DGX Spark Alternative: AMD Strix Halo vs GB10 (2026)

AMD's Ryzen AI Max+ 395, better known as Strix Halo, is the DGX Spark alternative this blog hasn't tested yet, and it's overdue. Every other 128GB desk-side box gets measured against GB10: RTX 5090, Mac Studio, a half-dozen GB10-badged OEM clones. Strix Halo is different. It's not another GB10 rebrand and it's not a discrete GPU squeezed into a tiny case: it's a separate silicon family with its own unified-memory architecture, its own software stack, and, after two rounds of DRAM-driven price hikes on both sides, a price gap that's shrunk enough to make this an actual decision rather than a foregone one.

This post puts real, cited numbers next to each other: decode speed, prefill, fine-tuning viability, and price. If you want the shortlist of every desk-side 128GB box side by side, including the Mac Studio angle, DGX Spark vs Mac Studio covers that comparison directly.

TL;DR: Is AMD Strix Halo a Good DGX Spark Alternative?

Strix Halo is a credible DGX Spark alternative for decode-heavy MoE chat at roughly half the price, as of 11 Oct 2026, but GB10 still wins on prefill speed and fine-tuning turnkey-ness.

Neither a $6,950 DGX Spark nor a roughly half-price Strix Halo build is refundable after a week of testing.

Spec and Price Snapshot: GB10 vs Ryzen AI Max+ 395, Both at 128GB

Both chips hit the same 128GB ceiling through unified memory, which is the whole reason they get compared. How they get there is where they split.

AMD's Ryzen AI Max+ 395 pairs 16 Zen 5 CPU cores with a 40-compute-unit Radeon 8060S RDNA 3.5 iGPU and an XDNA2 NPU, addressing up to 128GB of LPDDR5x-8000 unified memory on a 256-bit bus rated up to 256 GB/s.

NVIDIA's own 128GB/4TB DGX Spark launched at $3,999 about a year ago and now lists at $6,950, a jump of nearly 74% in roughly twelve months.

Both the DGX Spark and Strix Halo prices above are vendor list prices as of this post's date (11 Oct 2026) and move with memory-market conditions, the same DRAM-driven pressure covered later in this post. Check each vendor's own listing before buying.

Two things matter beyond the raw spec sheet. First, GB10's power budget is shared across the whole system, not just the GPU. Our own DGX Spark thermal throttling piece measured that constraint directly on real hardware, inside DGX Spark's 240W system envelope with headroom also covering the ConnectX-7 port, SSD, and USB-C peripherals. Strix Halo mini PCs don't carry that same 200Gb/s NIC inside their power budget, which is one reason OEM boxes built around it run cooler at a given wattage.

Second, Strix Halo is not a rebadged GB10. If you're shopping the GB10 side of this market and comparing OEM boxes from Dell, ASUS, HP, and Lenovo, DGX Spark alternatives: GB10 boxes compared covers that shortlist; every one of those runs the identical GB10 silicon NVIDIA ships in DGX Spark itself. Strix Halo buyers are choosing a genuinely different chip, not a cheaper case around the same compute. And if a discrete GPU with more raw tokens-per-second matters more to you than 128GB of shared memory, DGX Spark vs RTX 5090 measures that tradeoff on models that fit inside 32GB of GDDR7.

Tokens Per Second: Prefill vs Decode, Measured

"Tokens per second" hides two very different numbers, and the gap between GB10 and Strix Halo isn't the same size in both.

Qwen3 30B-A3B MoE Decode Speed

Decode speed, how fast the box generates each new token once it's already processing your prompt, is where Strix Halo looks genuinely competitive on MoE models.

NVIDIA doesn't publish a matching decode figure for GB10 on Qwen3 30B-A3B MoE, so there's no single apples-to-apples decode benchmark across both chips on that specific model. What NVIDIA's own GB10 announcement does commit to is up to 1 petaFLOP of FP4 (sparse) compute, a figure that matters more for prefill than for decode, which is memory-bandwidth-bound on both chips regardless of which one has the bigger FLOPs number on a spec sheet. For a single interactive chat session on a mixture-of-experts model that fits comfortably in 128GB, Strix Halo's 72 tokens/sec is fast enough that most users won't notice a gap either way.

Here's the sanity check behind that 72 tokens/sec number, worked from Strix Halo's own specs rather than taken on faith. Qwen3 30B-A3B's name states its own architecture: 30B total parameters, roughly 3B active per token once MoE routing picks the experts for that step. At Q4_K_XL quantization (roughly 4 bits, or 0.5 bytes, per parameter), decoding one token means streaming about 1.5GB of active weights through memory. Divide that into Framework's measured 215 GB/s real-world bandwidth and the bandwidth-only ceiling comes out to roughly 143 tokens/sec, about double what Framework actually logged. The gap between that ceiling and the measured 72 tokens/sec is a normal real-world tax: KV cache reads, router weights, and attention layers all compete for the same memory bus that the active-expert math alone doesn't account for. A measured result landing at roughly half of a back-of-envelope bandwidth ceiling is a sign the number is physically plausible, not an outlier.

Neither number requires taking Framework's word for it, either. Rent an RTX 5090 on Spheron at $0.78/hr on-demand, pull the same Qwen3 30B-A3B MoE checkpoint at Q4_K_XL quantization, and run it through llama.cpp or Ollama: both print the decode rate after every response, in the same tokens/sec units Framework's thread reports. A single 20-minute billing block is enough time to read your own third data point off the logs instead of refereeing two vendor-adjacent benchmarks against each other, and it puts a GPU with real GDDR7 bandwidth next to both desk-side boxes for comparison. Spheron's own fleet is cloud GPU rental, not desk-side boxes, so this rent-and-log path, not an in-house Strix Halo unit, is the direct way to add a number of your own to the two cited above.

Prefill and Time-to-First-Token: Where GB10 Pulls Ahead

Prefill, how fast the box chews through your prompt before it starts generating, is a different story, and it's the story Strix Halo's own community benchmarks don't really tell. Framework's test thread reports decode numbers in detail; it doesn't publish a comparable prompt-processing figure, which is itself a signal about where consumer-GPU benchmarking communities spend their attention and where Strix Halo's architecture has less to show off.

GB10, by contrast, is built around dedicated tensor cores and up to 1 petaFLOP of FP4 (sparse) compute, the kind of dense matrix-multiply throughput that compute-bound prefill actually taxes. AMD's own spec sheet for the Ryzen AI Max+ 395 publishes memory bandwidth and compute-unit counts but no equivalent FP4 throughput figure for the 40-CU Radeon 8060S iGPU. That's CUDA's dedicated tensor-core path doing what it's built for, against an RDNA 3.5 iGPU with no equivalent published throughput at that scale. For agentic workflows and long-context prompts, where most of the latency a user actually feels is prefill rather than decode, this architectural gap matters more than either vendor's headline tokens-per-second number.

Dense 70B Models: Both Hit the Same Bandwidth Wall

Dense 70B-class models are where both boxes stop looking competitive with anything bigger, and where they converge on each other instead.

That's the bandwidth wall showing up identically on both chips: a 70B dense model at Q4 needs to stream tens of gigabytes through memory for every single token, and neither chip's unified-memory bus, 273 GB/s on GB10, roughly 215 GB/s measured on Strix Halo, is fast enough to make that comfortable. If your workload is dense 70B+ models rather than MoE, both of these boxes are a worse fit than renting a multi-GPU cloud instance with real HBM bandwidth, which is the point where this comparison stops being about which desk-side box to buy.

To make that break-even concrete: the price gap between a roughly half-price 128GB Strix Halo desktop and a $6,950 DGX Spark buys a meaningful amount of rented GPU time to test which bottleneck you'd actually hit before committing to either box. At Spheron's on-demand RTX 5090 rate of $0.78/hr as of 13 Oct 2026, that price difference alone covers dozens of hours of testing against real HBM-class memory bandwidth, enough to run both the MoE decode test and the dense-70B test above on borrowed hardware before buying anything.

CUDA vs ROCm: Where the Software Stack Actually Decides the DGX Spark Alternative Question

Raw tokens-per-second is close enough on decode that the real DGX Spark vs Strix Halo decision usually comes down to software, not silicon. Our broader ROCm vs CUDA breakdown covers the general performance and cost gap between the two ecosystems at data-center scale; the same gap shows up at desk-side scale, just with higher variance because there's less engineering attention on a 128GB mini PC than on a cluster of MI300X nodes.

Backend Roulette: Vulkan vs ROCm on Strix Halo

Strix Halo runs on multiple inference backends, Vulkan, ROCm, and in some setups a Windows-specific path, and which one you pick changes your tokens-per-second by a wide margin. The dense 70B test above is the clearest example: the same mini PC, the same model, the same quantization, went from roughly 11 tokens/sec on Vulkan to about 18 tokens/sec after ROCm tuning, a jump of more than 60%. That's not a rounding error; it's the difference between a usable local model and a frustrating one, and it depends entirely on which backend the box happens to be running and how well it's been configured.

That's the honest state of Strix Halo's software stack in 2026: getting a benchmark number out of it is achievable, but getting a stable, repeatable setup takes real configuration work that GB10's CUDA path mostly skips.

Fine-Tuning: Possible on Strix Halo, Still Experimental

Fine-tuning is where the software gap widens the most. A documented community guide for fine-tuning 27B+ models on Strix Halo lays out the exact stack required: Linux mainline, a ROCm 7.13 nightly build, PyTorch 2.11, patched versions of bitsandbytes and flash-linear-attention, and out-of-process eval orchestration to keep multi-day training runs from dying partway through. Every one of those is a specific, non-default version pin, and the guide exists precisely because the default install doesn't work reliably for this job.

That matches what the Strix Halo fine-tuning guide implies by how much setup it requires: DGX Spark's CUDA stack has fewer moving parts for training jobs specifically, even though the two boxes are close on decode-only inference. If your Strix Halo fine-tuning job outgrows a single desk-side box, whether in model size or in training time, training LLMs on AMD GPUs at data-center scale covers the ROCm stack for MI300X and MI355X, the step up from a 128GB desktop to real HBM bandwidth and multi-GPU parallelism.

When Strix Halo Wins as a DGX Spark Alternative, When DGX Spark Wins, and When to Just Rent

Buy Strix Halo if your workload is decode-heavy chat or serving on MoE models that fit in 128GB, your budget is the deciding factor, and you're comfortable doing some backend tuning to get there. The price gap, roughly half of DGX Spark's $6,950, is real money, and Framework has publicly attributed its own repeated 2026 price increases, including a jump of up to $460 on the 128GB configuration, to DRAM costs tied to AI datacenter demand, meaning that gap could close further if memory prices keep climbing on both sides.

Buy DGX Spark if prefill speed matters (long-context prompts, agentic workflows, anything latency-sensitive on the first token), if you're fine-tuning models and want a CUDA stack that doesn't require nightly ROCm builds and patched libraries to stay stable, or if you expect to cluster two units over the 200Gb/s ConnectX-7 port for a larger model than either box can hold alone. Our DGX Spark to GPU cloud pipeline guide covers the break-even math for when local dev hours on a DGX Spark stop being cheaper than renting, and how to move the same vLLM setup to a cloud instance once a workload outgrows either desk-side box.

Just rent if you're not sure yet which workload you actually have. Neither a $6,950 DGX Spark nor a Strix Halo desktop is something you can return after a week of testing. Spheron bills GPU rentals on-demand, spot, and reserved by the minute after a 20-minute minimum, aggregating live rates across 5+ providers, which means testing a fine-tuning job or an inference workload against an RTX 5090, H100, H200, or MI300X costs a few dollars rather than a multi-thousand-dollar purchase decision. Spot capacity can be reclaimed without notice, so it's a better fit for exploratory testing than for the multi-day unattended fine-tuning run that a local Strix Halo or DGX Spark box would run uninterrupted with no meter running at all. On Spheron, RTX 5090 on-demand runs $0.78/hr and H100 on-demand runs $2.65/hr as of 13 Oct 2026, both well short of either box's purchase price for a few hours of testing.

Pricing fluctuates based on GPU availability. Spheron rates above are live as of 13 Oct 2026; Framework and NVIDIA pricing reflect their most recently published list prices and may have changed. Check current GPU pricing → for live rates.

Before you commit $3,000-$7,000 to a desk-side box, a few hours of rented GPU time on the same model class tells you which bottleneck, memory, bandwidth, or software stack, you'll actually hit.

Rent an RTX 5090 on Spheron →