Artificial Intelligence thread

tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
https://twitter.com/i/web/status/1887817659624120364

This morning I wake up and a whole bunch of more OEMs, chip makers and server providers are saying they are using or support DeepSeek models. This is a phenomenon. I have never seen anything like this where the entire Chinese social media are basically using their models for anything and everything.
 
tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
https://twitter.com/i/web/status/1887504843624833199

Very interesting, Phoenix Decentralized AI infrastructure, based in Shanghai is deploying Deepseek R1 & Qwen 2.5 on their Ascend high performance clusters, using 200+ nodes from China Telecom Global and Tensor

Again, seems to confirm the part about 910C delivering 60% of H100 for Inference.

And they are looking to use Ascend for both inference and training.
 
T

tokenanalyst

Brigadier
Registered Member
tphuang said:
Can you guys not post random takes online?

If you have your own take, then make a company.

I've harped on this for several pages now. Why are we still posting 1 line tweets or random videos without any commentary?
Click to expand...
I thought your post explained itself that people who were skeptical about how DeepSeek was going to get a return on its innovations were wrong. And that guy was the poster child of that skepticism debate.
 
tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
tokenanalyst said:
I thought your post explained itself that people who were skeptical about how DeepSeek was going to get a return on its innovations were wrong. And that guy was the poster child of that skepticism debate.
Click to expand...
I don't see the reason to post every random take online. We shouldn't care about what rando online think. It serves no purpose for this thread.
 
H

huemens

Junior Member
Registered Member
tphuang said:
https://twitter.com/i/web/status/1887504843624833199

Very interesting, Phoenix Decentralized AI infrastructure, based in Shanghai is deploying Deepseek R1 & Qwen 2.5 on their Ascend high performance clusters, using 200+ nodes from China Telecom Global and Tensor

Again, seems to confirm the part about 910C delivering 60% of H100 for Inference.

And they are looking to use Ascend for both inference and training.
Click to expand...

Just a thought regarding the 60% thing. There's a small chance this is a comparison of fp16 for Ascend and fp8 for H100.
The original weights supplied by Deepseek is fp8 and H100 cards have native support for fp8.

But the instructions for running on Ascend using MindIE requires the weights to be up-scaled to fp16 first. So either their software framework or the cards doesn't support fp8. At least the specs I have seen on the internet for 910B does not mention support for fp8. May be 910C has but software is not upgraded yet.
If they are running with fp16 the compute throughput will be much less than fp8 and it would consume double the RAM and memory bandwidth too.
 
tphuang

tphuang

Lieutenant General
Staff member
Super Moderator
VIP Professional
Registered Member
I think Ascend-910C is still a new platform and with customer feedback and plenty of extra hours, they should be able to increase its inference performance. I'm not too concerned.

Anyhow
Please, Log in or Register to view URLs content!

Looks like HarmonyOS's Xiaoyi assistance is using Ascend computation provided by SiliconFlow and Huawei Cloud. So, it looks like Huawei/Ascend have been putting a lot of overtime recently to get DeepSeek up and running on its platform.

Now if DeepSeek can just secure more compute and put them up on its servers so that it can serve all those requests
 
T

tokenanalyst

Brigadier
Registered Member
You must log in or register to reply here.
Top