AMD Ryzen AI Max+ 395 vs Nvidia DGX Spark: Which to Buy

Nvidia's DGX Spark and AMD's Ryzen AI Max+ 395 both cram enough unified memory into a shoebox-sized desktop to run large language models locally, but they get there in very different ways and at very different prices: DGX Spark pairs a 20-core Arm/Blackwell superchip with up to 128GB of LPDDR5x and a starting price of $4,999 for the new 64GB model, while Ryzen AI Max+ 395 mini PCs like the Framework Desktop pair a 16-core Zen 5 CPU with up to 128GB of LPDDR5x for as little as $1,959. For most people building a local-AI box on a budget, the Ryzen AI Max+ 395 is the better buy; for CUDA-dependent developers who need Nvidia's software stack and the fastest possible token throughput, DGX Spark justifies its premium.

What is the AMD Ryzen AI Max+ 395?

The Ryzen AI Max+ 395 is AMD's top "Strix Halo" chip, a single piece of silicon that combines 16 Zen 5 CPU cores (32 threads) with a 40-compute-unit RDNA 3.5 integrated GPU (branded Radeon 8060S) and an XDNA 2 neural processing unit, all built on a 4nm process. According to AMD's own product page, the chip has a default TDP of 55W with a configurable range of 45-120W, a 64MB L3 cache, and crucially a 256-bit LPDDR5x-8000 memory interface shared across the CPU, GPU and NPU, good for 256GB/s of bandwidth and up to 128GB of total capacity. AMD rates the NPU alone at up to 50 TOPS and the full platform (CPU+GPU+NPU combined) at up to 126 TOPS, a figure AMD also discussed in its own launch blog post for the chip.

AMD doesn't sell finished desktops itself; it sells the chip to OEM partners. The best-known implementation is the Framework Desktop, a Mini-ITX board in a roughly 4.5-liter case, but GMKtec, Asus (ROG Flow Z13 and desktop variants), HP and others also ship Ryzen AI Max+ hardware. Because the memory is unified and soldered rather than added as discrete VRAM, a Ryzen AI Max+ 395 machine can dedicate most of its RAM to a single large model, the same trick that makes Apple Silicon and DGX Spark attractive for local inference.

What is Nvidia DGX Spark?

DGX Spark is Nvidia's own "personal AI supercomputer," built around the GB10 Grace Blackwell Superchip: a 20-core Arm CPU (10 Cortex-X925 performance cores plus 10 Cortex-A725 efficiency cores) paired on-package with a Blackwell-generation GPU carrying 5th-generation Tensor Cores and 4th-generation RT Cores. Per Nvidia's own product page, DGX Spark delivers up to 1 petaFLOP of AI compute at FP4 precision, ships with either 64GB or 128GB of LPDDR5X coherent unified memory at 273GB/s of bandwidth, up to 4TB of self-encrypting NVMe storage, and networking that includes a 200Gbps ConnectX-7 NIC, 10GbE, Wi-Fi 7 and Bluetooth 5.4. It runs Nvidia's own DGX OS (an Ubuntu-based Linux distribution) with the full CUDA, cuDNN, TensorRT and NIM stack preinstalled, and the whole unit measures 150 x 150 x 50.5mm and draws up to 240W.

Two units can be linked over a single QSFP cable using what Nvidia calls Sync Cluster Assistant, pooling memory and letting a pair of 64GB Sparks act as a 128GB system with roughly 1.7x the inference throughput of one unit, according to Nvidia's own breakdown of the newer 64GB configuration.

Spec-by-spec comparison

The table below lines up the chip-level specs from AMD's and Nvidia's own product pages, plus Framework's configurator for a concrete AMD-based system, since AMD doesn't sell a complete desktop itself.

Memory and bandwidth: the real bottleneck

For local LLM inference, the two numbers that matter most are how much unified memory you get and how fast the system can move data through it, because token-generation speed on these architectures is usually bandwidth-bound rather than compute-bound. On paper, DGX Spark's 273GB/s edges out the Ryzen AI Max+ 395's 256GB/s by about 7%, a real but modest advantage. Both chips cap out at 128GB of total addressable unified memory on their largest configurations, enough to hold a roughly 70-billion-parameter model at 8-bit quantization with headroom for context, or a ~200-billion-parameter model at heavier quantization.

Where the two diverge is price-per-gigabyte: Framework's 128GB Ryzen AI Max+ 395 board lists at $3,449, while a 128GB DGX Spark MSRP'd at $4,699 before climbing further at retail as Nvidia and the rest of the industry absorbed a sharper DRAM price run-up through late 2026. That gap alone will decide the purchase for a lot of buyers.

Price and availability

Those Framework prices are live from the company's own configurator as of this week and are themselves up from the $1,599/$1,999 launch pricing Framework announced in 2025, which tells you the DRAM shortage squeezing Nvidia is hitting AMD's OEM partners too. Nvidia's own developer forum confirmed the Founders Edition price change in February, attributing it to "industry wide memory supply constraints," and the company's announcement of the 64GB model in October described the same dynamic shaping its decision to split DGX Spark into a cheaper 64GB tier and a pricier 128GB one rather than raise the price of a single SKU further.

AI performance: TOPS vs. petaFLOPS, explained

It's tempting to put "126 TOPS" and "1 petaFLOP" side by side and declare a winner, but the numbers aren't measuring the same thing. AMD's "Overall TOPS" figure is an INT8 throughput estimate that adds together the CPU, the Radeon 8060S GPU and the dedicated XDNA 2 NPU; the NPU alone is rated at up to 50 TOPS and is mainly used for low-power, always-on tasks like Windows Studio Effects rather than running a full LLM. The heavy lifting for large local models on a Ryzen AI Max+ 395 machine actually happens on the Radeon 8060S GPU through ROCm or Vulkan, not the NPU.

Nvidia's "1 petaFLOP" figure, by contrast, is a GPU Tensor Core number at FP4 precision with structured sparsity, a format specifically tuned for transformer-based LLM inference and training, and it's generated entirely by the Blackwell GPU portion of the GB10 die, with CUDA, TensorRT and Nvidia's NIM microservices built to exploit it directly out of the box. In practice this means DGX Spark's software stack is purpose-built for feeding that compute figure, while a Ryzen AI Max+ 395 box depends more on how well a given inference runtime (llama.cpp's Vulkan backend, ROCm, ONNX Runtime) has been tuned for RDNA 3.5. Neither company has published directly comparable third-party benchmarks as of this writing, so treat the headline numbers as architectural context rather than a scoreboard.

Which is better for local LLM and on-device AI work?

If your workflow lives inside Nvidia's ecosystem, DGX Spark is the easier, more supported choice: CUDA is still the default target for most open-source inference and fine-tuning frameworks, Nvidia ships a complete, tested software image, and the ConnectX-7 networking makes clustering multiple units for bigger models straightforward. That support and the extra 17GB/s of bandwidth come at a real cost, roughly $1,300-$1,500 more than an equivalently sized Ryzen AI Max+ 395 machine at today's prices.

If your priority is the most unified memory per dollar, more flexibility on OS choice, and you're comfortable running inference through Vulkan or ROCm rather than CUDA, the Ryzen AI Max+ 395 is hard to beat. A 128GB Framework Desktop costs about 30% less than the new 64GB DGX Spark while offering twice the memory, though with somewhat lower bandwidth and a less mature AI software stack than Nvidia's.

Who should buy which

• Buy Ryzen AI Max+ 395 if: you want the most memory for the least money, you're fine building your own inference stack, or you also want a capable general-purpose desktop (the Radeon 8060S handles 1440p gaming reasonably well, something DGX Spark isn't built for).
• Buy DGX Spark if: you're already working in CUDA/PyTorch/TensorRT, you need Nvidia's guaranteed software compatibility and support, or you plan to cluster multiple units for larger models.
• Consider neither if: your models fit comfortably in 24-32GB of VRAM, where a single discrete GPU like those compared in our RTX 5070 Ti vs. RX 9070 XT breakdown will likely run faster per dollar than either unified-memory box.

Frequently asked questions

See the FAQ section below the article for quick answers on pricing, memory, and which machine suits which workload.

What's next

Both products are moving targets right now. Nvidia's 64GB DGX Spark doesn't actually ship until October 23, 2026, so real-world benchmarks and reviewer impressions are still to come; AMD's OEM partners, meanwhile, are juggling the same DRAM shortage that pushed Nvidia to raise prices twice in eight months, so expect Framework and its rivals' pricing to keep moving too. Anyone buying either machine today should budget for further price volatility through the rest of 2026 and check current stock directly with the maker before ordering, since configurations have gone in and out of availability for both product lines this year.

Written byMara Lindqvist

Mara Lindqvist edits the hardware desk. She covers graphics cards, processors, memory and storage, the foundries and chip designers behind them, and what the numbers on a spec sheet mean for people choosing a PC. Specifications in her stories come from manufacturer spec pages and datasheets.

More from Mara LindqvistAll Hardware coverage