Should we (Intel, AMD, NVIDIA, Google, Tesla, OpenAI, Amazon, Apple) be worried?
1 Like
Not a surprise. With all those cyber invasions from China stealing corporate secrets you wonder if they are using code stolen from OpenAI.
As to AI chips, experience is they can get their hands on anything they want embargoed or not. If they are willing to pay for it someone will supply it.
And no doubt they are working on their own chips. How many of Nvidiaâs design secrets have they stolen? How long will it take them to learn to make Chinese copies?
1 Like
It appears this is a new open source (free to everyone to use) software algorithm that outperforms the existing algorithms. I am sure China has all the either patented or trade secret US invented algorithms reverse engineered and in use already.
AI hardware is actually pretty simple, as an offshoot of traditional GPU hardware. The secret sauce is in the software, and the reason NVIDIA is leading is they have the largest base of software developers out there (similar to why X86 is so dominant). AMD hardware will not run NVIDIA software without a translation layer, which is still a work in progress.
China has shown this algorithm, running on fairly ancient hardware, produces excellent results. This translates into a significant reduction in the AI hardware TAM (total available market).
3 Likes
They can (and do) copy various chip architectural techniques. But copying the EUV FAB equipment made by ASML â not so easy.
AFAIK the best that SMIC (Chinaâs best FAB) produces is 7nm (whatever that means to them) and uses DUV (deep ultra violet). TSMC has 5nm, 4nm, 3nm, 2nm (coming this year?). So maybe they are 5 years behind.
[Edit: they meaning SMIC is at least 5 years behind in process size, they are already ~7 years behind in not have EUV and no way to steal one from a ASML customer]
Mike
2 Likes
DeepSeek didnât hide what they did and keep that secret. They are sharing it with the world in an open model. Some should learn from that instead of adopting the mindset âyou steal my idea-it is my ideaâ. That is not the way.
Banning and restricting, that is simply being afraid of competition.
Yes, thereâs plenty of propaganda and deception at play on both sides. Perhaps propping up DeepSeek is just another way to divert attention or a strategy to exaggerate the enemyâs strength- enemy according to the govât. By inflating its significance, any eventual victory over it could be portrayed as far more monumental than it truly is.
It was hard but we won on the end. Glory!
tj
So, trying to be optimistic about AMD, is there any reason (or not) to believe that AMDâs chips will be able to run this new software faster than NVIDIAâs best chips? I expect there will soon be testing apace, if it hasnât already been done or is in progress.
1 Like
My impression is that to the first order AMD hardware is just as fast as NVIDIA. Both companies use the same base silicon technologies from TSMC so this should not be surprising.
The software problem still exists though. AMD has been getting some fairly high volume design wins from companies that need many instances of some sort of AI acceleration and the less expensive AMD silicon with special AMD software is a great solution for them. It doesnât make sense for smaller deployments to do custom software though. I donât see this particular algorithm changing that dynamic much as both AMD and Nvidia should be able to change the âunder the hoodâ code to this and leave the external interface the same.
Shorter answer⌠I think it will be a push.
1 Like
No.
The comments about deepseek supporting Int8 are meaningless. Both NVIDIA and AMD have supported int8 for almost a decade. Current models also use Int4 in areas where low precision will do.
FYI,
Alan
You are missing the point. Using Int8 allowed them to save on the memory. it is not about NVDA or AMD can support it.
AI has ânewâ formats more than int8. Such as FP8 and the Open Compute MX formats which are several bit depths that are small blocks of floating point that all share the same exponent that is stored just once per block.
see here:
https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
Mike
1 Like
My point was (and I think it is still valid) that using int8 (or other lower precision formats) should not be a reason the deepseek algorithm is more efficient. Other algorithms have been doing this for years.
Personally, I would like to see more about this algorithm before I would chop 10% off the market value of Nvidia or AMD. It appears that today the market might agree with that.
1 Like
The thread you quoted on Macro Economic Trends and Risks is huge. People can find it here : Deepseek, some random thoughts
I donât get it personally.
The world will always be constrained by compute power; if itâs there, it will be used.
Having new algorithms needing less power will allow more to be done, rather than needing less hardware.
The other thing I notice: AMD gets hit hard by the market when these industry scares happen (even if not really related) and doesnât recover, whereas other larger companies tend to recover quickly.
3 Likes
What about smaller companies like Intel? (lol!)
1 Like
A related question. This open source software was released to the world by Chinese software ⌠dare I say, geniuses? Why would China allow that? I would think Xi would want China to keep this wonderful new AI software to itself. What does China get out of this? âShock and Aweâ?
Re: AMD
Investors see AMDâs AI chips as a growth opportunity. Its other chips mostly for PCs are not very exciting. Maybe even in a slump.
Re: China.
Maybe national prestige. And rubbing it in as a response to US embargos on advanced chips as from Nvidia. Also demonstrates skill level of Chinese professionals.
2 Likes
DeepSeekâs R1 model is pretty impressive. Now, Hugging Face (the company running the popular hub for open-sourced AI models) is going to fully replicate it to validate the claims of its low training cost and dispel any doubts relating to the resources actually needed to achieve this level of performance.
Barely a week after DeepSeek released its R1 âreasoningâ AI model â which sent markets into a tizzy â researchers at Hugging Face are trying to replicate the model from scratch in what theyâre calling a pursuit of âopen knowledge.â [âŚ] âThe R1 model is impressive, but thereâs no open dataset, experiment details, or intermediate models available, which makes replication and further research difficult,â Elie Bakouch, one of the Hugging Face engineers on the Open-R1 project, told TechCrunch. âFully open sourcing R1âs complete architecture isnât just about transparency â itâs about unlocking its potential.â
As pointed out by many, including Intel ex-CEO Pat Gelsinger and Microsoft CEO Satya Nadella, the market has misunderstood the implications from this breakthrough by presuming demand for AI compute will be negatively impacted. The aforementioned CEOs referred to the Jevons paradox, but I think this is somewhat missing the point as well. The point is not as much that AI is going to become an affordable commodity, thereby increasing the size of the market.
In my view, the advance made by DeepSeek indicates how much performance is to be had by innovation in this rapidly evolving field, but it in no way detracts from the performance still to be had by further brute-force scaling. The big companies are going to incorporate the innovations made by DeepSeek into their own models and scale it up by magnitudes, as far as available compute resources allow. The race is still on towards Artificial General Intelligence â and Artificial Super Intelligence beyond.
I doubt ASI will be achieved on a budget.
While AMD cannot displace all the bespoke CUDA software out there (in universities and small business, in particular), the lucrative matrix multiplication backend of AI models is a fairly narrow workload to target.
The popular AI programming frameworks, such as PyTorch and Triton, come with AMD backends already, and the popular Hugging Face Hub repository of open-source models, including Metaâs Llama models, are continuously tested for compatibility with AMD hardware out of the box.
An interesting thing is that DeepSeek skipped CUDA (Nvidiaâs framework) for some functions and optimised the implementation using Nvidiaâs low-level PTX bytecode, reportedly (Tomâs Hardware). Of course, that makes the solution even more dependent on Nvidia hardware. But it shows the extent to which serious developers will go to optimise for the hardware theyâve got.
Positive for AMD is that their big customers, Microsoft and Meta in particular, are putting in similar effort to optimise for the AMD Instinct MI300X solutions they have deployed at scale in their data centres â with good results.
AMD just needs to edge ahead on GPU chip design. Arguably, MI300 is already the most advanced chip in the world when it comes to advanced packaging (championing chiplet design and leading-edge 3D stacking using dense hybrid bonding that no one else can match so far).
Can their HPC chip design lead be turned into material advantage in the AI field as well?
1 Like
uh oh.
DeepSeek releases new models Janus-Pro and JanusFlow on Lunar New Yearâs Eve ¡ TechNode
{ DeepSeek unveiled two new multimodal frameworks, Janus-Pro and JanusFlow, in the early hours of Jan. 28, coinciding with Lunar New Yearâs Eve. Janus-Pro is an upgraded version of Janus, designed as a unified framework for both multimodal understanding and generation. By decoupling visual encoding, the model improves adaptability and performance across various tasks. It has outperformed OpenAIâs image-generation model, DALL-E 3, in benchmark tests. Consistent with previous models in the Janus series, Janus-Pro is open-source. }
FWIW. Since Iâve not seen this ânewsâ anywhere else.
ralph
1 Like