Yesterday, Huang again took the stage in his iconic leather jacket, delivering a keynote speech at the GTC (GPU Technology Conference).

After listening to the whole speech, I realized that Huang did have quite a few big surprises up his sleeve this year.

This time, Huang finally didn't forget us "gamers".

Without hesitation, he dropped a bombshell: RTX Spark, the much-talked-about N1X processor that had been circulating for a while.

This product is a deep collaboration between NVIDIA, Microsoft and MediaTek, with a clear goal: breaking the 40-year limitations of traditional PC architectures and redefining the entire PC industry.

At present, the real situation is this: There is an irreconcilable rift between the traditional PC architecture and the demand for local AI.

Everyone wants to run a large model locally, but the harsh truth is that today's computers were never designed for local AI in the first place.

The video memory on the graphics card can run AI, but its capacity is really insufficient. Even the top-of-the-line 5090 graphics card only has 32GB of video memory, making it impossible to run a large model directly.

While the regular memory in a computer has a large enough capacity, its read and write speed is far too slow, making it really tough for it to run large models.

So running AI on traditional PCs has always been a headache.

Apple's pioneering unified memory architecture for SoC, integrates CPU, GPU and NPU into a single chip, breaks down the barriers between system memory and video memory. All cores share the same memory pool, completely freeing themselves from the limit of video memory capacity, and the memory resources available for AI computing are greatly improved.

So in the past two years, we can see that the Mac Studio equipped with full-power memory, with 8 channels and 512GB of large memory, offers a top-tier local AI running experience; AMD has also launched the AI Max+ 395. Although its performance is slightly weaker, it uses a similar architecture, and with 128GB of memory, allocating some to the graphics card is enough to run medium-scale models.

These products can indeed run AI without any issues, but their AI support always feels a bit lacking.

In contrast, NVIDIA, which has been deeply engaged in the CUDA ecosystem for decades, holds the world's most mature AI development and application system, a core advantage that Apple and AMD cannot catch up with in the short term.

Watching the local AI market being gradually carved up by competitors and amid the explosive growth of the AI agent track right now, Jensen Huang finally stopped sitting on the sidelines, took the initiative to enter the fray to break the industry deadlock, and RTX Spark was born.

Why can Apple and AMD do unified memory architecture, but I can't, Jensen?

The CPU part of RTX Spar is the custom Grace CPU co-developed by NVIDIA and MediaTek, consisting of 20 Arm cores. According to current leaked benchmark results, its performance is roughly on par with Apple's M3 Max from a few years ago.

The GPU is equipped with 48 stream processors and a total of 6144 CUDA cores, delivering performance equivalent to the desktop 5070 graphics card. This scale is by no means small. When it comes to the computing power that AI values more, it can reach 1P at NVFP4 precision, which is 1000 TOPS.

As a processor for the AI era, the RTX Spark also uses unified memory with a maximum capacity of 128GB, which is more than enough to run many models.

It's just that the read speed of this unified memory is only 273 GB/s, on par with AMD's AI Max+ 395, which is still a bit lower than Apple's.

However, server-grade NVLink is directly used between the CPU and GPU, providing a maximum bandwidth of 600 GB/s, which completely outperforms the PCIe interconnection used on traditional PCs.

Of course, NVIDIA's biggest trump card is still the CUDA ecosystem, which enables various AI applications to run quickly.

On the spot, Huang demonstrated such a scenario: by using an Agent to connect tools like ComfyUI and Blender, he could complete the entire process from room drawing, modeling, rendering to AI-generated preview images on a single personal computer.

Beyond AI, NVIDIA's traditional forte: gaming, has not been forgotten on RTX Spark.

With the power of RTX Spark, playing 2K games is completely hassle-free. Moreover, regarding the annoying anti-cheat issues on Windows on Arm before, NVIDIA and Microsoft have also put in a lot of effort to achieve ARM-native compatibility for the underlying components of mainstream online game anti-cheat systems such as Easy Anti-Cheat and BattlEye.

Huang also showed two laptops on the spot, one running the latest *007* and the other running *Horizon 6*.

Of course, besides catering to us regular consumers, the server market, which is where Nvidia truly rakes in huge profits, hasn't escaped his attention either.

In NVIDIA's view, current CPUs can't keep up with the pace of GPUs.

Has the CPU become a bottleneck for the GPU?

On-site, Huang used an analogy: If the GPU is an orchestra, then the CPU is its conductor.

For an orchestra to play a perfect piece, the conductor's tempo must keep up.

Nowadays, as Agent tools like Claude Code and Lobster grow increasingly popular, the computing speed of the CPU has become a serious bottleneck for the GPU.

Let's take a simple example: we ask an Agent to do something casually, like asking it to summarize NVIDIA's latest quarterly financial report.

At this point, the CPU first has to scrape data online to figure out which quarter the latest financial report is for, then conduct a search, lock onto the target, and run a download script to pull the financial report down.

Only after all these trivial tasks are done can we truly start analyzing the financial report.

Looking back at the entire process, you'll find that an Agent can't finish a task in one go. Basically, it's a cyclic relay mode where the GPU handles part of the work first, then the CPU takes over to do some tasks, and then the GPU takes the stage again.

So Lao Huang just came out and said it straight: the current CPU has become the bottleneck for GPU utilization.

This time, they specially created a CPU for Agent tools: NVIDIA Vera.

This product can be said to have been optimized around the four words "reduce latency" from start to finish.

In the past, the vast majority of server CPUs were actually made by piecing together several small chips. The advantage of doing this is that the chip yield rate is higher and the cost is lower.

But the downside is that the communication speed between cores isn't all that fast; sending a message from one core to another requires a long detour.

But Vera is free of such troubles. To make it run faster, Huang directly integrated 88 computing cores onto a single chip.

This directly increases the communication speed between cores by 50%, which is equivalent to upgrading from a two-lane road to a three-lane road.

After these moves, the speed at which Vera works has been tweaked by Huang to be quite impressive.

Huang gave an example with Starburst's SQL analysis test: in the same data analysis benchmark test, Vera runs three times faster than X86 CPUs.

In the real-time stream test at the New York Stock Exchange, the Vera CPU even managed to cut the computing latency down to one-sixth of its original value.

If I had this data center to trade stocks, even Warren Buffett would have to call me a stock god.

Of course, in addition to all the above, Huang also shared a lot of interesting things at this year's GTC.

There is the "Cyber Strategy" DSX that teaches you how to build a data center, which allows you to use simulation software to test the power, cooling and network environment of the factory before you actually break ground.

There is also a complete big gift package prepared for Agents: an enterprise-oriented Agent tool suite, and an OpenShell framework that ensures AI safety...

Finally, he also unveiled a world model for robotics and autonomous driving: Cosmos 3.

Among all the new products, RTX Spark is undoubtedly the one that has received the most public attention.

Over the past 40 years, the global PC market has long been monopolized by the duopoly of Intel and AMD's x86 architecture, with almost no new players able to break through.

Previously, Qualcomm tried to make inroads into the Windows Arm ecosystem, but due to issues such as insufficient GPU performance, poor adaptation of the DirectX game ecosystem, and maladjustment in the adaptation of software and hardware, it has always struggled to shake the mainstream market.

Moreover, the launch of a brand-new platform is often accompanied by concerns from software developers and OEM manufacturers that they will only "dip their toes in" rather than fully commit to it. This is why Windows laptops still predominantly rely on the traditional X86 architecture to this day.

Fortunately, NVIDIA is well aware of the difficulty in promoting the new platform, and Jensen Huang is mentally prepared for a long-term battle on the RTX Spark path. This time, it directly announced the technical roadmap up to 2030: Currently, it is Blackwell Spark, and in the future, it will be Rubin Spark and Rosa Feynman Spark.

What's more, with two ace cards, RTX and CUDA, in hand, even for the underlying adaptation work of software and games, the progress speed and developers' enthusiasm are far from comparable to the loose alliance between Qualcomm and Microsoft back then.

Now the baton has been passed, and Huang has done everything he needs to do. From now on, all the attention and pressure will fall on Microsoft.

At the end of the day, whether RTX Spark can truly gain a foothold hinges on two key points: first, the final pricing strategy; second, whether the Windows on ARM ecosystem can really take off.