products, which includes the B580. Perhaps someone from Intel will see the value in them, and contact you.

I’m not sure. I’ve made several videos about the B580 and tagged @Intel and @MJHolthaus, but I haven’t received any direct feedback from Intel.

One of my videos has thousands of views, clearly showing that many people are interested in using the B580 for machine learning and AI. I genuinely feel that, instead of sampling the cards to some YouTubers, Intel could consider sending samples to me or my school. We could test them and provide valuable feedback.

I really hope Intel's marketing team becomes more proactive in addressing the market and takes steps to actively prepare for Falcon Shores.

I work in a university robotic lab and we have and use quite a lot of GPUs.
 
It contains a 30-minute interview with Perplexity CEO Aravind Srinivas, which I found to be very informative.

I don't understand why this is a threat? And who said the US is dominant? Because you read it on the internet?
I’m not sure. I’ve made several videos about the B580 and tagged @Intel and @MJHolthaus, but I haven’t received any direct feedback from Intel.

One of my videos has thousands of views, clearly showing that many people are interested in using the B580 for machine learning and AI. I genuinely feel that, instead of sampling the cards to some YouTubers, Intel could consider sending samples to me or my school. We could test them and provide valuable feedback.

I really hope Intel's marketing team becomes more proactive in addressing the market and takes steps to actively prepare for Falcon Shores.

I work in a university robotic lab and we have and use quite a lot of GPUs.

Who at Intel did you contact? I may be able to help. Send me private email through SemiWiki.
 
I don't understand why this is a threat? And who said the US is dominant? Because you read it on the internet?


Who at Intel did you contact? I may be able to help. Send me private email through SemiWiki.
Thank you. I’ll think about that. I believe the lab I work in is definitely open to collaborations.

Recently, I tagged Intel and MJ on Twitter in a post suggesting they should do more to promote the B580 to the ML/AI audience. I also shared a video I created analyzing the B580 for machine learning purposes. I felt this should be Intel’s responsibility, not mine. Today, I tagged Intel and MJ again with a new video.

Additionally, I tagged Robert Hallock on Bluesky regarding AMD's misleading tweet about the Linus/Jimmy Fallon show. I suggested he consider being more active on X (formerly Twitter) since misinformation can spread quickly there. I also mentioned that Intel should work to address and correct such issues.

However, it feels like this communication is entirely one-sided...
 
You can test model on Your computer. Download and install Ollama , then enter ollama run deepseek-r1 into command line. It should download and run 7B model.

You can also test different sizes or different model from library. https://ollama.com/library/deepseek-r1

I think this model is pretty good but size is still limiting factor at least for personal use (locally). I spend 30 minutes forcing it to fix one function (unsuccessfully). Copilot (o1) fixed same issue instantly (literally just "fixt it" prompt). But again, it is more issue with size and 600B+ is probably better...
The Ollama deepseek-r1 model is a distilled version, it's not the deepseek V3 R1. The name chosen by Ollama is very misleading.
 
The Ollama deepseek-r1 model is a distilled version, it's not the deepseek V3 R1. The name chosen by Ollama is very misleading.
I think it’s fine. My understanding is that R1 stands for Reasoning model 1. Depending on the parameter size, the base models vary. I used the benchmark table to select the model, which I discussed in my video.

1738065654329.png
 
You can test model on Your computer. Download and install Ollama , then enter ollama run deepseek-r1 into command line. It should download and run 7B model.

You can also test different sizes or different model from library. https://ollama.com/library/deepseek-r1

I think this model is pretty good but size is still limiting factor at least for personal use (locally). I spend 30 minutes forcing it to fix one function (unsuccessfully). Copilot (o1) fixed same issue instantly (literally just "fixt it" prompt). But again, it is more issue with size and 600B+ is probably better...
The repeated monologue responses wear on me quickly.
 
Back
Top