Microsoft Surface Laptop Ultra for Business Review (Condensed, -40% Word Count)

The Microsoft Surface Laptop Ultra for Business is the first Surface laptop powered by NVIDIA silicon and the first RTX Spark device we have benchmarked. Our review unit features the full N1X superchip: a 20-core Grace Arm CPU, 6,144-core Blackwell RTX GPU and 128GB LPDDR5X unified memory, with up to 96GB allocatable to the GPU. Consumer SKUs range from $2,599.99 (18-core, 24GB, 512GB SSD) to $5,899.99 (20-core, 128GB, 1TB SSD). Pricing for this business model was unavailable at publication.

This benchmark review evaluates local AI, CPU, memory, gaming, thermals and storage, comparing against four modern x64 laptops: Intel Panther Lake / Arrow Lake H, AMD Strix Point, plus laptops with NVIDIA RTX 5090 and RTX PRO 2000. Physical design, display, keyboard and battery assessment will come in a later chassis-focused follow-up; this piece covers platform performance and raw numbers.

The 20-core N1X dominated nearly all multi-threaded CPU benchmarks, often with a 30%+ lead. Its ability to assign 64GB memory to the NVIDIA GPU — more VRAM than any prior NVIDIA laptop GPU — enables usable-speed inference for 35B-parameter models. However, sustained all-core double-precision computation triggered three full power cuts on our test unit in one evening. An earlier sample failed identically and required replacement, making thermal and power reliability a critical concern.

Key Takeaways

• ---GPU memory adjustable from 64GB up to 96GB unlocks larger local AI models: gpt-oss:20b and Qwen3.6-35B-A3B load and generate tokens at 58–89 tok/s. The Razer Blade 16’s RTX 5090 runs 2.3× faster for models fitting inside its 24GB VRAM but cannot handle these bigger models.
• ---Top multi-core CPU in the test group: 6,963 Cinebench 2026 multi-core score, 34% ahead of the Blade 16 and 47% faster than the Dell Pro Precision 5 16S. It compiled LLVM in 206 seconds; the next-fastest competitor needed 327 seconds.
• ---122GB/s memory bandwidth comes with latency tradeoffs: STREAM Triad hit 122,114 MB/s, 48% higher than the Precision 5 16S. Yet its 251ns memory latency and 172ns inter-core latency were worst-in-class, hurting single-threaded workload performance.
• ---One heavy workload triggers power loss: all-core OpenBLAS LU factorization at N=20000 shut the machine down three times on October 14. The chip previously sustained 98°C for 31 minutes in other tests before hitting 95°C here. Two separate Surface Laptop Ultra units exhibited identical failures; logs point to power delivery or firmware protection rather than overheating.
• ---Windows on Arm limits random I/O: the Gen5 SK hynix SSD hits 14.1GB/s sequential read and 1.67M raw-device IOPS. But with NTFS and BitLocker enabled, 4K random I/O tops out near 300K IOPS, less than half the x64 systems.

RTX Spark N1X Platform Overview

RTX Spark is NVIDIA’s Windows PC platform built around the N1X superchip, a laptop and compact desktop counterpart to the GB10. NVIDIA offers two laptop N1X variants: the high-end 20-core Grace CPU paired with a 6,144-core Blackwell RTX GPU and up to 128GB LPDDR5X, and an 18-core version with a 5,120-core GPU capped at 64GB. Both share a combined 45–80W TDP for CPU+GPU; desktop N1X/GB10 run at 140W. Our Surface uses the 20-core CPU, reported by Windows as 10 performance cores + 10 efficiency cores without SMT.

Unified memory differentiates this platform from all tested x64 laptops. A single 128GB pool of LPDDR5-9400 serves both chip halves over NVLink-C2C on a 256-bit bus with ~300GB/s theoretical bandwidth. Windows exposes it similarly to discrete graphics, with fixed dedicated and shared memory pools. We configured our test unit for a 64/64GB split; benchmarks saw 62.5GB usable GPU memory, qualifying it for both our local AI benchmark tiers. Users adjust GPU memory allocation via NVIDIA’s app, limited by total system RAM selected at purchase. Smaller 32GB/48GB Surface configurations cannot run these large AI models.

The Blackwell GPU includes 5th-gen Tensor Cores, 4th-gen RT Cores, one 9th-gen NVENC encoder and one 6th-gen NVDEC decoder, with up to 1 petaflop theoretical sparsity-enabled FP4 compute. All NVIDIA-side benchmarks ran CUDA via native Arm64 llama.cpp and Ollama using the factory driver version 32.0.16.1742.

The laptop runs Windows 11 Pro 26H1 for Arm64. x64 applications run via Prism emulation. We built our benchmark suite with native Arm64 binaries to compare raw silicon performance instead of emulation translation overhead. Where mature Arm64 builds do not exist or only x64 binaries are available, we note this in results.

Microsoft Surface Laptop Ultra for Business Specifications

Test Setup & Methodology

Local AI Performance

Local AI performance hinges on two factors: model inference speed and whether the model fits into GPU memory. The Blade16’s RTX5090 delivers 2.3× higher token throughput for models fitting inside its 24GB VRAM. The Surface’s strength is supporting larger models: 64GB allocated GPU memory makes it the sole test system running our tier-M large models at usable speeds. The Pro Max14’s 8GB RTX PRO2000 matches the Surface on small models but cannot load larger ones. The two integrated x64 GPUs run all models slowly.

MLPerf Client v2.0 runs Llama3.1 8B and Phi-4-mini across five prompt categories, measuring time-to-first-token and token generation rate. Hardware acceleration stacks differ by vendor: Intel OpenVINO on Precision5 16S, AMD Windows ML DirectML on Pro7 14, llama.cpp+CUDA on NVIDIA laptops. The Blade16 hit 128.8 tok/s on Llama3.1 8B and 222.7 tok/s on Phi-4-mini. The Pro Max14’s RTX PRO2000 stays within 5–12% of the Surface for these small models. The Precision5 16S’s Arc Pro B390 starts fast but yields half the Surface’s token speed; the Pro7 14 DirectML implementation is slowest.

MLPerf agent benchmarks use Llama3.1 8B for software engineering and data analyst agents with multi-turn tool calls. Blade16 completed both agent workflows in under 28 seconds. The Pro7 14 failed both agent tasks, hitting out-of-memory errors despite 32GB reserved GPU memory, marked DNF.

Model loading tests measure Ollama cold (empty cache) and warm model load time plus SSD read throughput. Tier S uses gpt-oss:20b, tier M adds Qwen3.6-35B-A3B. Cold loads reflect storage speed: the Gen5 Precision5 16S loaded gpt-oss:20b in 5.98s, the Surface in6.57s. The Surface dominates warm reloads and tier-M testing; after model swap, reloading Qwen3.6 took only 5.0s versus 13.4s on the Precision, thanks to both models resident in 64GB GPU memory. Blade16 reads model files fastest at 3.9GB/s but incurs PCIe transfer overhead moving weights into VRAM.

Coding Agent testing uses Codex CLI on Ollama-served local models on a fixed code task, passing when 11 repository tests succeed. We report median wall time, pass count and generation speed. A concurrency test runs the coding agent alongside Speedometer web benchmark to quantify performance impact during parallel workloads. Blade16 finished tier-S coding tasks in64 seconds at137.9 tok/s, three times faster than Surface, but only passed two out of three runs; Surface and Precision passed all three tier-S trials. For tier-M workloads, the Surface finished both tasks under90 seconds, vastly quicker than the x64 integrated-GPU laptops. Concurrency exposes iGPU shared-resource limitations. Blade16’s discrete GPU isolates workloads: browser performance dropped just6% with agent active. On the Surface, agent workloads share GPU and memory bandwidth, cutting Speedometer browser score nearly in half. The coding agent itself saw minimal slowdown on every laptop.

Retrieval-Augmented Generation (RAG) test embeds 20,963 SQuAD passages, retrieves top five matching chunks for 2000 questions and answers with Phi-4-mini, scoring F1 and exact match accuracy. Identical small models run on all systems, isolating raw speed differences. Accuracy was nearly identical across devices; only latency varies. Blade16 ingested the corpus in131 seconds and answered typical queries in0.16s. Pro Max14 finished ingestion second at154s, ahead of Surface’s190s.

CPU Performance

The Surface N1X dominates nearly all multi-threaded CPU benchmarks, often by 30%+. The Dell Precision5 16S’s Core Ultra X9 388H leads single-threaded, Python, browser and OpenBLAS workloads. Razer Blade16’s Ryzen AI9 HX370 wins Stockfish and GROMACS and ranks second in rendering and compiling. Dell Pro7 14 and Pro Max14 lag all-core performance.

Cinebench2026: Surface multi-core score of6,963 beats Blade16 (+34%) and Precision5 16S (+47%). Its single-core score 558.5 narrowly outperforms Precision5 16S’s 534.8. AMD chips trail single-thread performance. A 30-minute sustained loop showed Surface maintained steady performance at 98°C, no performance degradation. Blade16 also stayed stable; Precision5 16S throttled 6.2% after roughly10 minutes.

7-Zip compression benchmark: Surface leads multi-threaded compression at151.0 GIPS, 20% ahead of Blade16. Precision5 16S leads single-threaded work. The Surface has a single-thread quirk: cores under load for several minutes drop to ~80% speed, triggered even at moderate temperatures, affecting 7-Zip, PyPerformance and Stockfish single-thread runs.

Blender4.5 Cycles CPU rendering: Surface took first place across Monster, Junkshop and Classroom scenes. It held a 47% lead on Monster and an 18% lead on the heavy Classroom scene.

HandBrake video encoding: Surface led x265 HEVC encoding at41.7fps (+25% over competitors). However its native Arm64 SVT-AV1 encode hit only15.4fps, much slower than x64 builds, caused by immature Arm64 software implementation rather than hardware limits. Precision5 16S led SVT-AV1 among x64 systems.

LLVM build test compiles full LLVM source tree: Surface finished in206.4s, 37% faster than Blade16’s 327s. Precision5 16S’s result required retesting after Smart App Control blocked the compiler.

Stockfish chess benchmark: Blade16 beat the Surface on Stockfish multi-threaded and single-threaded runs. primesieve showed Surface fastest on multi-thread prime counting; AMD CPUs led single-thread primes. Precision5 16S led Python benchmarks and Speedometer browser tests, with Pro Max14 very close behind. The Surface finished last here, compounded by its core slowdown and Arm64 Edge build overhead. Combined single-thread results lag comparable x64 laptops by15–30%, partially explained by high memory latency.

GROMACS molecular dynamics: Blade16 delivered best GROMACS performance. The Surface was slowest here; its Arm64 build lacks AVX2 optimizations present on x64 versions. OpenBLAS double-precision linear algebra: Surface failed all OpenBLAS tests (DNF). The N=20000 all-core LU solve repeatedly triggered full power loss. Among x64 laptops, Precision5 16S dominated smaller DGEMM and LU workloads, Blade16 won the largest LU solve.

Memory Bandwidth & Latency

STREAM benchmark: Surface LPDDR5-9400 reached122,114 MB/s Triad bandwidth and130,516 MB/s Copy bandwidth, 48% higher than Precision5 16S. This high bandwidth drives strong compile and rendering results, but latency is a major weakness: 251ns memory latency and 171.6ns average core-to-core latency, worst of all tested laptops. Precision5 16S achieved lowest memory latency at115.5ns; Pro7 14 had tightest inter-core mesh at49.7ns.

Gaming & GPU Encoding

SuperTuxKart gaming benchmark (1080p Ultra): native Surface run averaged235.9fps; x64 binary via Prism emulation hit264.6fps, above Blade16’s 200fps. Blade16 is limited by CPU and OpenGL driver, not RTX5090 GPU performance. Hardware HEVC encoding via NVENC/QSV/AMF: all hardware encoders deliver similar throughput around128–137fps. Surface’s single 9th-gen NVENC reached135fps. Pro Max14 returned no NVENC results due to outdated GPU driver.

Thermals & OpenBLAS Power Cut

During long workloads the Surface held 92–97°C; in 30-minute sustained Cinebench loops it maintained 98°C and stable clocks without throttling. Blade16 peaked 98–99°C and also stayed stable. Precision5 16S uses timer-based thermal reduction. The Surface’s OpenBLAS failure was not graceful thermal throttling: the machine hard power-cuts under heavy FP64 all-core load. System logs show Kernel-Power 41 events, no thermal trip records, no crash dumps. Temperatures at failure sat around95°C, after 31 minutes of stable 98°C operation in other benchmarks. The failure repeats consistently 14–16 seconds into N=20000 LU solve. Two separate Surface Laptop Ultra units failed identically under different memory split configurations, indicating a systemic issue related to power delivery or firmware protection, not a defective individual unit. We have reported this behaviour and will update upon vendor resolution.

Storage Performance

The Surface Gen5 SK hynix SSD achieved sequential read 14.1GB/s and write11.5GB/s. Raw 4K random read performance hits1.67M IOPS bypassing OS layers, but Windows on Arm + NTFS + BitLocker caps filesystem 4K random I/O around 300K IOPS, less than half x64 laptops. The bottleneck is Windows on Arm storage stack, not the SSD hardware, with BitLocker software encryption consuming CPU resources. The Precision5 16S, Pro Max14 and Blade16 reached 648K–780K filesystem 4K IOPS. The Pro7 14 QLC SSD suffered extremely poor write speeds around120MB/s. Storage latency testing showed Surface excels at sequential tail latency, while Precision5 16S achieved lowest average 4K random latency. The Windows on Arm stack introduced severe long-tail I/O latency spikes for random workloads.

Conclusion

The RTX Spark N1X inside Surface Laptop Ultra for Business delivers class-leading notebook multi-threaded CPU performance paired with high memory bandwidth and up-to-96GB GPU memory allocation. It dominates Cinebench, 7-Zip, Blender, x265 encoding and LLVM compilation, sets sequential SSD speed records and uniquely runs 35B local coding agents within 90 seconds, while x64 integrated-GPU laptops require many minutes.

However, the platform carries notable drawbacks. Windows on Arm limits random storage IOPS heavily, memory latency is elevated, single-thread performance falls behind x64 competitors, and several Arm64 software ports are poorly optimized. The shared-memory iGPU halves browser responsiveness while running AI agents. For models fitting inside 24GB VRAM, Razer Blade16 RTX5090 offers 2.3× faster token generation with smoother multitasking. The Surface shines specifically for large local 20B–35B AI models, big code compilation and CPU rendering.

Most users will never run heavy HPL double-precision linear algebra workloads; the laptop sustains 98°C in every other benchmark without crashing. Still, a workstation-grade laptop should survive its heaviest valid compute tasks, and the unexplained hard power cut remains an unresolved flaw.

For developers and analysts running 20B–35B local models, compiling large codebases or doing CPU rendering on a laptop, this Surface N1X delivers capabilities no x64 notebook in our comparison can match, especially in the 128GB configuration. Users working on single-threaded, browser-heavy or small discrete-GPU AI workloads will get better value from x64 alternatives, either Precision5 16S or Blade16. We will publish chassis quality and battery testing later, plus vendor responses regarding the power shutdown defect. The Surface Laptop Ultra for Business is listed on our local-AI laptop leaderboard and business laptop leaderboard with these benchmark results.

Beijing Qianxing Jietong Technology Co., Ltd.
Sandy Yang/Global Strategy Director
WhatsApp / WeChat: +86 13426366826
Email: yangyd@qianxingdata.com
Website: www.qianxingdata.com/www.storagesserver.com
Business Focus:
ICT Product Distribution/System Integration & Services/Infrastructure Solutions
With 20+ years of IT distribution experience, we partner with leading global brands to deliver reliable products and professional services.
“Using Technology to Build an Intelligent World”Your Trusted ICT Product Service Provider!