The number that stopped me was not 300 billion parameters. It was 160GB of configurable graphics memory.
GMKtec has just shown the EVO-X5 Pro at IFA 2026 with AMD’s new Ryzen AI Max+ PRO 495. The machine supports 192GB of unified LPDDR5X memory, and GMKtec says up to 160GB of that pool can be assigned to the Radeon 8065S graphics. Put that beside an RTX 5090 with 32GB of GDDR7 and the comparison looks almost silly. Five times the usable GPU memory, inside a compact desktop.
Then I looked at the other number.
The EVO-X5 Pro has about 273GB/s of memory bandwidth. The RTX 5090 has 1,792GB/s. So the NVIDIA card has roughly 6.6 times the bandwidth even though it has one fifth of the memory capacity.
That is the whole story in two numbers. The GMKtec can potentially hold models a 5090 simply cannot keep in VRAM. The 5090 can move the models it does hold much, much faster. Local AI hardware is now splitting into two different ideas of what a “fast” machine should be.
And this is where things get interesting.
Access without medium partner: Mini PC Now Give Its GPU 160GB Memory
160GB sounds like VRAM, but it is not RTX 5090 style VRAM
GMKtec’s wording is easy to misread. I had to read the launch material twice because “160GB VRAM” sounds like someone somehow put five RTX 5090 memory pools inside a mini PC.
That is not what happened.
The Ryzen AI Max+ PRO 495 uses a 256-bit unified memory architecture. CPU and GPU share the same LPDDR5X pool instead of having normal system RAM on one side and dedicated graphics memory on another. AMD officially supports up to 192GB on this processor, and GMKtec lets as much as 160GB be configured for graphics use.
This matters a lot for local LLMs because the GPU does not hit the same hard 32GB wall that a 5090 does. A large model can live inside one shared memory pool without being pushed across PCIe into normal system RAM every time the dedicated GPU runs out of space.
But the memory itself is LPDDR5X-8533, not GDDR7 sitting directly beside a monster discrete GPU. It is wide for an integrated design, yet it is still built around a very different power and performance target.
So yes, 160GB is real in the sense that the GPU can use a huge chunk of the unified pool.
No, it is not the same thing as putting 160GB of RTX-class GDDR7 on the Radeon 8065S.
That difference is going to confuse a lot of buyers.
The RTX 5090 loses the capacity fight badly
NVIDIA’s flagship consumer card has 32GB of GDDR7. For gaming, that is already a lot. For local AI, 32GB can feel strangely small the moment model size starts climbing.
A 7B model is easy. A 14B model is easy. Many 30B-class models fit comfortably with sensible quantization. Once you get into larger dense models, big multimodal models, long context, or several models running together, 32GB becomes the thing you start planning around.
This is why the RTX 3090 stayed relevant for so long. Local AI users cared about its 24GB long after gamers had moved on. Memory capacity changes what you can load, and loading the entire model into accelerator memory can matter more than another gaming benchmark.
The EVO-X5 Pro pushes that logic to an extreme.
With 160GB available to graphics, the machine can hold models that would require several consumer NVIDIA cards, heavy CPU offload, or a professional GPU with a very different price tag.
That is a real advantage.
The interesting part is that GMKtec is doing it inside a compact x86 system with 16 Zen 5 CPU cores, Radeon 8065S graphics, an XDNA 2 NPU, up to 24TB of storage, USB4, and enterprise management features. AMD rates the Ryzen AI Max+ PRO 495 itself from 45W to 120W configurable TDP. An RTX 5090 alone is rated at 575W total graphics power.
Power figures are not directly comparable because one is a whole processor and the other is a discrete GPU, and the complete GMKtec system will obviously draw more than the CPU package alone. Still, the physical direction is clear. One design is trying to make a huge shared memory workstation compact. The other is trying to move data as fast as possible through a high-end GPU.
The RTX 5090 wins the bandwidth fight by an absurd margin
Here is the part that stops the “160GB destroys 32GB” headline from becoming nonsense.
The RTX 5090’s memory bandwidth is 1,792GB/s.
The EVO-X5 Pro is around 273GB/s.
That is not a small gap. It is roughly 6.6 times.
For single-user LLM generation, memory bandwidth often matters because the system repeatedly needs to read model weights while producing tokens. This is especially true for dense models at batch size one, where the workload can become more about moving weights than doing fancy math.
So imagine two models.
The first is a 30B or 32B model that fits completely inside the RTX 5090’s 32GB. The second machine can also load it easily.
Which one would I expect to feel faster?
The RTX 5090, by a lot, assuming the software path is good on both systems.
The 5090 does not only have far more bandwidth. It also has NVIDIA’s fifth-generation Tensor Cores, mature CUDA tooling, heavily optimized inference kernels, and an ecosystem that most AI software still treats as the default target.
The GMKtec’s 160GB advantage does almost nothing if the model already fits comfortably inside 32GB.
Capacity only helps when capacity is the problem.
That sentence sounds obvious, but local AI hardware marketing keeps making us forget it.
A 300B model is where the GMKtec starts making sense
GMKtec says the EVO-X5 Pro is designed to run LLMs with up to 300 billion parameters fully offline.
That claim is the reason this machine will get attention.
It also needs a giant asterisk.
Take the simplest possible memory calculation. A dense 300B model stored at exactly 4-bit quantization needs around 150GB just for raw weights. Real quantized formats add scales, metadata, alignment, and other overhead, so the actual file can be larger than the clean four-bit math suggests.
Now remember that the machine can assign up to 160GB to graphics.
Suddenly the famous “300B” claim looks much tighter. A 4-bit dense model can eat almost the entire graphics allocation before KV cache, runtime buffers, context, the operating system, or other services get involved.
That does not mean the claim is fake.
It means the model matters.
A 300B Mixture of Experts model can have hundreds of billions of total parameters while activating only a much smaller set for each token. A more aggressive three-bit or two-bit quantization changes the memory math again. Some runtimes can also use unified memory in ways that do not map neatly to a fixed “VRAM” carve-out.
This is exactly why parameter count by itself is becoming a bad hardware benchmark.
“Runs a 300B model” tells me something about capacity.
It tells me almost nothing about speed.
The simple bandwidth math gets uncomfortable very quickly
Here is a rough way to think about a dense model.
If a quantized 300B model occupies around 150GB and the system has 273GB/s of memory bandwidth, a naive bandwidth-only calculation gives you less than two complete weight passes per second. Real inference is more complicated than this, of course. Kernels, caching, model architecture, quantization format, active experts, batch size, and software efficiency all change the result.
Still, the mental model is useful.
A machine can technically hold an enormous model and still generate text slowly.
That is why I want to see independent benchmarks before repeating GMKtec’s 300B claim as if it means “300B at desktop-chat speed.” As of this writing, GMKtec has publicly announced the machine and its specifications, but I could not verify an independent shipping-unit benchmark that shows a specific 300B model, quantization, context length, and sustained tokens per second.
That missing benchmark matters more than the launch slide.
If the 300B model is sparse and activates a small fraction of its total weights, performance can be far better than the dense-model example above. If it is dense and aggressively quantized, quality may change. If it needs large context, memory use climbs again.
There are too many moving parts to turn “300B” into one performance number.
I would rather see five boring benchmark rows than one giant parameter count.
The NPU is not the reason this machine can run huge LLMs
The Ryzen AI Max+ PRO 495 includes an XDNA 2 NPU rated at up to 55 TOPS. AMD lists total processor AI capability up to 131 TOPS when different engines are counted together.
Those numbers will appear everywhere.
For the giant local LLM story, they are not the main reason the EVO-X5 Pro is interesting.
The big reason is memory.
A 300B model does not become possible because an NPU has a nice TOPS number. It becomes possible because the system can physically keep an enormous amount of model data close enough to the compute engines to use it.
This is the same mistake the AI PC market has been making for two years. It keeps selling accelerator TOPS when local LLM users are asking a more basic question: how much model can I keep in fast memory?
GMKtec finally has an answer that sounds ridiculous.
Up to 160GB for graphics.
That is much more interesting to me than 55 NPU TOPS.
But AMD still has a software problem
Hardware is only half of local AI.
NVIDIA owns the other half better than anyone.
CUDA is still the path that gets first support in a lot of AI software. New kernels, serving frameworks, quantization tools, research code, and performance work often arrive on NVIDIA before anything else. Developers know the environment. Documentation is everywhere. If something breaks, somebody has probably already opened the issue.
AMD has improved a lot. ROCm support is better, llama.cpp gives AMD users several useful paths, Vulkan works well for many local inference setups, and Windows support has become less painful than it used to be.
Less painful is not the same as CUDA.
A 160GB memory pool is useless if the exact workload you need expects NVIDIA-only kernels. This is where people buying these high-memory AMD boxes need to be careful. The specification can be perfect and the software can still decide the purchase.
I think this is the biggest reason the RTX 5090 keeps winning recommendations even when its 32GB looks weak beside these new unified-memory systems.
NVIDIA gives you less room.
It gives you fewer surprises.
The EVO-X5 Pro is not really an RTX 5090 replacement
The title comparison is useful because the numbers are so different.
The actual products solve different problems.
If you run a 14B, 30B, or 32B model every day and care about fast token generation, CUDA compatibility, image generation, GPU rendering, and gaming, an RTX 5090 workstation makes much more sense. You are paying for bandwidth and compute, and the model fits.
If you want to experiment with 70B, 100B, 200B, or very large MoE models without building a multi-GPU tower, the EVO-X5 Pro becomes much more interesting. You accept lower bandwidth in exchange for a huge shared memory pool and a much smaller physical system.
And if your workload runs several local agents at once, the 192GB total memory becomes useful in another way. You do not have to spend all 160GB on one giant model. You can keep a model loaded while giving the CPU plenty of memory for databases, vector stores, containers, browsers, and agent processes.
That may be a better use of the hardware than chasing the biggest model you can barely fit.
A 300B screenshot is fun.
A 70B or 100B model with lots of headroom may be the reason to actually buy the machine.
The 192GB number may matter more than the 160GB number
There is another way to look at this system.
Forget the “VRAM” marketing for a minute.
The EVO-X5 Pro has one large 192GB pool shared across CPU and GPU. That means local AI software can work with data without the same rigid split between system memory and discrete graphics memory.
Traditional PC builds can have 128GB or 192GB of DDR5 plus a powerful NVIDIA GPU. The problem is the boundary between them. Once model data spills out of dedicated VRAM, performance can fall sharply because the GPU starts pulling data across PCIe from system RAM.
Unified memory avoids that specific boundary.
It does not remove the bandwidth problem.
That trade is important.
AMD is basically saying: what if the GPU gets access to almost all the memory, even if that memory is much slower than top-end GDDR7?
For local AI, that is a very reasonable question.
Apple has already shown why people care about the same idea. Mac Studio systems became popular with some local LLM users because large unified memory pools let them load models that would otherwise need several GPUs.
Now AMD is bringing that memory philosophy into x86 machines that can run normal Windows and Linux software.
This could be a bigger deal than the EVO-X5 Pro itself.
Mini PCs are becoming memory machines
Mini PCs used to compete on size and price.
Now they are competing on how absurd a model they can load.
That shift happened fast.
First 64GB looked large. Then 96GB and 128GB unified-memory systems became the interesting local AI tier. Now AMD’s new platform supports 192GB, and vendors are immediately turning 160GB graphics allocations into the headline.
This is the opposite of the gaming PC market.
Gaming keeps asking for more GPU speed.
Local AI keeps asking for more memory first, then speed.
The perfect machine would obviously give us both. Something like 160GB of memory with RTX 5090-class bandwidth and CUDA support would solve a lot of arguments.
It would also cost a small fortune.
So buyers are choosing which compromise hurts less.
The RTX 5090 says: fit a smaller model and run it fast.
The EVO-X5 Pro says: fit a much larger model and accept that memory bandwidth is not in the same league.
Neither answer is stupid.
The price could decide whether any of this matters
GMKtec has announced the EVO-X5 Pro, but final retail pricing was not listed in the launch material I checked.
That number is going to matter a lot.
At around $2,000, a 192GB Ryzen AI Max+ PRO 495 machine would be extremely hard to ignore for local AI users. At $3,000, people start comparing it more seriously with used multi-GPU systems, Mac Studio configurations, and other Strix Halo machines. At $4,000 or more, specialized NVIDIA systems enter the conversation.
High-capacity LPDDR5X is not cheap right now. Other 128GB and 192GB systems have already shown how quickly memory pushes these compact AI computers into workstation pricing.
So I am not calling the EVO-X5 Pro a bargain before GMKtec gives us the number.
The hardware idea is good.
The value is still unknown.
What I would benchmark first
I do not want another chart showing NPU TOPS.
I want the boring local AI tests.
Load a 32B model at the same quantization on the EVO-X5 Pro and an RTX 5090. Show prompt processing, generation speed, power draw, and memory use. Then move to 70B. Then 120B or 150B. Finally, show the exact 300B model GMKtec used for its claim, including quantization, context length, active parameter count if it is MoE, and sustained tokens per second after the prompt is loaded.
That would answer almost everything.
I also want to know how ROCm and Vulkan compare on the same machine, whether 160GB graphics allocation creates any ugly Windows behavior, and how performance changes when the CPU and GPU are both fighting over the unified pool.
This part is still unknown.
And that is okay. The product was only just unveiled at IFA. The problem starts when a capacity claim gets repeated online as a speed claim before anyone has done the boring testing.
Five times the memory does not mean five times the computer
The EVO-X5 Pro is one of the most interesting local AI machines announced this year because it attacks the problem consumer GPUs keep refusing to solve.
Memory capacity.
An RTX 5090 gives you 32GB. GMKtec says its little desktop can give Radeon 8065S graphics up to 160GB from a 192GB unified pool. That is a five-times capacity advantage on paper.
Then NVIDIA hits back with 1,792GB/s of bandwidth versus 273GB/s.
Now the comparison makes sense.
The RTX 5090 is the machine I would want when my model fits and I care about speed. The EVO-X5 Pro is the machine I would want to test when the model does not fit at all.
That second case is becoming more common.
Open models are getting larger. Agents are keeping more state. Context windows are growing. Multimodal models need extra memory. People want to run several services locally instead of paying an API for every call.
Suddenly 32GB can feel like a strange limit on a flagship GPU.
GMKtec and AMD are showing the other extreme. Give the GPU a huge memory pool, accept lower bandwidth, and let software decide how useful that trade really is.
I do not think 160GB makes the RTX 5090 obsolete.
I think it makes the RTX 5090’s 32GB look much harder to ignore.
And for local AI hardware, that may be the more important change.