A year ago, putting 128GB of unified memory into a box roughly the size of a stack of books sounded a bit silly. Now NVIDIA and AMD both sell one. NVIDIA wants $4,699 for a 4TB DGX Spark. AMD asks $3,999 for the 2TB Ryzen AI Halo developer platform. Both companies say their tiny machines can work with local AI models up to 200 billion parameters.
That makes the buying decision look easy. Save $700 and buy AMD, or pay extra and get NVIDIA. I wish it was that simple. I had to read these spec pages two times because both companies use very different performance numbers, very different software stories, and very different ideas of what a desktop AI computer should be. Put the headline specs next to each other and they look almost equal. Start using them for real work and the gap changes depending on what you are trying to do.
My short answer is this. If I wanted a small machine mainly for local LLM inference, coding agents, image models, normal desktop work, and the freedom to use Windows or Linux, I would look very hard at Ryzen AI Halo. If my work already lives inside CUDA, TensorRT LLM, NVIDIA NIM, PyTorch on NVIDIA, or I expect the same project to move to NVIDIA servers later, I would pay more for DGX Spark. The expensive machine is not automatically the better machine. That is where this gets fun.
These two boxes look similar until you read the small print
DGX Spark uses NVIDIA’s GB10 Grace Blackwell Superchip. The CPU side is a 20 core Arm processor with ten Cortex X925 cores and ten Cortex A725 cores. The Blackwell GPU has 6,144 CUDA cores and fifth generation Tensor Cores. NVIDIA gives the system 128GB of LPDDR5X unified memory with 273GB per second of memory bandwidth. The box measures 150 by 150 by 50.5mm and weighs around 1.2kg. It runs DGX OS, which is based on Ubuntu. NVIDIA currently sells the 4TB model for $4,699 in the US.
AMD took another route. Ryzen AI Halo uses the Ryzen AI Max+ 395, a 16 core and 32 thread Zen 5 x86 processor. Graphics come from the integrated Radeon 8060S with 40 RDNA 3.5 compute units. There is also an XDNA 2 NPU rated up to 50 TOPS. The developer platform has 128GB of LPDDR5X memory at 256GB per second, a 2TB SSD, 10Gb Ethernet, Wi Fi 7, and Windows 11 or Linux support. Its dimensions are 150 by 150 by 45.4mm, so it is even slightly shorter than Spark. AMD lists the current retail price as $3,999.
Pretty close, right?
No. The processors are built around different software worlds. Spark is Arm plus Blackwell plus CUDA. Halo is x86 plus Radeon plus ROCm, with an NPU sitting beside it. That architecture choice affects the boring stuff nobody puts in the hero image: whether an old dependency has an Arm build, whether your container assumes x86, whether a CUDA only library needs replacing, whether Windows matters to you, and whether your model runner has good support for the hardware today.
And that boring stuff can waste an entire afternoon. I know the spec sheet is more fun to discuss about, but dependency trouble is what you actually end up staring at when a demo refuses to start.
128GB unified memory is the reason both machines exist
Forget the petaflop headline for a minute. The 128GB memory pool is the part I care about most.
Consumer GPUs are fast, but VRAM gets expensive very quickly. A 24GB or 32GB graphics card can run plenty of useful local models, especially with quantization, but bigger models and longer contexts eat memory fast. Once you want 70B class models with room left for context and runtime overhead, normal gaming hardware starts becoming awkward. You can split a model across devices, offload parts to system RAM, or rent cloud GPUs. All of those options work. None feels as clean as putting the model in one shared memory pool.
Both NVIDIA and AMD now claim support for local models up to 200 billion parameters on these 128GB systems. Read that sentence carefully. “Supports 200B” does not mean every 200B model will fit at every precision, run at a good token rate, or leave enough space for a huge context window. A 200B model at 4 bits is already around 100GB just for raw weights before you count runtime overhead. Different quantization formats change this. Mixture of experts models change the compute pattern too. Context length can make things ugly. So the 200B number is a ceiling claim, not a promise that a 200B model will feel fast.
NVIDIA gives Spark a small memory bandwidth lead, 273GB per second against AMD’s 256GB per second. That gap is around 6.6 percent. I would not choose a $4,000 machine from that number alone. Real token speed also depends on the model, precision, backend, kernel quality, prompt length, batch size, and how mature the software path is for that exact workload. This is why cross vendor AI benchmark charts make me nervous. Change two settings and you can get a different story.
Still, 128GB changes what you can try locally. You stop asking, “Can I squeeze this into VRAM?” quite so often. You start asking whether the model is fast enough to be useful. That is a better problem to have.
AMD has a very awkward benchmark for NVIDIA
AMD did something smart with the Ryzen AI Halo launch. It tested the box directly against DGX Spark instead of comparing it only with a laptop.
On AMD’s current Ryzen AI Halo page, the company reports higher token rates than DGX Spark on four selected LLM workloads. AMD claims a 14 percent lead on GLM 4.7 Flash 30B A3B, 7 percent on GPT OSS 120B, 12 percent on Qwen 3.5 122B A10B, and 4 percent on Qwen 3.6 35B A3B. The same page uses retail prices of $3,999 for Halo and $4,699 for Spark, so AMD also claims better tokens per second per dollar.
That is a nasty chart if you are NVIDIA.
But I would not buy Halo because of that chart alone. AMD ran the tests. The Halo system was a preproduction developer platform. The tests used the software available on May 6, 2026. More important, AMD says tokens per second were measured with a 100 token context. That is a very short context for the kind of large local models people buy 128GB machines to run. Longer prompts, bigger KV caches, different quantization, or another inference engine can move the numbers.
So AMD deserves credit for publishing direct numbers. It also deserves the normal vendor benchmark warning label. NVIDIA marketing gets the same treatment from me.
NVIDIA’s own performance material shows why Spark cannot be reduced to one LLM chart. NVIDIA has published local fine tuning results for Llama models, image generation tests, and TensorRT LLM workloads. The company says Spark can fine tune models up to 70 billion parameters and run inference on models up to 200 billion. It also has a very specific FP4 path on Blackwell. If your workload can use those Tensor Cores well, the machine is doing something different from an AMD box running another backend at another precision.
Raw AI compute numbers make this even messier. NVIDIA advertises up to 1 PFLOP of FP4 AI compute with sparsity. AMD advertises up to 60 FP16 TFLOPS from the GPU and up to 50 TOPS from the NPU. Those are different data types and different measurement conditions. Putting “1 PFLOP vs 60 TFLOPS” in a thumbnail would get clicks, but it would be a useless comparison.
I would rather know my token rate on the exact model I plan to use.
CUDA is the reason DGX Spark can still justify $4,699
This is where NVIDIA earns part of the extra money.
DGX Spark comes with NVIDIA’s AI software stack already set up around DGX OS. NVIDIA lists PyTorch, TensorRT LLM, NIM, CUDA libraries, and its broader AI tooling as part of the intended workflow. If your code already expects CUDA, that matters more than a few percent on a vendor token chart. You are buying compatibility with a software path your team may already know.
Spark also looks more natural if the desktop box is the first stage of a project that later moves to bigger NVIDIA hardware. You can test locally, then send heavier work to NVIDIA based cloud or data center systems without changing the whole mental model of your stack. That does not mean every workload moves with zero work. It means the tools and deployment direction are familiar.
There is a catch, and it is not a small one. DGX Spark is Arm64. NVIDIA has an entire porting guide for moving x86 Linux applications to Spark. The company points out instruction set differences, memory ordering differences, compiler changes, and the need to check dependencies. NVIDIA even documents cross compilation from x86 systems to the Arm based Spark target. If everything you run comes from containers and libraries that already publish Arm builds, you may barely notice. If you depend on some old x86 binary that nobody has touched in four years, congratulations, you have found your weekend project.
I think reviewers sometimes skip this because “CUDA support” sounds like the end of the software discussion. It isn’t. CUDA support can be excellent and an Arm dependency can still ruin your day.
Spark is also Linux only. DGX OS is based on Ubuntu, and that is fine for a lot of AI work. For someone who wants the same box to serve as a Windows workstation, run Windows only apps, or sit under a normal desktop setup, Halo has an easier pitch.
And Spark has one hardware trick AMD’s tiny box does not match in the same way: ConnectX 7 networking. NVIDIA says two Spark systems can be connected to work with models up to 405 billion parameters. The hardware guide lists two QSFP connections and ConnectX 7, while NVIDIA’s multi node material goes much further with distributed workloads. If you think you may buy a second box later, this is not a decorative port.
Ryzen AI Halo feels more like a computer you can live with
AMD’s advantage starts with x86.
That sounds boring because x86 has been around forever. Here boring is useful. The Ryzen AI Max+ 395 gives you 16 Zen 5 cores and 32 threads, so Halo looks less like a special appliance and more like a very dense workstation that happens to have 128GB of unified memory. AMD supports Windows 11 and Linux. It names common local AI tools such as PyTorch, vLLM, llama.cpp, Ollama, ComfyUI, and LM Studio in its current developer material. For people coming from a normal PC, that is an easy idea to understand.
Basically, you can use it as a PC.
That matters for independent developers, creators, small teams, and people who don’t want a second computer beside their AI box. You can run normal x86 desktop software, use familiar local model apps, keep your files on the same machine, and still have enough shared memory to play with models that are impossible on a typical laptop. The 2TB SSD is smaller than the 4TB DGX Spark NVIDIA currently sells directly, but storage is not the reason I would choose either box.
The Radeon 8060S is the part doing most of the heavy local AI work. AMD also includes an XDNA 2 NPU with up to 50 TOPS, which gives Halo another route for workloads designed for the NPU. I would not overvalue that today for large LLM inference. The GPU and memory pool are the bigger attraction for this class of model. The NPU becomes more useful when apps are built to send smaller always on AI jobs there instead of waking up the bigger GPU path.
ROCm has also moved forward on Ryzen hardware. AMD’s current ROCm documentation lists official production support for the Ryzen AI Max+ 395 on Linux, and ROCm 7.2.1 brings updated PyTorch support on Windows. But here’s the annoying bit, because every platform has one: AMD’s own limitations page says only PyTorch is currently available on Windows and the rest of the ROCm stack is still Linux only. So “Windows support” should not be read as “everything you can do on Linux works identically on Windows.”
I spent too long trying to make sense of that wording because AMD’s product page sounds broader than the ROCm limitation page. Both can be true. The hardware supports Windows, several AI tools work on Windows, and the complete ROCm experience is still better on Linux. If you are buying Halo for serious development, I would probably use Linux unless a Windows app is part of the reason for buying it.
AMD has another current detail hanging over this comparison. Its Ryzen AI Halo page says a Ryzen AI Max+ PRO 495 version with support for 192GB memory is coming. There is no reason to pretend that does not matter. A 192GB option would give local model users more room than either 128GB system discussed here. Price and shipping details decide whether it is a smart buy, and those are not the point of this current comparison. Still, anybody about to spend $3,999 should know it exists on AMD’s roadmap.
The $700 gap is real, but the software tax can be bigger
A $700 saving sounds big until you are paying an engineer to fight the wrong stack for two weeks.
This is the part I would use to make the decision. Start with the software you need, then buy the hardware. If your project uses TensorRT LLM, NIM, custom CUDA kernels, NVIDIA containers, or a deployment target built around NVIDIA GPUs, buy Spark. The extra $700 is easy to defend because changing all of that just to save money on the box would be silly.
If your workflow is llama.cpp, Ollama, LM Studio, local coding agents, ComfyUI, PyTorch that already works on ROCm, or x86 software mixed with AI work, Halo makes more sense to me. You pay less. You get Windows as an option. You still get 128GB of shared memory. AMD’s own LLM tests suggest the Radeon path is not automatically slower than Spark on large quantized models, though I would want to test my own models before spending four grand.
Which brings me to cloud cost. Both companies sell these machines partly on the idea that local inference can reduce API or rented GPU bills. That can be true, but the math gets abused. A local box has electricity cost, idle time, maintenance, model setup work, and a big payment on day one. Cloud has usage fees, but you can rent far more compute for a short burst and then shut it down. If you only run a big model for two hours every Friday, buying either box is probably a weird financial decision.
For daily local work, privacy sensitive data, constant agent runs, or repeated testing where token bills keep climbing, the calculation changes. AMD even publishes a modeled three year cost claim against cloud API use. I would treat that as a scenario, not a universal saving. Your usage pattern decides it.
And there is another cost nobody writes on the product page: your time. A machine that is 8 percent slower but works with your existing code on the first day can easily be the cheaper machine.
Who should actually buy DGX Spark?
I would buy DGX Spark for a small AI team already living in NVIDIA software.
It makes sense for developers prototyping models that will later run on NVIDIA infrastructure. It makes sense for people who need CUDA first software, want NVIDIA’s local NIM and TensorRT LLM path, or care about connecting multiple Spark units. It also makes sense if local fine tuning is a serious part of the job and the published NVIDIA recipes match what you plan to do.
There is another buyer: somebody who wants an appliance. Spark is not trying very hard to be your normal PC. It is a compact DGX box. Turn it on, stay in the NVIDIA software world, and use it for AI work. For a lab this can be nicer than a custom workstation with a giant GPU, a huge case, driver choices, and whatever janky power adapter situation you built under the desk.
But I would not buy Spark just because the product page says “1 PFLOP.” I would not buy it because 200B looks good in a headline either. The Arm CPU needs consideration, Linux only can be a deal breaker, and 273GB per second memory bandwidth is not magic. A huge model can fit and still be slower than you hoped.
That last point matters. Capacity and speed are different things.
Who should buy Ryzen AI Halo?
Halo is the one I would recommend to more independent local AI users, assuming the software they need is supported.
The reasons are pretty plain. It costs $700 less at current listed retail prices. It uses a familiar x86 CPU. It gives you Windows or Linux. It has the same 128GB memory capacity, almost the same memory bandwidth, and AMD says it supports models up to 200B parameters. You also get 16 Zen 5 CPU cores, a 40 CU Radeon 8060S, and an NPU in a box that is smaller than a normal mini PC workstation setup with external GPU hardware.
So why wouldn’t everybody buy it? ROCm is the answer I would check first. Support is far better than the old jokes about AMD AI software suggest, but it still doesn’t mean every CUDA focused GitHub repo will run after changing one install command. Windows ROCm also has limits today. Some AI libraries still arrive with NVIDIA as the first path. If your work depends on one of them, $700 stops looking like much money.
For someone using common local model runners, though, Halo is difficult to ignore. AMD launched the developer platform through Micro Center in July 2026, so this is not some prototype board people are waiting to see in two years. It is a current product. That timing also means the software is still moving quickly, and some of the experience six months from now will probably be different from what buyers see in August.
If it were my $4,000, I would buy based on one boring question
What software do I already use every week?
If the answer contains CUDA over and over, I would buy DGX Spark and stop trying to save $700. I would be paying for a stack I already trust, Blackwell features, NVIDIA’s tooling, and a path toward larger NVIDIA systems. The Arm issue would be something I check before purchase, especially for old binaries or custom dependencies.
If the answer is Ollama, llama.cpp, LM Studio, normal x86 apps, local coding agents, ComfyUI, and a mix of Windows and Linux work, Ryzen AI Halo would be my first pick. I like that it behaves more like a normal computer. I like the lower price. I also like that AMD is putting direct pressure on DGX Spark instead of pretending the machines belong in separate categories.
Neither machine is a cheap toy. Four thousand dollars buys a serious desktop, a lot of cloud tokens, or months of rented GPU time depending on what you use. I would not recommend buying one because local AI is trendy. Buy one when you have a workload that keeps hitting memory limits or a cloud bill that has become annoying enough to measure.
The part I think matters six months from now
The interesting fight is no longer only NVIDIA GPU versus AMD GPU. These boxes are turning local AI into a complete computer category.
NVIDIA is betting that developers want a tiny version of its data center world on the desk. AMD is betting that developers want a normal x86 machine with a very large shared memory pool and enough GPU compute to keep serious models local. Both ideas make sense. I lean toward AMD for a personal local AI workstation and toward NVIDIA for a team already built around CUDA.
And the category is moving fast. AMD is already talking about a 192GB Halo configuration. NVIDIA can pair two Sparks for much larger models. Software updates on both sides are changing token rates after the hardware has shipped. A benchmark from May can already be old by August.
That is why I would not crown a permanent winner here.
For the machine you can buy today, Ryzen AI Halo is the better value for a broad local AI setup. DGX Spark is the safer buy when NVIDIA compatibility is part of the job itself. The wrong choice is buying either one from a spec sheet without checking the exact model runner, libraries, operating system, and deployment path you use.
Before spending $3,999 or $4,699, install your stack on hardware from the same family if you can. Run the actual model. Use your actual context length. Check the token rate you can live with.
That boring test will tell you more than a petaflop badge ever will.