Surface Laptop Ultra Review: RTX Spark N1X Delivers 128GB AI, but OpenBLAS Power-Offs — Megathread

According to a detailed benchmark review published by StorageReview, the new Microsoft Surface Laptop Ultra for Business—the first Surface notebook built on NVIDIA silicon and powered by the NVIDIA RTX Spark N1X superchip—rewrites the rules for multi-core CPU performance and local large language model (LLM) capacity in a Windows laptop. Armed with a 20-core Grace Arm CPU, a 6,144-core Blackwell RTX GPU, and 128GB of unified LPDDR5X memory on the top-end configuration, the machine crushed nearly every multi-threaded CPU test against four competing Intel and AMD x64 notebooks while running a 35-billion-parameter coding agent locally at up to 89 tokens per second.

Yet StorageReview's lab testing also uncovered sharp platform trade-offs on Windows 11 Pro 26H1 for Arm64. Memory latency is more than double the fastest x64 competitor, Windows on Arm's storage stack throttles random 4K SSD throughput to less than a fifth of the raw drive's capability, and—most concerning of all—two separate Surface Laptop Ultra review units suffered sudden, hard power-offs during a heavy all-core double-precision floating-point workload.

Inside the RTX Spark N1X Platform and Surface Specifications​

When Microsoft opened pre-orders on October 7, 2026, for the Surface Laptop Ultra alongside the desktop Surface RTX Spark Dev Box, the company pitched the lineup as a ground-up rethinking of the high-performance Windows PC for developers, creators, and AI builders. While Microsoft's marketing materials tout theoretical FP4 AI performance up to 1 petaflop (using sparsity) and claim the ability to run models exceeding 120 billion parameters locally, StorageReview's benchmark analysis puts hard numbers to how the silicon actually behaves in a shipping configuration. (Physical chassis impressions, keyboard feel, display testing, and battery life were deferred by StorageReview for a subsequent follow-up once lab benchmarking concludes.)

At the heart of the system is NVIDIA's RTX Spark N1X, the laptop and compact desktop sibling of the GB10 superchip found in the DGX Spark. NVIDIA produces two mobile N1X bins, both operating within a combined 45W to 80W thermal design power (TDP) envelope for the CPU and GPU together (compared to 140W for desktop N1X and GB10 implementations):

• Top-tier N1X bin (as tested by StorageReview): A 20-core NVIDIA Grace Arm CPU (reported in Windows as 10 faster cores and 10 efficiency cores, with no simultaneous multithreading) paired with a 6,144-core Blackwell RTX GPU and up to 128GB of LPDDR5X unified memory.
• Entry N1X bin: An 18-core Grace CPU paired with a 5,120-core Blackwell RTX GPU, supporting up to 64GB of memory.

What makes the N1X fundamentally different from traditional x64 notebooks is its memory subsystem. A single 128GB pool of LPDDR5-9400 serves both the Grace CPU and Blackwell GPU across an NVLink-C2C interconnect. Across a 256-bit interface, 9,400 MT/s yields roughly 300GB/s of theoretical bandwidth. Windows 11 presents the GPU similarly to a discrete card with a fixed dedicated allocation alongside a shared pool, and users can adjust that dedicated split via the NVIDIA app—up to 96GB for the GPU on a 128GB machine.

Section Summary​

The Surface Laptop Ultra for Business introduces NVIDIA's 45–80W RTX Spark N1X superchip to Windows 11 Pro 26H1 (Arm64). Its defining architectural advantage is up to 128GB of LPDDR5-9400 unified memory connected over NVLink-C2C, allowing users to assign 64GB—or up to 96GB—directly to the 6,144-core Blackwell GPU.

Local AI Benchmarks: Capacity vs. Raw Discrete Speed​

How does an integrated superchip with massive memory stack up against traditional x64 laptops? StorageReview ran its automated sr-client 1.2 benchmark suite against four current x64 notebooks, all plugged in and set to their highest OEM performance profiles:

1. Dell Pro Precision 5 16S: Intel Core Ultra X9 388H (16C/16T), 64GB LPDDR5-8533, integrated Intel Arc Pro B390 GPU.
2. Dell Pro 7 14: AMD Ryzen AI 9 HX PRO 470 (12C/24T), 64GB LPDDR5-8000, integrated AMD Radeon 890M (with 32GB Variable Graphics Memory reserved).
3. Razer Blade 16: AMD Ryzen AI 9 HX 370 (12C/24T), 32GB LPDDR5-8000, discrete NVIDIA GeForce RTX 5090 Laptop GPU (24GB VRAM).
4. Dell Pro Max 14 Premium: Intel Core Ultra 9 285H (16C/16T), 64GB LPDDR5-8400, discrete NVIDIA RTX PRO 2000 Blackwell Laptop GPU (8GB VRAM; 7.7GB usable).

Every local AI test reveals a clear dichotomy: if a model fits comfortably inside 24GB of dedicated GDDR memory, the Razer Blade 16's discrete RTX 5090 Laptop GPU is roughly 2.3 times faster than the Surface Laptop Ultra. However, as soon as the workload scales up to "Tier M" models requiring around 40GB or more of GPU-addressable memory, the Blade 16 and Dell Pro Max 14 Premium are disqualified outright, leaving the Surface Laptop Ultra in a class of its own.

MLPerf Client v2.0 and RAG Throughput​

Using native Arm64 builds of llama.cpp with CUDA on the Surface (while Intel used OpenVINO and AMD used Windows ML with DirectML), StorageReview tested Llama 3.1 8B and Phi-4-mini in MLPerf Client v2.0:

• Token Generation Rates: The Surface generated 56.4 tokens/s on Llama 3.1 8B and 97.5 tokens/s on Phi-4-mini. The RTX 5090-equipped Blade 16 hit 128.8 tokens/s and 222.7 tokens/s, respectively. Interestingly, the Dell Pro Max 14's 8GB RTX PRO 2000 came within 5% of the Surface on Llama 3.1 8B and within 12% on Phi-4-mini, showing that on small models, the N1X GPU performs much like an entry-level discrete workstation GPU.
• Agentic MLPerf Scenarios: In multi-turn Data Analyst and Software Engineering (SWE) agent tests using Llama 3.1 8B, the Blade 16 finished in 27.8s and 24.6s, roughly twice as fast as the Surface's 52.6s (48.9 tok/s) and 46.0s (46.4 tok/s). The Dell Pro Precision 5 16S trailed at 84.7s and 65.6s, while the AMD-based Dell Pro 7 14 failed to finish (DNF) because DirectML ran out of memory despite having 32GB reserved.
• Retrieval-Augmented Generation (RAG): Indexing 20,963 SQuAD v1.1 passages with nomic-embed-text and answering 2,000 queries with Phi-4-mini took the Surface 189.8 seconds for ingestion and 0.235 seconds (median question-to-answer), trailing both the Blade 16 (131.0s ingest / 0.159s answer) and the 8GB RTX PRO 2000 in the Dell Pro Max 14 (153.9s ingest / 0.341s answer), though vastly outpacing the integrated x64 systems.

Model Swapping and the 35B-Parameter Coding Agent​

Where the Surface Laptop Ultra flexes its 64GB GPU allocation is in Ollama model loading and the Codex CLI coding agent benchmark (which requires an LLM to pass 11 unit tests in a pinned software repository):

Because the Surface can keep both the 12.8GB gpt-oss:20b model and the 21.1GB Qwen3.6-35B-A3B model resident in its 64GB GPU allocation simultaneously, switching between them and reloading Qwen3.6-35B-A3B took just 5.01 seconds. Moreover, because unified memory avoids copying weights over a PCIe bus into discrete VRAM, the weights are ready as soon as they are read.

In the Tier M coding agent tasks, the Surface generated between 58 and 89 tokens per second, finishing the 35B-parameter Qwen3.6-35B-A3B 64K-context coding challenge in a median 88.6 seconds (passing 3 of 3 runs). The integrated x64 notebooks took between nearly 6 minutes (Dell Precision 5 16S) and over 11 minutes (Dell Pro 7 14).

However, buyers should note two critical caveats from StorageReview's testing:

1. Lower-RAM configurations won't replicate this: A 24GB, 32GB, or 48GB Surface Laptop Ultra cannot allocate 64GB to the GPU, meaning it will not qualify for these large Tier M multi-model workloads.
2. The Unified Memory Bus Tax: When you run a heavy coding agent in the background while working in a browser, the integrated GPU and CPU fight over the same memory bus. Running the coding agent alongside the Speedometer 3.1 browser benchmark cut the Surface's browser responsiveness nearly in half (from 34.7 down to 18.1, a 1.92x slowdown). On the Razer Blade 16, where the agent runs cleanly on the isolated RTX 5090, browser performance dipped by only 6%.

Section Summary​

For models that fit in 24GB of VRAM, a discrete mobile RTX 5090 is more than twice as fast and avoids bogging down the system's CPU/memory bus during multitasking. But for developers running 20B to 35B+ parameter models with massive context windows (128K) or keeping multiple large models hot in memory simultaneously, the 128GB Surface Laptop Ultra finishes agentic coding loops in under 90 seconds that discrete-GPU laptops cannot load and x64 integrated laptops take up to 11 minutes to complete.

CPU, Memory Latency, and the Windows on Arm Software Factor​

Using native Arm64 binaries compiled specifically for its benchmark suite, StorageReview found that the 20-core NVIDIA Grace CPU in the N1X is a multi-threaded juggernaut—frequently beating the 16-core Intel Core Ultra X9 388H and 12-core/24-thread AMD Ryzen AI 9 chips by 30% to 47%.

• Cinebench 2026: The Surface scored 6,963 in multi-core, beating the second-place Razer Blade 16 (5,179) by 34% and the Dell Pro Precision 5 16S (4,739) by 47%. It also narrowly took first place in single-core at 558.5 to the Precision's 534.8. In a brutal 30-minute sustained Cinebench loop, the Surface showed zero thermal throttling, finishing at 7,016 (1.1% faster than its first pass) while holding 98°C and ~3.28GHz.
• LLVM 21.1.4 Compilation: Compiling the massive LLVM tree with Clang and Ninja took the Surface just 206.4 seconds, blowing past the Razer Blade 16 (327.0s) by 37% and the Dell Precision 5 16S (340.1s, retested after disabling Windows Smart App Control, which initially blocked the tool).
• Blender 4.5 & 7-Zip: The Surface swept all three Blender CPU scenes (leading the Blade 16 by 47% in Monster at 207.6 samples/min) and topped multi-threaded 7-Zip at 151.0 GIPS total (and 156.9 GIPS in compression, 35% ahead of second place).

Why Single-Threaded and Specialized Workloads Lag Behind​

Why, then, did the Surface finish dead last in Python (PyPerformance), web browsing (Speedometer 3.1 on native Arm64 Edge, scoring 34.7 vs. 42.7 on the Precision 5 16S), SVT-AV1 video encoding, and GROMACS molecular dynamics? StorageReview's diagnostics point to three distinct culprits:

1. High Memory and Core-to-Core Latency: While the Surface's LPDDR5-9400 crushed the STREAM Triad bandwidth test at 122,114 MB/s (48% ahead of the Precision 5 16S and ~2.6x the AMD laptops), its latency was the worst in the group by a wide margin. Memory latency measured 251.1 ns (versus 115.5 ns on the Intel-based Precision 5 16S), and mean core-to-core latency hit 171.6 ns with a 362.6 ns max (compared to 49.7 ns mean on the AMD Pro 7 14). Latency-sensitive, single-threaded pointer-chasing workloads feel that penalty immediately.
2. A Peculiar Single-Core Clock Drop: StorageReview discovered that when a single N1X core is pinned under continuous load for several minutes—as happens during single-threaded 7-Zip—it drops into a slower state roughly 20% below its initial pace, even though core temperatures are only 47°C to 67°C. Because PyPerformance and single-thread Stockfish immediately followed 7-Zip on the same pinned core in the automated suite, they inherited that downclocked firmware state.
3. Unoptimized Arm64 Code Paths: In HandBrake, the Surface won x265 software encoding by 25% (41.7 fps), yet collapsed to 15.4 fps in SVT-AV1—roughly one-fifth the speed of the Precision 5 16S (80.8 fps). Similarly, it trailed in GROMACS (17.29 ns/day vs. 26.41 ns/day on the Blade 16) because the Arm64 build lacks the hand-tuned AVX2 kernels that accelerate x64 processors. Interestingly, in the lightweight 1080p SuperTuxKart gaming test, running the x64 build under Windows 11's Prism emulation actually yielded a higher average frame rate (264.6 fps) than the native build (235.9 fps).

Section Summary​

The N1X's 20 Grace cores and 122GB/s of memory bandwidth dominate heavy parallel workloads like C++ compilation, 3D rendering, and file compression. However, 251ns memory latency, a strange single-core sustained clock step-down, and unoptimized Arm64 SIMD paths in certain open-source tools (SVT-AV1, GROMACS) leave it 15% to 30% behind Intel's Core Ultra X9 388H in everyday single-threaded and browser responsiveness.

Storage Bottlenecks: PCIe Gen5 Hardware vs. the Windows on Arm Path​

The tested Surface Laptop Ultra ships with a removable 1TB SK hynix (HFS001TFM9X189N) PCIe Gen5 x4 NVMe SSD. In sequential throughput (fio 128K at QD32 across a 25GB footprint), it set an outright record for the group: 14,134 MB/s read and 11,461 MB/s write.

Yet when testing 4K random I/O through the standard NTFS file system with BitLocker enabled as shipped, performance hit an unexpected wall:

• File-System 4K Random I/O (QD32): The Surface peaked at 307,738 IOPS read and 284,680 IOPS write—less than half the 648K to 780K IOPS achieved by the x64 Dell and Razer laptops.
• Raw Device 4K Random Read (Bypassing NTFS/BitLocker): When StorageReview bypassed the Windows file system and tested the raw physical disk directly, the SK hynix drive screamed to 1,670,736 IOPS—a 5.6x difference.
• Tail Latency Spikes: In 16-job random 4K sets through the OS path, the Surface exhibited extreme p99.99 tail latencies of 4.8 seconds on reads and 5.6 seconds on writes (meaning roughly 1 in 10,000 I/O requests stalled for seconds).

StorageReview notes that the 300K IOPS ceiling is imposed by the Windows on Arm storage path rather than the SSD, pointing to software-based BitLocker XTS-AES-128 encryption combined with Memory Integrity (HVCI)—which consumed 57% of the CPU during the 4-job random test—as the likeliest culprit.

Section Summary​

While the Surface's Gen5 SSD delivers blistering 14.1 GB/s sequential reads for loading AI models, the Windows 11 Arm64 storage and software encryption stack currently bottlenecks small 4K random file-system operations to roughly one-fifth of the hardware's raw capability.

The OpenBLAS Power-Off Mystery: What Happened in the Lab?​

The most urgent finding in StorageReview's evaluation involves system stability under extreme floating-point stress. During the OpenBLAS benchmark—which executes an HPL-style double-precision (FP64) LU matrix factorization and solve—the Surface Laptop Ultra abruptly lost power three times in one evening whenever the all-core test transitioned to the N=20000 problem size across all 20 threads.

Even more telling: this was StorageReview's second retail/review unit to fail under this exact workload. Their first review unit hung in the same OpenBLAS test, developed an OLED display defect near the exhaust edge alongside a reproducible CPU fault, and had to be replaced. Furthermore, the two units used different unified memory splits (96GB/32GB on the first unit; 64GB/64GB on the second).

Here is what the diagnostic telemetry revealed during the October 14 failures on the replacement unit:

• It does not look like a standard thermal shutdown: Although the N1X runs hot—regularly sitting at 97°C to 98°C during Blender, Stockfish, and the 31-minute Cinebench loop without losing a single point of performance—the CPU was at 95.2°C to 95.7°C ( around 3.1–3.2GHz) when power cut out during OpenBLAS.
• Instantaneous power loss: Windows Event Viewer logged Kernel-Power Event ID 41 with bugcheck code 0 and no power-button press. No crash dump, WHEA hardware error record, or thermal trip event was written.
• Repeatable timing, yet intermittent: In two consecutive reproduction runs, the machine shut off 42 to 46 seconds after the workload started, and exactly 14 to 16 seconds into the 20-thread N=20000 solve. All preceding N=10000 residual checks passed without mathematical corruption. However, on one retry immediately following the first reboot, the unit successfully completed the entire suite up to N=40000 (scoring 336 and 370 GFLOPS), failing on three out of four attempts overall while plugged into AC power with a 100% battery.
• Platform counter clues: The PSYS platform thermal zone counter climbed from ~1,000–1,200 at idle to peak at 4,419 and 4,406 at the exact moment of the two reproduction power cuts (though the first replaced unit failed at a lower reading of 2,808).

StorageReview noted that other OEM laptops running the RTX Spark N1X in their lab also run in the mid-to-high 90s Celsius under load without faulting, leading the reviewers to suspect that the Surface Laptop Ultra's power delivery, cooling design, or firmware protection limits are tripping under sustained all-core FP64 transient loads. They have reported the findings to the vendors and plan to update their coverage once Microsoft or NVIDIA responds.

How Technical Users Can Test for the OpenBLAS Issue​

For engineers or enterprise evaluators who receive a 20-core Surface Laptop Ultra and want to verify whether their firmware and power delivery handle sustained FP64 load, StorageReview published the exact reproduction procedure. Because the official Windows on Arm OpenBLAS binary caps itself at 12 threads, you must compile version 0.3.34 from source inside an MSYS2 CLANGARM64 shell to load all 20 cores:

pacman -S --needed mingw-w64-clang-aarch64-clang mingw-w64-clang-aarch64-cmake mingw-w64-clang-aarch64-ninja
cmake -S. -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DTARGET=NEOVERSEN1 -DNUM_THREADS=256 -DNOFORTRAN=1 -DC_LAPACK=1 -DBUILD_BENCHMARKS=ON -DBUILD_TESTING=OFF
cmake --build build --target benchmark_linpack_DOUBLE

With the laptop plugged into AC power and Windows 11 set to the Best performance power mode, run the warming loops followed by the 20-thread N=20000 solve:

cd build
OPENBLAS_LOOPS=6 OPENBLAS_NUM_THREADS=20./benchmark_linpack_DOUBLE.exe 10000 10000 1
OPENBLAS_LOOPS=6 OPENBLAS_NUM_THREADS=10./benchmark_linpack_DOUBLE.exe 10000 10000 1
OPENBLAS_NUM_THREADS=20./benchmark_linpack_DOUBLE.exe 20000 20000 1

The Verdict for Enterprise IT and Developers​

Should you deploy the Microsoft Surface Laptop Ultra for Business? The answer depends entirely on whether your workflow hits the VRAM wall of traditional x64 mobile workstations.

• Who should buy it: AI engineers, software developers prototyping local multi-agent workflows on 20B to 35B+ parameter models, C++/Rust developers compiling massive codebases like LLVM, and CPU-bound 3D renderers. Provided you invest in the 128GB unified memory configuration, the RTX Spark N1X lets you run local AI models in under 90 seconds that simply cannot fit into a 24GB RTX 5090 Laptop GPU.
• Who should wait or stick to x64: Users whose day-to-day work revolves around single-threaded responsiveness, heavy browser multitasking alongside local AI, database-style 4K random disk I/O under BitLocker, or smaller AI models (<24GB) where a discrete RTX 5090 is more than twice as fast.

Above all, enterprise IT administrators considering a fleet rollout for scientific or engineering teams should watch closely for an official firmware or EC update from Microsoft and NVIDIA addressing the OpenBLAS FP64 power-off behavior. Putting supercomputer-grade unified memory into a 4.4-pound Windows laptop is an extraordinary milestone—now the platform's power protection firmware just needs to catch up with the silicon.