DeepSeek + Huawei Just Went After Nvidia's CUDA Moat

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: DeepSeek + Huawei Just Went After Nvidia’s CUDA Moat · Channel: Fahd Mirza · 09:27 · Watch on YouTube

Topics: AI infrastructure and hardware, AI models, Open source and software ecosystems

Summary

The video explains DeepSeek’s release of DeepGEM Ascend and TileLang, presenting them as an effort with Huawei to make AI software work efficiently on Ascend chips and reduce reliance on Nvidia’s CUDA ecosystem. DeepGEM Ascend is described as a port of DeepSeek’s GPU library that preserves its API, while TileLang offers a Python-style way to write kernels for multiple hardware platforms. DeepSeek reports performance close to the stated hardware limits in several tests, though the video notes these figures come from the project’s own repository. The release is presented as a meaningful starting point, not proof that Nvidia’s lead or CUDA’s ecosystem will soon be displaced.

Key points

DeepSeek has open-sourced DeepGEM Ascend, which ports its DeepGEM library to Huawei’s Ascend chips.
The release announcement reports performance at 99.8% of the hardware limit on GEMM and 98% on a mixture-of-experts workload.
DeepGEM Ascend keeps the same API and package name as the existing library, aiming to make adoption easier for developers.
The library supports BF16, FP8, and FP4 matrix operations, along with attention scoring and mixture-of-experts operations.
DeepSeek’s repository reports results close to stated hardware limits on large BF16, FP8, and FP4 tests.
TileLang is an open-source language for writing AI kernels; it already targets Nvidia CUDA, AMD, and Apple Metal, and has added an Ascend 950 backend.
The video frames TileLang as an attempt to make high-performance kernel development more accessible without requiring developers to specialize in one chip.
Huawei provided technical backing during development, and the video describes a super node built from 128 Ascend 950 chips.
The video cautions that approaching a chip’s stated limit does not establish how its performance compares with Nvidia’s newest hardware.
CUDA’s established tools, documentation, and existing code remain significant advantages; adoption will determine whether the new tools gain traction.

Chapters

Start	Chapter	Summary
00:01	DeepSeek and Huawei’s push against CUDA	The video introduces the attention around DeepSeek and Huawei’s work and explains the open-source DeepGEM Ascend announcement.
02:38	DeepGEM Ascend and its reported performance	The video describes how the library works with Ascend chips, preserves DeepGEM’s API, and reports results close to hardware limits.
05:49	TileLang and the Ascend software stack	The video explains TileLang’s role in writing AI kernels and discusses its Ascend backend alongside Huawei’s hardware plans.
08:25	What the release does—and does not—prove	The video stresses that strong utilization does not prove parity with Nvidia or guarantee adoption, while calling the open-source release a promising start.

Quotes

Deep Jam Ascent is now open source. — 01:05

That is the real mode. — 02:08

It keeps the same API, same package name, same way of calling it. — 03:43

Those are two separate questions and this release answers the first one, not the second. — 08:25

They are handing you the code and saying run the test and check for yourself. — 09:27

Concepts and entities

DeepSeek, Huawei, NVIDIA, CUDA, DeepGEMM Ascend, TileLang, Ascend 950, TVM, Mixture of Experts

Full transcript (original language)

00:01 AI social media is going crazy about this some new magical tool which is going to not only break Nvidia’s monopoly on hardware, but more importantly on CUDA software. The thing is that almost a year ago I made this video, Chinese chips are rising. It was about a programming language called Tile Lang and the idea that China’s AI chips might one day stop depending on Nvidia’s software. Not many people watched it as you can

00:32 see. Today my timeline looks very different. Everyone is talking about it as you can see from this tweet and there are heaps of others. Deep Seek and Huawei have joined hands. And they are going after Nvidia on two fronts at once. The hardware and the software. So I want to go through what actually happened, what these tools are, and what they do and do not prove. Look at this tweet from an infrastructure engineer from Deep Seek.

01:05 An engineer who I think works at the core of their infra. He has talked about Deep Jam. Again, we have covered this on the channel already months ago. Deep Jam Ascent is now open source. That is the big and actual news. And he says that they have hit 99.8% of the hardware limit on Jam and 98% on mega mixture of expert. That was the whole announcement. No

01:36 stage, no keynote, just a link to the GitHub repo, which I will drop the link in video’s description too. But if you know what that sentence means. And it means a lot. It is a big deal. And the reason is a thing I think most people are ignoring at the the amongst all the hype. Why does Nvidia own AI at the moment? Everyone says the chips and the chips are great, but a chip is just a silicon

02:08 until software tells it what to do. For almost more than two decades, that software has been CUDA. Look, the New York Times this week put the number of CUDA developers at an estimated 4 million worldwide. 4 million people who learned one way of working, built their tools around it, and would have to start over if they left. That is the real mode. You can build a faster chip, you cannot easily build 4 million people’s worth of habit.

02:38 And this is exactly the problem China has. Export limits cut off the best Nvidia hardware, so Huawei built its own chips, the Ascend line. But having a chip is only half of it, and we have been covering Huawei’s chips for a long time. The thing is that a new software ecosystem needs a language that is universal, easy to program, and still reaches the hardware’s full performance. Read that again. You are reading a description of what CUDA became.

03:10 Now, let me tell you what exactly this new Deep Seek stuff is in as simple words as possible. So, this is the first piece, Deep GEM Ascend. GEM means general metrics multiplication. It is the operation sitting at the heart of a neural network, multiplying giant grids of numbers again and again, billions of times. If that one operation is slow, your whole model is slow. Deep Seek already had a library for this on Nvidia’s GPUs called Deep GEM, which

03:43 again we have covered on the channel. What they have done now is to port it to Huawei’s Ascend chip. And here is the the part. It keeps the same API, same package name, same way of calling it. A developer who already uses Deep GEM installs the Ascend version and carries on working the way they always did. The biggest thing keeping people on CUDA is the pain of relearning everything and this is built to remove that pain. Underneath, they wrote a thin layer over

04:14 Ascend’s matrix multiply instructions. Anyone who has written low-level kernels knows the misery of memory layouts, alignment rules, and address calculations. That layer hides most of it, so the kernel stays short and readable. The library covers BF16, FP8, and even FP4 matrix math, plus the attention scoring and mega mixture of expert operations that Deep Seek models rely rely on and various others including Kimmy.

04:46 And even um mini max and various others as they go about. Now the numbers, I want to be very careful here. Everything I’m about to say comes from Deep Seek’s own repo. I’m just quickly going to show you that repo. This is the repo. And if I go down, these are all the operations which I just mentioned in simple words. And then you can just go through it and check that out. But the numbers are really really crazy. So, I’ll just stop here to show you some

05:18 of the stuff which which they have shared around the numbering. And these numbers cover various facets, but I’m just going to stick with some of the salient ones. On a large BF16 test, they have reported 431 teraflops against a hardware limit of 432. For FP8, 861 against 865. And you can see that for FP4, 1701 against 1730. When a kernel lands that close to the

05:49 limit, the software is leaving almost nothing on the table. The chip is doing nearly everything it physically can. Mega mixture of expert is the one from that post, I think, which fuses the whole mixture of experts step into a single operation, and they have reported it passing 800 teraflops on the larger system. Now, let me show you the second part which they have released, and that is primarily this TileLang. That is the

06:19 second piece. And this is the one, I think, matters most in the long run, even more than this DeepJam. It is an open-source language for writing AI kernels. You write short Python-style code describing how data moves in tiles and how the matrix math runs, and the compiler turns that into fast code for the chip. It is built on TVM, and it already targets Nvidia CUDA, AMD, and Apple Metal. Yes, all of them.

06:50 Now, if you check out its readme, and they have a very good repo, which we already have covered. So, this basically is saying that it added an Ascend 950 backend with native code generation and automatic scheduling and synchronization. That is a long way of saying a developer no longer has to hand-tune every detail for this chip. Now, one of the kernels inside DeepJam Ascend, the one for Deep Seek’s

07:20 hyperconnection module, is written in TileLang. The DeepJam Ascend shows the chip can go fast, and TileLang is the attempt at what actually built CUDA’s empire, a way for ordinary developers to get that speed without being expert in one specific chip. And that is a huge thing. Then there is a hardware side we already have covered in another video, but very quickly Huawei gave full technical backing during development and the two companies worked on a super node built from 128

07:53 Ascend 950 chips turning both the compute and the communication between them. This basically comes right after Huawei unveiled its next generation of AI processors and super node systems and said it expects wide adoption for model training over the coming years and I think that is happening. Chips, clusters, and software all arriving in the same window. That is what a coordinated push looks like. So I think uh still long way to go for

08:25 this. Now you might be asking what about Nvidia’s grip on this. I don’t think so that is going to break like tomorrow. And I’m not going to pretend otherwise. Reaching 99% of a chip limit is impressive. But it only tells you the software is good. It does not tell you how high that limit is compared to Nvidia’s newest hardware. Those are two separate questions and this release answers the first one, not the second. And an ecosystem is more

08:56 than one library. CUDA has years of debuggers, profilers, documentation, and millions of lines of existing code. A new language only matters if developers actually adopt it. So that is where the push will come in. But again, if you look at this deep jump, look at how it was released. I think it’s a great starting point and even well, you know, way uh beyond that. MIT license, open source, same API as the existing library, built by the same team that built the models,

09:27 and nobody asking you to trust a press release. They are handing you the code and saying run the test and check for yourself. That is how ecosystem really start or flourish. Not with a declaration that CUDA is over but with a team quietly proving the other road can be walked and that is the best thing. So check it out. Let me know your thoughts in the comments. Recently I have started this free weekly AI newsletter at wahabmirza.substack.com. If you want to help out the channel, please become a member. Thank you.