What Dell has actually proven
StorageReview's UL Procyon AI Text Generation tests put the Pro Max 18 Plus, with its RTX PRO 5000 Blackwell (24GB GDDR7), at 185.1 tokens per second on Phi, 140.5 on Mistral and 119.7 on Llama 3. Time to first token was under 0.35 seconds on every model. StorageReview calls it the fastest laptop it has tested. That is a claim about one outlet's lab, not a universal ranking.
The price of that speed is portability. StorageReview lists the machine at 7.17 pounds with 3 hours 39 minutes of measured battery life, the shortest runtime in its laptop dataset. Those battery figures come from PCMark 10 Modern Office runs. Local inference is a much heavier sustained load, so plan on wall power for real sessions.
TechRadar says Dell's top configuration cost $9,245.52 at launch and now sells for almost $18,600. I could not corroborate that price history independently. StorageReview lists the tested system at $9,245 at review time, but treat the later figures as TechRadar's reporting.
What RTX Spark is, and what it claims
The platform pairs a 20-core Arm-based Grace CPU with a Blackwell RTX GPU and up to 128GB of unified memory in a single package. Nvidia's headline claims are:
- Up to 1 petaflop of FP4 AI performance, 6,144 CUDA cores and up to 128 GB of shared memory.
- Local LLMs of about 120 billion parameters with a one-million-token context window.
These are vendor claims. Treat "128GB" as "up to". The line-up is not uniform:
- Laptops come in two variants, a 5,120-core GPU with an 18-core CPU capped at 64GB, and a 6,144-core GPU with a 20-core CPU and up to 128GB. Both are rated at 45 to 80W.
- Dell's XPS 16 Creator Edition starts at $3,799.99 for preorder, with the 18-core chip, 5,120-core GPU and 32GB of unified memory.
- Some base Surface and ProArt models offer just 24GB.
The petaflop figure needs a caveat. Microsoft's Surface footnotes describe it as theoretical FP4 performance using sparsity. TechRadar likewise calls it an upper limit that everyday workloads may not reach. It is not comparable to a tokens-per-second score.
Capacity versus speed
Nvidia's case rests on memory. Dell's discrete GPU has 24GB of VRAM. Spark's pool is shared with the CPU, so larger models can load at all. StorageReview's own explainer makes the distinction: fit determines whether a model runs; the silicon determines how fast. Its leaderboard shows the same trade. The HP ZBook Ultra G1a, with up to 96GB of assignable unified memory, ran DeepSeek-R1 70B but trailed every RTX PRO machine on Procyon speed.
Spark's published numbers lean the same way:
- Nvidia's Windows-on-Arm porting guide lists a 256-bit LPDDR5 interface with 300 GB/s of bandwidth. TechRadar sets that against 896 GB/s for Dell's RTX PRO 5000.
- TechRadar also notes Spark's 6,144 CUDA cores against Dell's 10,496.
- Token generation is typically bandwidth-sensitive. That is general industry knowledge rather than a measured Spark result, but it is why the 300 GB/s figure matters.
The sensible expectation is that Spark may win on model size and efficiency, not on raw speed with models that fit in Dell's 24GB. That is an inference, not a finding.
What to watch after October 16
StorageReview says three of the first Spark systems are confirmed for its lab, including the Dell XPS 16 Creator Edition and the HP OmniBook Ultra 16. It adds that they enter the rankings when its numbers are in. It also notes a Microsoft chart putting a 137-billion-parameter model at roughly 60 tokens per second on a Surface Laptop Ultra. StorageReview describes that as a vendor result it has not reproduced.
For a fair verdict, look for:
- The same Procyon text-generation suite on Phi, Mistral and Llama 3, with matching software.
- Which configuration was tested. A 24GB or 32GB unit is a different proposition from a 128GB one.
- Larger-model results, such as 70B-class models, where Dell's 24GB cannot compete.
- Battery life under inference load, not just office benchmarks.
- Windows on Arm compatibility for your tools. Nvidia's porting guide exists because software needs to be ported.
Bottom line
Dell holds the measured lead as of October 13. Spark's larger unified-memory ceiling could make bigger local models practical in a lighter, cooler machine. Nothing published yet shows it beating Dell's token rates. For IT buyers, the decision between the two is about which models you actually run, not which logo is on the lid.
Update: Reported 24GB model could broaden RTX Spark’s large-model options (October 14, 2026)
According to Tech Times, Nvidia’s October 13 software announcements included Nemotron 3.5 Bolt 74B, a model scheduled for release on October 15 that reportedly fits in 24GB of memory. The outlet says it combines 2-bit quantization-aware training with a mixture-of-experts architecture that activates 9 billion of its 74 billion parameters per token. Launch support reportedly includes llama.cpp, Ollama and LM Studio.
That introduces an important qualification to the capacity comparison: a 70B-class parameter count does not necessarily require Spark’s highest-memory configuration. If the reported footprint holds in practical testing, lower-memory systems could gain another large-model option. It does not establish equivalent reasoning quality, usable context length or faster inference, however; buyers still need measured results and clarity about memory available to the model.
Tech Times also reports that Nvidia’s PAIR router reached general availability with RTX Spark integration. It reportedly distributes complete inference jobs across compatible networked machines rather than pooling their GPU memory. For Windows users with multiple supported PCs, that could improve parallel workflow throughput—not make one oversized model fit across those machines.
Update: Launch-event hands-on adds pricing detail, but no comparable AI benchmarks (October 15, 2026)
According to Gizmodo’s October 15 hands-on report from an Nvidia-hosted event, the Surface Laptop Ultra starts at $2,600 with 24GB of unified memory, while a 128GB configuration with 1TB of storage costs $5,900. The outlet also reports a $2,800 starting price for the Asus ProArt P16. Those figures sharpen the purchasing distinction: entry-level RTX Spark pricing does not buy the memory capacity behind its largest-model claims.
Gizmodo also reports that Nvidia presented efficiency claims positioning Spark against a discrete RTX 5070 laptop GPU at around 80W of combined CPU and GPU power, including for AI prompt processing. These remain vendor claims, not independent measurements establishing a lead over Dell’s Pro Max 18 Plus.
The hands-on gaming demonstration offered limited additional evidence: according to Gizmodo, its writer played Gears of War: E-Day at medium and high settings, but only with DLSS 4.5 and frame generation enabled. The report supplies no comparable local-AI token-generation results. For buyers awaiting the October 16 launch, the new information clarifies configuration costs without resolving the article’s central performance question.
Update: Independent RTX Spark AI benchmarks arrive, but Dell comparison remains open (October 15, 2026)
According to HotHardware’s October 15 review, independent RTX Spark N1X benchmarks are now available, superseding this article’s earlier statement that performance evidence was limited to vendor claims. The outlet tested a top-end Surface Laptop Ultra with 20 CPU cores, 6,144 GPU cores and 128GB of unified memory.
HotHardware reportedly found the Surface ahead of every system in its Procyon AI Text Generation comparison across Phi 3.5, Mistral 7B, Llama 2 and Llama 3.1. Its chart, however, compares Spark with Intel, AMD and Qualcomm integrated-graphics systems—not Dell’s Pro Max 18 Plus or its RTX PRO 5000. The chart also reports overall benchmark scores rather than the tokens-per-second figures cited above. These results therefore establish independent evidence of Spark’s AI performance, but do not establish that it beats Dell.
HotHardware also reports 14 hours and 35 minutes of battery life for the tested Surface in its PCMark 10 battery test. That is not a sustained local-inference runtime measurement and should not be treated as one.
For Windows AI buyers, the evidence has moved beyond demonstrations and theoretical throughput. The remaining purchasing question is narrower: how Spark performs against discrete RTX PRO hardware under matched models, software and test settings.
Update: Early compatibility testing exposes limits to Spark’s “everything runs” claim (October 15, 2026)
According to The Verge’s October 15 hands-on report, Nvidia says RTX Spark already supports the anti-cheat software used by Fortnite, Roblox, Rust, Rainbow Six Siege and MapleStory. Support for other titles, including Call of Duty, Valorant, League of Legends and Apex Legends, is promised for later—not confirmed as available at launch. That distinction qualifies Nvidia’s broader claim that all apps and games will run on Spark.
The Verge also reports performance limitations when Microsoft’s Prism emulator translates x86 software for Arm. Its reviewers reportedly could not push Horizon Zero Dawn Remastered or Assassin’s Creed Shadows above 30fps, although the outlet describes CPU translation overhead as a likely explanation rather than a proven diagnosis. Conversely, it reports smooth 1080p gameplay in Cyberpunk 2077, without supplying a directly comparable workstation benchmark.
For buyers considering Spark as both a local-AI workstation and an everyday Windows laptop, these findings add a compatibility caveat: software launching successfully does not establish equivalent performance, and promised anti-cheat support should not be treated as shipping support. The report does not resolve the outstanding AI-performance comparison with Dell’s Pro Max 18 Plus.
Update: Spark token-rate results add a concrete comparison with Dell (October 15, 2026)
According to Tom’s Guide’s October 15 review, its Surface Laptop Ultra with RTX Spark produced 74.4 tokens per second on Phi 3.5, 54.2 on Mistral 7B and 50.9 on Llama 3.1 in UL Procyon AI Text Generation testing. Reported time to first token was 0.28, 0.49 and 0.44 seconds respectively. These measurements add the token-rate evidence missing from the earlier HotHardware comparison.
Spark’s reported rates are numerically below all three Dell results cited above, but this is not yet a controlled head-to-head verdict. The outlets’ model labels differ, and the new report does not establish matching model versions, quantization, software or test settings. It therefore does not demonstrate that Spark beats Dell—or establish a precise performance gap.
Tom’s Guide also reports 10 hours and 15 minutes in its web-browsing battery test, versus 1 hour and 38 minutes in PCMark 10’s gaming battery test. Neither measures sustained local inference. For Windows AI buyers, the new results strengthen the distinction between memory capacity and generation speed: Spark now has published token rates, but its larger-model advantage still needs workload-specific testing rather than inference from memory size alone.
Update: RTX Spark laptops reportedly go on sale with native development tools (October 16, 2026)
According to Windows Report’s October 16 launch coverage, RTX Spark laptops are now available globally from six manufacturers, starting at $2,599, with configurations and availability varying by region. This moves the platform from preorders and review samples into reported retail availability; buyers should still check the memory capacity of the specific configuration rather than assume the advertised 128GB maximum.
Windows Report also reports native CUDA, PyTorch-CUDA for Windows on Arm, Windows Subsystem for Linux compatibility, and native Python and Conda tools. The outlet attributes a claim of compatibility with 97% of the top 15,000 Python packages to Nvidia. For developers, that is a more specific software-support commitment, but not proof that every dependency in an existing AI workflow works unchanged.
For creative workflows, Photoshop and Premiere reportedly receive native RTX Spark support through launch-day beta releases, according to Windows Report, while Substance support remains forthcoming. These additions broaden the reported launch software offering, but the source supplies no new matched AI benchmarks against Dell’s Pro Max 18 Plus.
Update: PCMag tests put 64GB RTX Spark behind M5 Max in local AI speed (October 17, 2026)
According to PCMag’s October 16 first tests, a 64GB Asus ProArt P16 RTX Spark generated text at roughly half the speed of its 128GB Apple MacBook Pro M5 Max comparison system. This adds independent measurements from another Spark laptop, although it does not resolve the direct comparison with Dell’s Pro Max 18 Plus.
In MLPerf Client 1.6.1, PCMag reportedly measured 50 tokens per second on Llama 3.1 8B Instruct for the Asus versus 105 for the MacBook, and 28 versus 58 on Phi 4 Reasoning 14B. Both machines delivered their first token in under a second. PCMag says testing with the newer, Arm64-capable MLPerf Client 2.0.1 preserved Apple’s approximate two-to-one generation-speed lead; the Asus also outpaced HP’s ZBook Ultra G1a on the Llama workload.
These are configuration- and benchmark-specific findings, not a platform-wide ranking. PCMag notes differing memory capacities and software optimization as unresolved factors. Its MLPerf results should not be directly compared with the Procyon figures above.
For Windows buyers, the report adds another reason to evaluate actual workloads rather than maximum memory capacity alone. PCMag has not yet completed battery testing, including verification of Nvidia’s claimed performance consistency when unplugged.
Update: New Spark tests claim faster AI generation than M5 Pro (October 20, 2026)
According to Principled Technologies’ October 20 press release distributed through EIN Presswire, three 64GB RTX Spark N1X laptops reportedly outperformed a 64GB MacBook Pro with M5 Pro in selected AI workloads. The company tested Microsoft’s Surface Laptop Ultra, HP’s OmniBook Ultra and Asus’s ProArt P16, but says the devices were pre-production units and retail performance may vary.
Principled Technologies reports up to 5.7 times faster video generation in ComfyUI and up to 2.1 times faster time to first token using Qwen3.5 27B with an 8,192-token prompt in llama-bench. These claims concern specific generation tasks and initial responsiveness—not a demonstrated lead in sustained text-generation speed.
The comparison does not overturn PCMag’s earlier M5 Max findings: it uses a different Apple chip and different workloads. Nor does it settle the outstanding comparison with Dell’s Pro Max 18 Plus.
For mobile AI users, Principled Technologies also reports that Spark systems stayed within 3% of plugged-in performance during Blender and llama.cpp testing on battery. That adds an unplugged-performance claim, not a battery-runtime measurement. Buyers should treat these as results announced by the testing company through a press release, rather than independently corroborated retail benchmarks.