Tom's Hardware benchmarks RTX Spark on how long local agents take to finish a task
Tom's Hardware tested Nvidia's RTX Spark in a 128GB Surface Laptop Ultra against an M5 Max MacBook Pro on agent task time. Its headline calls local AI a strength and gaming a weak spot, but the text reviewed here ends before the AI timings.
The Engineer · Build desk
What happened
- With no mobile Ryzen AI Max+ 395 system in its labs, Tom's Hardware used AMD's compact Ryzen AI Halo as the AMD comparison.
- The test uses Artificial Analysis's AgentPerf-local, which replays inference traces across several agentic AI tasks.
- In Geekbench, the RTX Spark's single-core and multi-core CPU scores greatly trail Apple's M5 Max.
Why it matters
- What it limits constraint The RTX Spark and MacBook ran DFlash2 speculative decoding while the Ryzen AI Halo ran built-in MTP, so each system got its own best setup. A team reproducing the numbers has to match that stack.
- Who's exposed exposure The AMD figures come from a different machine than the mobile Ryzen AI Max+ 395 a laptop buyer would get, so they say little about AMD laptops.
- Who has to decide decision Battery life, build and display quality are withheld until separate per-laptop reviews, so a purchase decision for a specific model has to wait.
- What doesn't add up contradiction The CPU trails Apple's chip, yet the headline rates local AI a strength. The AI timings that would reconcile the two are not in the text reviewed.
The most reusable part of this review is the method. Tom's Hardware's past local-AI tests scored prompt processing and token generation separately, on the view that a system fast at both gives the best experience [10]. An agent finishes a job over many inference turns, and the outlet says that split does not capture the total time [10]. So the review reports time to completion, the way it would for a CPU or content-creation test [16]. It also calls local AI benchmarking a work in progress and asks readers for feedback on the approach [15]. In our view, that is the right unit for anyone sizing hardware for agents.
For the timings to transfer to your decision, several things have to hold. Your agent traces have to resemble the replayed ones. Your models have to sit near a 27B dense model or a 35B mixture-of-experts model with 3B active parameters [11]. Unsloth's Q4_K_M quantization has to meet your quality bar [13]. And your runtime has to be llama.cpp built from source for CUDA, Metal or ROCm [12]. A team on a packaged runtime would be measuring something this test did not.
The CPU result sits awkwardly beside the headline. Tom's Hardware says the MediaTek Arm CPU complex may be the platform's weakest performance point, though the system does not feel pokey in regular use [17]. Whatever strength the headline refers to has to come from somewhere other than Geekbench.
Gaming has the thinnest support. The only gaming statement in the text reviewed is the headline's own: gaming "isn't its strong suit" [14]. The passage breaks off mid-sentence in the Qwen 3.8 27B discussion, before any AgentPerf-local results appear [1].
The review also holds back anything tied to one laptop. The outlet calls the launch embargoes "a bit peculiar" and says battery life, build and display quality will come in separate reviews of other RTX Spark laptops [5]. The platform pairs the GB10 Superchip with a Windows on Arm build tailored for secure, contained execution of local AI agents [3]. Nvidia CEO Jensen Huang announced it at Computex 2026 with the ambition to "reinvent the PC" [2].
What to watch
- The AgentPerf-local time-to-completion results for Qwen 3.8-27B and the Qwen 3.6 35B A3B model on all three systems.
- Tom's Hardware's per-laptop RTX Spark reviews with battery life, build and display results.
- Any gaming benchmarks that back the headline's verdict on gaming.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence55
- Adoption
- Not scored
- Hype gap+10
- Incentives
- Not scored
- Confidence50
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
The review text available breaks off mid-sentence in the Qwen 3.8 27B discussion, before any AgentPerf-local results are shown.
- [2]
Nvidia CEO Jensen Huang announced the RTX Spark platform at Computex 2026 with the grand ambition to "reinvent the PC".
- [3]
RTX Spark pairs the GB10 Superchip with a version of Windows on Arm tailored for the secure and contained execution of local AI agents, with partners Microsoft, Dell, HP, MSI, Lenovo and Asus.
- [4]
The Surface Laptop Ultra tested has the full RTX Spark Superchip with 20 CPU cores and 6144 CUDA cores alongside 128GB of LPDDR5X-9400 RAM.
- [5]
Tom's Hardware calls the RTX Spark launch embargoes "a bit peculiar": it shares broad experience and competitive performance but not system-specific results such as battery life, build or display quality, and detailed reviews of other RTX Spark laptops in its labs are coming soon.
- [6]
The first comparison system is a 14-inch MacBook Pro with Apple's M5 Max SoC, a 40-core GPU, an 18-core CPU and 128 GB of LPDDR5X RAM.
- [7]
Tom's Hardware had no mobile Ryzen AI Max+ 395 system in its labs, so it used AMD's compact Ryzen AI Halo, with the latest ROCm stack and llama.cpp builds, as the AMD data source.
- [8]
In Geekbench, both single-core and multi-core CPU performance of the RTX Spark greatly trail Apple's M5 Max.
- [9]
Artificial Analysis developed a benchmark, AgentPerf-local, that replays inference traces across several agentic AI tasks.
- [10]
Tom's Hardware's past testing exercised prompt processing and token generation separately; in an agentic workflow that technique does not capture the total time to complete a task across all the inference turns needed.
- [11]
Tom's Hardware ran AgentPerf-local with two models: the dense Qwen 3.8-27B and the Qwen 3.6 35B A3B mixture-of-experts model.
- [12]
The inference backend was the latest llama.cpp build, built from source on each test system to use CUDA, Metal or ROCm acceleration as appropriate.
- [13]
For Qwen 3.8 27B, Tom's Hardware used Unsloth's Q4_K_M quantization, with DFlash2 speculative decoding on the RTX Spark and MacBook Pro and the model's built-in MTP on the Ryzen AI Halo.
- [14]
Tom's Hardware's headline says Nvidia's Windows on Arm platform impresses in local AI, but gaming isn't its strong suit.
- [15]
Tom's Hardware calls local AI benchmarking a continuous work in progress and asks readers for feedback on the approach.
- [16]
Instead of raw time-to-first-token and tokens-per-second figures, Tom's Hardware reports time-to-completion values, as it would for traditional CPU and content creation benchmarks.
- [17]
Tom's Hardware says the MediaTek Arm CPU complex may be the Spark's weakest performance point, though the platform does not feel pokey in regular use.
Get this ledger inside ChatGPT. Ask about any of our stories and the answer brings its figure, each claim’s verdict and the sources. Free, no account.
Sources
1 independent publisher whose own reporting we read for this story.
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.