AMD shows Radeon 7900 XTX outperforms Nvidia RTX 4090 in DeepSeek benchmarks

D

DragonSlayer101

Posts: 577   +3
Staff
TL;DR: AMD has published benchmarks that appear to suggest its Radeon RX 7900 XTX graphics card is significantly faster than Nvidia's RTX 4090 and RTX 4080 Super when it comes to DeepSeek R1 benchmarks. Team Red also released instructions for DIY enthusiasts to run the R1 on Ryzen AI CPUs and Radeon GPUs.

According to an X post by David McAfee, the VP and GM of Ryzen CPU and Radeon graphics at AMD, the company tested the 7900 XTX against the RTX 4090 and RTX 4080 Super using DeepSeek R1, and the RDNA 3 GPU was between 13 percent and 34 percent faster than Nvidia hardware depending on the LLM and number of parameters.

With seven billion parameters, the 7900 XTX trumped the RTX 4090 by 13 percent in Distill Qwen, while it was 11 percent faster in Distill Llama with eight billion parameters. With 14 billion parameters, the AMD card beat the Ada Lovelace flagship by two percent in Distill Qwen. However, when the number of parameters went up to 32 billion, the RTX 4090 had a four percent advantage over the RDNA 3 flagship in Distill Qwen.

AMD also compared the 7900 XTX to the RTX 4080 Super, showing its card was faster in every test. The AMD product outperformed the Nvidia GPU by 34 percent using DeepSeek R1 Distill Qwen with seven billion parameters and 27 percent using Distill Llama with eight billion parameters. With the number of parameters increased to 14 billion, the card from Team Red was still 22 percent faster using Distill Qwen.

In addition to the benchmarks, AMD also published detailed instructions on how to run DeepSeek locally on your PC. According to the company, all its RDNA 3 desktop GPUs are compatible with several LLMs using R1. Most of these models can also run on select Ryzen CPUs with XDNA NPUs.

AMD also released a list of the maximum supported LLM parameters and provided instructions on how LM Studio's one-click installer can be tuned for its hardware. According to the company, the Radeon RX 7900 XTX supports up to 32 billion parameters, while the entry-level RX 7600, with its 8GB of VRAM, supports up to eight billion parameters in Distill Llama.

Permalink to story:

AMD shows Radeon 7900 XTX outperforms Nvidia RTX 4090 in DeepSeek benchmarks

 
As someone who made a hefty chunk on AMD stock when the Ryzen architecture was released... this is interesting if verifiable... and if the news gains traction... NVIDIA will need to respond.. maybe a price drop... but what I suspect the will do is rush the roll out of the rest of the 50 series cards.

 
  • Like
Reactions: m3tavision, riev90, Clynt and 1 other person
Ashford said:
As someone who made a hefty chunk on AMD stock when the Ryzen architecture was released... this is interesting if verifiable... and if the news gains traction... NVIDIA will need to respond.. maybe a price drop... but what I suspect the will do is rush the roll out of the rest of the 50 series cards.
Click to expand...
Theyll probably allocate as much of their TSMC space to server grade chips and create an artifical shortage of the 50 series knowing full well they offer almost nothing over the 40 series if you aren't into AI.
 
What is a useful range of parameters?

A lot of the amuse models suck and when there is one that is good, it barely runs on the XTX.
 
How long to this is fundamental feature of your PC/GPU.
It's only going to grow as PC get more types of processing serial ,parallel, neural , quantum etc etc
You will have to have your own model to communicate with all the other models you sub to

Oh they cancelled your favourite show after one year - no worries someone got a model to create season two

This is the next fight who controls it , us to the government or mega corps

In 20 years people will pay big money to go to sanctuaries with no AI or modern tech

big brother/mother/other is really coming, EU has rights not to tracked I believe or be forgotten - let's see how that works out

Will we not only get black markets but black nets like a huge faraday cage , can't see how to stop 24 hour near perfect monitoring except with lots of effort

People here make fun of AI saying it isn't real or a fad, Nah we haven't even started
 
  • Like
Reactions: gdavid65
I’m just asking what the article means. In my experience with Amuse it’s not all the same, the details make all the difference.
 
kiwigraeme said:
How long to this is fundamental feature of your PC/GPU.
It's only going to grow as PC get more types of processing serial ,parallel, neural , quantum etc etc
You will have to have your own model to communicate with all the other models you sub to

Oh they cancelled your favourite show after one year - no worries someone got a model to create season two

This is the next fight who controls it , us to the government or mega corps

In 20 years people will pay big money to go to sanctuaries with no AI or modern tech

big brother/mother/other is really coming, EU has rights not to tracked I believe or be forgotten - let's see how that works out

Will we not only get black markets but black nets like a huge faraday cage , can't see how to stop 24 hour near perfect monitoring except with lots of effort

People here make fun of AI saying it isn't real or a fad, Nah we haven't even started
Click to expand...


Ai is kewl and all..
2 Humans talking is considered quantum speech. It's incalculable to derive a life experience x billion.

Ai is a tool, just like a stick.
 
m3tavision said:
Ai is kewl and all..
2 Humans talking is considered quantum speech. It's incalculable to derive a life experience x billion.

Ai is a tool, just like a stick.
Click to expand...
An imprecise tool at best, even our language can be imprecise as a tool.
That's why I never got the anti-AI - if shouldn't take over your executive control or editorial control .
Like any tool, you need to learn it, and know it's limitations.

I see some models now quote sources and or links

But really with PC we sometimes do a lot of boring admin.
Hate going various tax returns as a small business, Vat/gst , year end etc .
I mean it can look at 10 years of data banksheets , returns , credit card statements and fill out new form in seconds with colour coded certainty for your review

Lots of files on lots of drives, need to cleaned up .
I mean people will use it for creative stuff , feed a model every rpg module they can legally or illegally get and have it run a RPg as a DM - that can adapt on the fly and not get annoyed if players go off script

But I think just getting it do the mundane things

You will soon know it people are using them when a Media creator is "responding" to lots of comments and emails etc and remembers all your comments for years
 
MasterAce said:
I’m just asking what the article means. In my experience with Amuse it’s not all the same, the details make all the difference.
Click to expand...

You are confusing two different programs.

https://www.amuse-ai.com/

This is used for AI picture generation.

https://lmstudio.ai/

This is used for LLM's

This article is about how the 7900XTX outperforms both nvidia cards in LLM performance on deepseek.

Except for the 4090 having a 4% lead at 32B.
 
You must log in or register to reply here.

Similar threads

D
RTX 5090 early benchmarks show underwhelming performance uplift over the RTX 4090
2 3 4
Replies
78
Views
679
EdmondRC
E
zohaibahd
After DeepSeek stuns the AI world, Alibaba responds with an allegedly more powerful model
Replies
9
Views
59
kiwigraeme
K
S
AI startup DeepSeek rivals OpenAI models using far fewer resources, shocks AI industry
2
Replies
38
Views
293
mbk34
M

Latest posts

Back