Speaker A
This is my MacBook Pro. I bought it a few years ago for a fortune, and I don't regret it at all. It's a joy to work on, but in 2026, doing work on a computer yourself will be a chore.
Explore the future of AI PCs in 2026, featuring unified memory, local LLMs, and innovations from Apple, NVIDIA, and AMD.
Key Takeaways
- AI PCs rely on unified memory to efficiently run large local AI models beyond traditional GPU limits.
- Apple, NVIDIA, and AMD are adopting different architectures to meet the demands of next-gen AI workloads.
- Memory capacity and bandwidth are critical bottlenecks for running advanced AI models locally.
- Distributed computing and clustering of compact PCs enable scaling beyond single-machine memory limits.
- The future of personal computing will integrate AI deeply, making manual work on computers less necessary.
What the video covers
- The video discusses the evolution of personal computers into AI PCs by 2026, focusing on multi-agent systems and autonomous assistants.
- It highlights the limitations of traditional NPUs and the importance of VRAM capacity and bandwidth for running large local AI models.
- Unified memory architecture (UMA) is introduced as a key innovation, combining system RAM and video memory for better performance and efficiency.
- Apple’s approach with Apple Silicon unifies memory across CPU, GPU, and NPU, enabling larger local LLMs on machines like Mac Studio and Mac Mini clusters.
- NVIDIA’s ecosystem, based on CUDA, supports AI workloads with powerful GPUs and new designs like the DGX Spark and RTX Spark PCs.
- AMD’s Ryzen AI Max Plus chip and compact PCs offer scalable AI performance with local clustering capabilities.
- The video explains how AI PCs differ from traditional laptops by supporting larger models with unified memory and distributed computing.
- It covers practical examples like running GPT OSS 120B and the challenges of memory requirements for advanced AI models.
- The video also touches on the importance of updated hardware like routers for maintaining high-speed internet connectivity in AI PC setups.
- Overall, it presents a comprehensive view of how major manufacturers are addressing the architectural shifts needed for future AI computing at home.
Chapters
- 00:00Introduction to AI PCs and future computing
- 01:48Limitations of traditional NPUs and hardware upgrades
- 03:15Dom Rave router line for reliable internet
- 04:40Memory bottlenecks and unified memory architecture
- 06:12Differences in AI PC memory design and model requirements
- 07:43Apple’s unified memory approach and Mac Studio innovations
- 09:12NVIDIA’s CUDA ecosystem and DGX Spark solutions
- 12:20NVIDIA RTX Spark and compact AI PCs
- 15:28AMD Ryzen AI Max Plus and scalable PC clusters
- 16:58Future outlook and AI PC ecosystem development
Full Transcript — Download SRT & Markdown
Speaker A
Multi-agent systems, autonomous assistants, and heavy local neural networks should do it for you. And for this task, you need a completely new type of computer. For example, here's the NVIDIA DJX Spark and the Asus Pro Art laptop. Compact machines with 128
Speaker A
gigabytes of memory in the heaviest local model. And this is the Rockflow Z13 gaming tablet, powered by the monstrous AMD Ryzen AI Max Plus chip with equally massive memory. Here's the new Mac Studio, where the processor and graphics can handle hundreds of
Speaker A
gigabytes of combined memory. And here's a cluster of four Gecom mini PCs, connected via USB 4, which in 2026 launched a single gigantic trillion-class model. And all this is AI PC. AI PC. But what does that even
Speaker A
mean? So today we'll figure out what an AI PC is, how it differs from the classic computers we're used to, and how Apple, NVIDIA, and AMD answer this question differently. This is Droider.
Speaker A
Valery Istishev with you. Let's get started. Let's get started. What is this AI PC of yours? Is it just a regular computer running a neural network? Basically, yes, but with a big "but." Until recently, we were sold regular laptops with a dedicated NPU
Speaker A
chip that had an AI chip. And the more powerful the NPU—40 trillion operations per second, 50-60—the more AI it has. When was the last time you upgraded your smartphone, updated your laptop, or maybe bought a TV? You probably remember it, and even name the
Speaker A
year and exact model. What about your router? Can you remember its manufacturer, model, and the year you bought it? In general, many internet problems are solved by simply rebooting the router, but it, like other equipment, needs to be updated. I mean,
Speaker A
buying something up-to-date. Routers also become outdated, and this can limit your home internet capabilities.
Speaker A
You can upgrade your plan and increase the speed, but an old router won't be able to handle it, and you'll be stuck with what you've got. Dom Rave is a line of routers for gigabit and reliable internet, consisting of three
Speaker A
models. The basic Dom Rave is suitable for small apartments; its sleek design allows for placement or installation anywhere, and it works with all providers. The advanced Dom Rave Plus is faster and more powerful. It's suitable for gaming and watching 8K
Speaker A
video. And finally, the Dom Rave Plus system. It can cover up to 200 m² of space. The advantages of this solution include seamless transitions without reconnecting and stable operation when moving between rooms, for example, during video calls. The
Speaker A
Dom Rave router line is a simple and elegant solution for always staying connected. Details are available at the link in the description, and with a promo code, you get even better deals.
Speaker A
So, let's move on. But what if it turns out that power isn't the only issue? If we want to run a large LLM, an image generator, a coding agent, or several models simultaneously directly on the computer, the bottleneck becomes a
Speaker A
completely different component: VRAM, or video memory, or more precisely, its capacity and bandwidth. And yes, you need a lot of this memory. A simple calculation lifehack: for each billion parameters a model has, you need approximately the same number of
Speaker A
gigabytes of video memory to run it with adequate four-bit compression. So, roughly speaking, for a model with 30 billion parameters, you'll need 30 GB RAM. It would seem that taking the flagship NVIDIA GeForce RTX 5090 is the answer. It has a monstrous 32 GB of
Speaker A
ultra-fast GDDR7 memory with a mind-boggling 1792 GB/s of bandwidth. For any model that fits into those 32 GB, it's ideal. But here's the rub.
Speaker A
There's a consensus in the industry. Models are truly becoming smarter and more intelligent, and are less likely to hallucinate between 70 and 90 billion parameters. This means you need at least 70 GB of memory to run this kind of model. And if you give it a
Speaker A
long context, you'll need a full 128 GB. And your top-end $5,000 graphics card, or whatever it currently costs, becomes a jack-of-all-trades. And this is where the major architectural shift occurs, around which the new class of AI PC devices is currently being built.
Speaker A
Instead of putting small, albeit ultra-fast, video memory on a discrete card, manufacturers combine system RAM and video memory into a single array, called unified memory (UMA). This memory is shared equally by the processor, GPU, and NPU. While unified
Speaker A
memory is significantly inferior to fast GDDR7 in pure peak speed, it can accommodate the entire model, which is more important. This gives us a key advantage: the zero-copy principle.
Speaker A
Data no longer needs to be constantly shuttled from RAM to video memory back and forth through the bottleneck of the PCI Express bus. The result is a noticeable performance boost and a significant reduction in power consumption. A good example is the GPT
Speaker A
OSS 120B model. This is a large, open Open AI model with 117 billion parameters. And thanks to MOE, about 5.1 billion are active at each current.
Speaker A
Officially, the MXFP4 version is designed to run with approximately 80 GB of memory. It physically doesn't fit into the 32 GB of the RTX 5090, but a modern machine with 128 GB of unified memory can handle it like a charm. This
Speaker A
means we understand that these computers are accessing a completely different class of models. And this is precisely how the new wave of AI PC differs from the old idea of a laptop with a dedicated NPU. Manufacturers are starting to redesign their memory
Speaker A
architecture for local LLM agent generators and other heavy AI tasks. And then the fun begins. As we've already mentioned, Apple, NVIDIA, and AMD are solving this problem in completely different ways. Moreover, a suitable architecture emerged long before the current AI boom. In 2020,
Speaker A
Apple released the M1 and ported the Mac to its own SoC architecture, calling it Apple Silicon. Did anyone think back then that this would revolutionize LLM? At least, no one said so out loud. But it was then that
Speaker A
Apple took the path of memory unification. Starting with Apple Silicon, the CPU and GPU share a single, unified memory. All blocks within the System on Chip (SoC) can access a single data layer without constantly copying between separate system and
Speaker A
video memory. At the time, unified memory was explained by the benefits of performance, power efficiency, graphics, and editing. Only later did it suddenly become clear that this design is very convenient for large local models. After all, the GPU sees a huge
Speaker A
amount of memory and can utilize it. If a Mac has 64, 128, or several hundred gigabytes of unified memory, a significant portion of this memory can be utilized by the GPU. This allows the Mac to run models that simply don't fit
Speaker A
in the video memory of a typical consumer graphics card. After all, a Mac with 128 GB of unified memory can run a significantly larger model entirely. Its GPU may also perform slower calculations, but the model still fits. A whole new Mac use case
Speaker A
quickly developed around this. People began buying Mac Minis and Mac Studios specifically as compact machines for local LLM. MLX versions of popular Apple Silicon models appeared: LM Studio, Lama, CPP, and other tools.
Speaker A
When a single machine's memory isn't enough, enthusiasts combine several Mac Minis into a cluster and distribute the memory across them. Of course, the memory of several regular Mac Minis isn't physically combined into a single memory stick. Here, the software
Speaker A
divides the memory between the nodes and transfers the data over the network. With the new Mac Studio, Apple has gone further. The M5 Ultra supports clusters via Thunderbolt 5 and RDMA.
Speaker A
Apple explicitly describes this as a shared memory pool across multiple systems for distributed AI. In other words, the architecture Apple developed, not for the sake of having GPT chat on your desk, unexpectedly turns out...
Speaker A
there are limitations. Large memory alone doesn't solve everything. For AI, software is crucial. And here, Apple has its own Metal MLX ecosystem and proprietary tools. For popular LLMs, this is already quite sufficient, but most professional and iStack software
Speaker A
has historically been built around a different platform, around KUDA. And here Nvidia enters the picture. It had a practically reversed situation. The software ecosystem is excellent. Koda has been the standard for a huge number of tools, libraries, and research
Speaker A
software for years. But consumer graphics cards are limited by memory capacity. It's a rather strange situation. The hardware is incredibly fast, the software is perfect, but a larger model simply doesn't fit in the memory, and Nvidia decided to address
Speaker A
this very limitation. That's how the DJX Spark was born. We even made a separate video about it. It's a computer approximately 1515 cm in size and weighs 1.2 kg. That's comparable to a Mac Mini, but Nvidia doesn't market
Speaker A
it as a Mac Mini. They bill it as a " little-little AI supercomputer." What makes it so super? Well, for example, it has a GB10 Grace Blackwall chip inside. This is a twenty-core ARM Grace processor and a Blackwall CPU. They are
Speaker A
connected via Envlink C2C and work with 128 GB of combined LP DDR5X memory. Memory bandwidth is 273 GB per second.
Speaker A
And this beauty currently costs just over $ 4,500. This is already slightly cheaper than a graphics card. Logically , this is very similar to what Apple is doing with its Unified Memory. The CPU and GPU see one large pool. Only here,
Speaker A
NVIDIA built the Skuda Tenzer RT and the rest of the company's legacy memory around it. And the need for Spark is clearly demonstrated by the same GPT OSS 120B. In the public Lama CPP1 benchmark, DGX Park produces approximately 55-60 tokens per second
Speaker A
of generation with small and medium contexts. As context increases, the speed decreases, so this is a benchmark rather than a universal figure, but it's still quite powerful. A regular graphics card wouldn't be able to handle this. And from this, it quickly
Speaker A
becomes clear why Spark was created at all. DJX Spark is a rather specific box for developers and researchers. And Nvidia positions it as a solution for developing and inferring models with up to 200 billion parameters. In other words, it addresses the scenario where
Speaker A
a single consumer graphics card no longer has enough memory, and it's still too early to upgrade to a full-fledged server. But it's hardly a mainstream home PC for $ 4,500.
Speaker A
Moreover, it's impossible to run regular tasks on it. Hmm, what if a separate box isn't needed at all? What if your regular work laptop gets this architecture on its own? IF recently took place, and before that, Nvidia's presentation, and the world of EPC has
Speaker A
changed again. We've already covered the chip itself in detail, but at the exhibition, we showed the first devices already built on NVIDIA RTX Park.
Speaker A
Simply put, RTX Park is an attempt to shoehorn the DJX Park idea into a regular computer. It's the same approach with Grace Blackwall, with a large shared memory pool, but now in laptops and compact Windows PCs. In other words, it's no longer a separate
Speaker A
box for developers and researchers, but a fully-fledged computer capable of running regular tasks. Learn more about this in our video about SRPK. And the top-end version has up to 128 GB of shared LPDDR 5X memory, a twenty-core Blackwall GPU with 614 cores, plus up
Speaker A
to one petaflop of FP4 performance for AI. You get the idea—that's power. Enough about the numbers. Let's look at the devices we showed. For example, the thin and powerful Asus Pro P16 or the Mini PK from Asus. There's also the
Speaker A
Yoga Pro 9 from Lenovo or this compact RTX Park system from Acer. NVIDIA also claims to support systems from Dell, HP , Microsoft Surface, MSI, and other partners. Basically, there's plenty to choose from. It's up to you to decide
Speaker A
what to run LLM on at home. I ran it directly on my laptop. As for hardware, like MacMi, Mac Studio, or Spark laptops, look for them on BigGeek when they become available. They always have the most up-to-date products with a
Speaker A
warranty and delivery, and you can get a discount using the Droid code. Link in the description. There's one problem , though. No one has really tested these Spark laptops yet. ASUS, for example, announced 120GB models and local 4K video generation. For now,
Speaker A
we'll leave that in the "yes, we believe it" category. Nvidia itself promises the RTX Spark PC will go on sale in October. We don't have a complete picture of performance, heat, and battery life yet, but I think it
Speaker A
will be available soon, including on Android devices. And there's another unusual thing here. Grace is an ARM processor, meaning RTX Park changes several familiar things at once.
Speaker A
Instead of X86, it's ARM. Instead of dedicated video memory, there's a large shared pool. AMD, however, is taking a much more familiar route. It keeps the standard X86 Ryzen processor, but heavily boosts the integrated graphics and gives it a huge chunk of shared
Speaker A
memory. And it turns out to be a completely different answer to the same problem. Take a look. Ryzen AI Max Plus 395, aka Strix Halo. At first glance, it's just another AMD mobile processor.
Speaker A
Inside, there are 16 Zen5 cores and a Redeon 8060S with 40 compute units. But the main focus is again on the memory.
Speaker A
It's the same 128 GB. And now let's look at where they put this thing. Yes, it's a tablet. It's last year's Rockflow Z13. A very strange device, even without any AI. That is, in shape, it's almost a Surface, but the hardware
Speaker A
inside is already a small workstation. And yes, as you can see, there's 128 GB of memory. Yes, 128 GB of memory in a tablet. I don't know who needs a tablet like this, to be honest, but it's a
Speaker A
really cool demonstration that a powerful AI machine can be built in such a compact case. The desktop framework is even more interesting. The box is only 4.5 liters in volume. The 128 GB version allows you to dedicate up to 96 GB of the integrated Redeon to
Speaker A
graphics memory, and in Linux you can go even further. On this machine, we ran the already familiar GPT OSS 120B VM Studio without any problems and got about 38 tokens per second. This means a 120 billion parameter model runs on
Speaker A
an X86 computer without a discrete GPU. That's already impressive. AMD's approach is evolving even further. The new Ryzen AI Max Plus Pro 495 supports 192 GB of total memory, of which up to 160 GB can be allocated to graphics.
Speaker A
AMD is talking about running three billion models in four-bit quantization . So, the memory race is literally happening before our eyes. 32 GB of memory on a powerful gaming GPU, 128 GB in VStrix Halo, 192 GB in the new
Speaker A
Gorgon Halo, 512 GB in the new Mac Studio. Yes, everything is very interesting with OMD memory, but then comes the software. And here we need to explain two acronyms. NVIDIA has one, we already mentioned it. It's a software platform through which
Speaker A
applications use Nvidia GPUs for calculations. Over the years, a huge AI ecosystem has grown around it.
Speaker A
PtoOrchge, libraries for training and optimization, ready-made kernels, and developer tools. AMD has its own equivalent, called Rockm. It's AMD's open computing platform for working with RON and Instinct accelerators. It gives AI frameworks access to GPUs, supports libraries for matrix computing
Speaker A
, and so on. Basically, it's Nvidia's world, and Rockm is AMD's world. And a few years ago, this was AMD's weakest point. A lot of software was primarily written for this purpose. The situation is much better now, but differences
Speaker A
still remain. For example, on Windows, AMD itself warns that the entire stack isn't supported yet. Many features are still only available on Linux. But overall, for running the popular LLM through LM Studio, the difference between Nvidia and AMD may not be
Speaker A
significant: download the model, select the GPU, and go. However, with some new or experimental models, things will be more complicated. And don't forget that AMD offers a lot of memory in a compact X86 machine. 192 GB and off we go. But
Speaker A
what if even 192 GB of memory isn't enough? AMD's answer is simple: buy another PPC and connect them. Yes, the idea of clusters itself is nothing new.
Speaker A
What's more interesting is something else. Previously, such a home cluster was more of an experiment for enthusiasts, but now manufacturers are actually designing compact PCs with this kind of scaling in mind. And the Ryzen AI Max Plus turned out to be
Speaker A
almost the ideal unit for such a construction. One framework desktop has 128 GB of memory. Take four such computers, and you get almost 480 GB of available GPU memory. AMD did just that . They even published a guide. Four
Speaker A
framework desktops on Ryzen AI Max Plus 395 were connected via a regular 5 Gbps network and Kimi K2 was launched via the Lama CPP. This is a trillion-parameter-class model, and a specific quantized version takes up 375 GB. How does it work? Let's take a look
Speaker A
. One machine controls the process, and the model is split into parts upon loading. Yes, there are transmission delays, but it's still interesting. And in 2026, Geekom showed a more complete version of this idea. They took four A9
Speaker A
megahertz Razen EI Max + 395, Ron 8060S , and 128 GB of memory, and connected them via USB 4 at 40 Gbps. According to Geekom itself, this system can run Deeps V4 flash and delivers about 14 tokens per second. The speed, frankly,
Speaker A
isn't record-breaking, but the very fact that a model of this size runs locally on four commercial mini PCs is interesting. And the flexibility of this approach is, of course, amazing.
Speaker A
We don't even need any additional interfaces for connection. So, for example, today you buy one box, run out of memory, add a second, then a third.
Speaker A
And from the application side, everything can be configured as a single Open AI-compatible IPi. The program sends a request and doesn't even need to know that the model is currently distributed across four computers. This raises the next logical
Speaker A
question. If a local AI can be distributed across several boxes, and we'll still communicate with it over the network via IP, then why should it be inside your main computer at all?
Speaker A
What if you remove the AI from your main computer altogether? Take the Mini Sform N5 Max. On the outside, there are five HDD bays, another five for SSDs with up to 200 TB of storage, and two 10-gigabit Ethernet ports. Inside,
Speaker A
there's the same Ryzen AI Max Plus 395 and up to 128 GB of shared LP DDR R5X memory. And this design has one specific purpose. Let's imagine, for example, a local assistant. It indexes the entire document archive, builds a
Speaker A
backbone for work files, searches photos by content, recognizes videos, launches an agent, and accepts requests from a laptop and phone. But a laptop alone isn't very convenient for this.
Speaker A
You close the lid and leave. And ours is already at home 24/7, connected to the network. It stores all the data the AI needs to work with. Mini Sorum builds Mini Scloud OS directly around this. Currently, it features a local
Speaker A
assistant, image and video search, Docker, and virtual machines. The company promises to add search for all file types by the end of Q3 2020. So, the software is currently evolving behind the hardware, but the feature itself looks very reasonable. So, what
Speaker A
should you ultimately choose for the future? If you plan to work with smaller models that fit into 32 GB of VRAM, then a classic PC with one RTX 5090 or several older 3090s is sufficient. This is the optimal option.
Speaker A
You'll get maximum performance and versatility, although video cards are currently expensive. A Mac is worth considering if you're comfortable with Mac OS and the software you need works with LM Studio, MLX, Metal, and so on.
Speaker A
There's a wide selection of hardware, from the smallest Mac Mini to the largest Mac Studio with 512 GB of shared memory. The maximum configuration will be expensive, but it will be capable of handling trillion-dollar models. If you need
Speaker A
Windows or Linux, AMD is probably the most interesting option right now. The Ryzen i Max Plus 395 offers up to 128 GB of shared memory paired with a classic X86 processor, meaning you get the best of both worlds. You can play
Speaker A
games and run simulations on such a PC. And Ryzen-based computers are generally cheaper than similar solutions from Nvidia. But if you need it and can't do without it, then you have no choice.
Speaker A
Sorry. Buy RTX Park as soon as it becomes available, or DJX Park, if only for neural networks. Basically, everything Nvidia offers. And that's all for today. Thanks for watching.
Speaker A
This is Droider. This was Valery Esteshev. And see you in the future. เฮ
Topics:AI PCunified memorylocal LLMApple SiliconNVIDIA DGX SparkAMD Ryzen AIGPT OSS 120BVRAMAI computingmulti-agent systems
