Performance · Evidence checked October 5, 2026
RTX Spark Benchmarks
Reported CPU scores are available, but they are not a retail-device verdict. This page separates leaked results, historical samples, vendor specifications, and hands-on impressions so you can see what each actually establishes.
CPU · Reported results, not a ranking
Geekbench 6 Scores
| Device / evidence | Test version | Single-core | Multi-core | OS / driver | Power / source |
|---|---|---|---|---|---|
| 20-core N1X; OEMQAJ 766_MIS Product Name DV Leak — OEM test-system listing, not verified retail | Geekbench 6 family; exact version Unknown in the retrieved article text | 2,570 | 23,126 | OS/build: Unknown in retrieved text GPU driver: Unknown | Power limit/mode: Unknown TweakTown: reported listings |
| N1x, HP 8EA3 motherboard; final device identity Unknown Historical — pre-release Linux sample | Geekbench 6.2.2 preview, Linux AArch64 | 3,096 | 18,837 | Ubuntu 24.04.1 LTS GPU driver: Unknown | Power limit/mode: Unknown Original Geekbench listing · Wccftech, May 31, 2026 |
TweakTown reports a second, 18-core listing but does not give exact scores in its article text. We have not reconstructed those scores from its relative comparison. Wccftech describes the historical Linux run as dating to June 2025; its 2026 coverage date is not the test date.
The different single-core and multi-core outcomes do not isolate a firmware improvement or regression: these are different samples with unmatched conditions. Geekbench measures a CPU workload suite, not gaming FPS, CUDA throughput, or the speed of every developer application. A buyer-facing comparison needs the same benchmark version, disclosed power modes, and repeat runs on named devices.
Cinebench 2026: Prototype CPU Result
| Device / evidence | Test | Single-core | Multi-core | OS / driver | Power / source |
|---|---|---|---|---|---|
| Surface Laptop Ultra pre-sample; reported 20-core RTX Spark, JMVW0A-Generic-CPU Leak — prototype screenshot | Cinebench 2026; exact build Unknown | 540 | 5,771 | Windows 11; build Unknown GPU driver: Unknown | 80W PL1 / 95W PL2 shown; OEM performance mode Unknown Notebookcheck, July 18, 2026 (updated July 19) · Screenshot source credited by Notebookcheck |
The reported 540 is the single-core score, not multi-core. Cinebench 2026 CPU rendering scores cannot be substituted for Cinebench R23 or 2024 scores, nor compared numerically with Geekbench. The displayed power limits are settings, not measured consumption throughout the run. A single screenshot does not establish sustained rendering speed, thermal stability, or production-unit performance.
Current Official RTX Spark Specifications
Vendor specification — NVIDIA labels both laptop configurations N1X. These are platform limits, not the RAM, power mode, or speed of every OEM SKU.
| Component | N1X laptop (20-core) | N1X laptop (18-core) | N1X desktop |
|---|---|---|---|
| CPU | 20-core Grace Arm CPU | 18-core Grace Arm CPU | 20-core Grace Arm CPU |
| Blackwell RTX GPU | 6,144 CUDA cores | 5,120 CUDA cores | 6,144 CUDA cores |
| Unified LPDDR5X | Up to 128 GB | Up to 64 GB | Up to 128 GB |
| Listed power (TDP) | 45–80W | 45–80W | 140W |
| Features | DLSS 5 · Reflex 2 · fourth-generation RT cores · fifth-generation Tensor cores · DirectX 12 Ultimate | ||
| Listed OS | Windows 11 | ||
NVIDIA markets the CPU as Grace. Its porting guide identifies Cortex-X925 performance and Cortex-A725 efficiency cores: 10+10 in the 20-core model, 9+9 in the 18-core model. It does not describe these as custom NVIDIA CPU cores. Unified memory capacity is not a measured memory-bandwidth or CPU–GPU interconnect result.
Sources: NVIDIA product and feature tables · NVIDIA CPU-specific porting guide. Checked October 5, 2026.
Gaming Performance: Features Are Not FPS
NVIDIA's DLSS 5, Reflex 2, and hardware ray-tracing support are feature specifications, not measured frame rates. A 6,144-core GPU does not establish universal RTX 5070-equivalent performance: power allocation, clocks, memory access, game build, and chassis cooling all matter. The 5,120-core configuration must also be tested separately.
Manufacturer demo / event hands-on
PC Gamer's Gamescom 2026 hands-on reports smooth gameplay, but says FPS was hidden, frame generation was enabled, and the writer could not run benchmarks or inspect settings. That is a useful observation of the demo experience, not a native-rendering FPS measurement or a promise that every game works.
Source: PC Gamer: Gamescom RTX Spark hands-on · More labeled reviews and hands-on coverage
Useful gaming results should report native rendered FPS separately from frame-generated output, average FPS, 1% lows, latency, resolution, ray-tracing preset, upscaling mode, and whether the executable is native Arm64 or emulated. Anti-cheat, launchers, and drivers need title-specific checks; an emulation percentage cannot be forecast for all games. See the compatibility guide.
AI Performance and Local Models
NVIDIA advertises up to 1 petaflop of FP4 AI performance and native CUDA on RTX Spark.[1] This is a vendor peak-throughput claim, not measured tokens/second. It is not an apples-to-apples ranking against INT8 NPU TOPS: precision, sparsity, operation counting, and workload differ. The official specifications cited here do not establish a separate NPU specification.
Up to 128 GB of shared memory can make larger local workloads feasible, but model weights, quantization, context/KV cache, backend overhead, and OS memory use determine whether a model fits and how fast it runs. We do not assign maximum LLM parameter counts to RTX Spark or competing laptop families from memory capacity alone. See local-AI testing requirements; measured inference results remain pending in this evidence set.
Creative App Benchmarks: What to Measure
NVIDIA lists Studio support, hardware 4:2:2 encode/decode, and AV1 encoders on its product page.[1] These capabilities do not demonstrate a fixed speedup in Premiere Pro, Photoshop, Blender, or Substance 3D. An export test needs an app version, identical project/assets, codec and output settings, acceleration backend, plugin compatibility, and elapsed time. Timeline responsiveness and AI-effect speed should be measured separately from export throughput.
No reproducible creative-app timings are included in the sources summarized here. Rather than a synthetic scorecard, we will add source-linked results when those conditions are disclosed.
Where RTX Spark Stands Against Apple, Intel, AMD and Snapdragon
The leaked CPU scores answer narrower questions than “which laptop is fastest?” A retail MacBook comparison also changes the OS and software backend; an x86 or Snapdragon comparison needs a specific chip, device, power mode, and app build. CUDA availability, game compatibility, memory capacity, sustained performance, battery runtime, and price can point to different choices. Our platform comparison explains those trade-offs without extrapolating a winner from core counts or peak AI figures.
Evidence Labels and Benchmark Methodology
What each label means
- Vendor specification: announced capabilities or limits; no measured result implied.
- Manufacturer demo / event hands-on: selected hardware and workloads; observational unless instrumentation and settings are disclosed.
- Leak / preproduction test: listing or screenshot with uncertain provenance; not verified retail hardware.
- Community report: an individual experience, not necessarily an editorial lab review.
- Retail lab review: named shipping device with disclosed methods and measured outcomes; none established by the score rows above.
Requirements for comparable results
- Record device/SKU, chip, RAM, firmware, OS/build, driver, test version, and native/emulated execution.
- Disclose AC/battery operation, OEM mode, power limits, ambient conditions, and warm-up duration.
- Repeat runs and report spread, sustained clock behavior, temperatures, noise, and measured power alongside scores.
- Use identical workloads/settings; do not mix benchmark generations or frame-generated and rendered FPS.
- Link the original result and identify who tested it. Keep unknown fields explicit rather than filling them with platform defaults.
Open Questions Before Buying
- Sustained CPU/GPU sharing: how does simultaneous load affect rendering, compilation, and gaming in different chassis?
- Battery runtime: what do matched brightness, battery capacity, and workloads show, rather than architecture-based forecasts?
- Software readiness: do required plugins, anti-cheat drivers, peripherals, and CUDA packages work on the specific Arm64 stack?
- Local AI: what are time-to-first-token, prompt-processing and decode rates, context length, memory use, and wall power for a named model/backend?
- Value: what exact configurations are shipping in your region, at what price, with what warranty?
Sources: [1] NVIDIA RTX Spark product page. Score and hands-on sources are linked directly beside the corresponding evidence above.
Device specifications · Reviews & hands-on · Compatibility · Local AI · Latest news
Last updated: . This is a curated evidence register, not an in-house test report or a comprehensive claim that no other benchmarks exist.