NVIDIA Brings Its Superchip Architecture to the Windows PC. The Same Silicon That Powers a $4,699 AI Workstation Is Coming to Slim Laptops This Fall.
At its Computex 2026 keynote in Taipei on June 1, NVIDIA CEO Jensen Huang announced the RTX Spark — the company’s first processor designed for Windows laptops and compact desktops. The chip pairs a 6,144-core Blackwell RTX GPU with a 20-core Arm-based Grace CPU and up to 128 GB of unified LPDDR5X memory on a single TSMC 3nm package, delivering what NVIDIA calls 1 petaFLOP of FP4 AI compute.
The architecture is not new — it is a close consumer-market cousin of the GB10 Grace Blackwell chip that already powers NVIDIA’s DGX Spark, a compact AI workstation that sells for $4,699. The key differences are the operating system (Windows instead of DGX Ubuntu), the form factor (14–16-inch slim laptops and mini desktops versus a standalone unit), and the target buyer: creators and knowledge workers, not just AI researchers.
ASUS, Dell, HP, Lenovo, Microsoft, and MSI have all committed to RTX Spark devices for fall 2026, with Acer and GIGABYTE to follow. No pricing has been announced. The launch put direct competitive pressure on Apple’s M5 silicon, Qualcomm’s Snapdragon X Elite, and x86 incumbents Intel and AMD — NVIDIA’s stock rose roughly 5% on announcement day while Qualcomm fell approximately 7%.
- 1PetaFLOPFP4 AI compute (with sparsity) — equivalent to ~500 TFLOPS dense FP4 — NVIDIA Newsroom · June 1, 2026
- 128GBunified LPDDR5X memory, CPU and GPU sharing the same pool at 300 GB/s bandwidth — NVIDIA Newsroom · Tom's Hardware
- 70Btransistorson TSMC 3nm — the same process node as Apple M4 (Apple M5 uses TSMC N3P, a refined 3nm variant) — Tom's Hardware · TechSpot
- 120Bparametersmaximum local LLM size RTX Spark can run, with up to 1 million token context — per NVIDIA — NVIDIA Newsroom
- 30+laptopsand ~10 compact desktops committed by OEM partners for fall 2026 launch wave — NVIDIA Newsroom · Tom's Hardware
The N1X chip — long rumored — is real. It is not a discrete GPU. It is NVIDIA’s first full system-on-chip for the Windows PC.
For forty years, NVIDIA made discrete graphics cards. The RTX Spark is something different: a unified system-on-chip (SoC) in which the GPU, CPU, memory, and interconnect are all integrated on one package. The underlying silicon has been known internally as the N1X. At Computex 2026, NVIDIA officially named the platform and its CPU-GPU pairing the “RTX Spark Superchip.”
The Blackwell RTX GPU inside the chip carries 6,144 CUDA cores, fifth-generation Tensor Cores (capable of FP4 precision), and RT Cores for ray tracing — the same generation of GPU architecture found in NVIDIA’s flagship desktop RTX 50-series cards. The Grace CPU is a 20-core Arm processor: 10 high-performance Cortex-X925 cores clocked up to 4.1 GHz and 10 efficiency-focused Cortex-A725 cores. The CPU design was co-developed with MediaTek. CPU and GPU are connected via NVLink-C2C at 600 GB/s — faster than any PCIe lane between a discrete GPU and a CPU in a conventional laptop.
The memory pool — up to 128 GB of LPDDR5X running at 300 GB/s — is shared between the CPU and the GPU. That “unified memory” design is the same architectural principle behind Apple Silicon: no separate VRAM pool, no memory-copy overhead between CPU and GPU. The maximum 128 GB configuration is large enough to hold a 120-billion-parameter language model, which NVIDIA has used as a marketing benchmark for the platform. That claim comes directly from NVIDIA’s own newsroom and has not yet been independently verified under production conditions.
DGX Spark launched as a $4,699 AI workstation running Ubuntu. RTX Spark is its consumer Windows sibling.
To understand RTX Spark’s significance, it helps to understand what came before it. Earlier in 2026, NVIDIA shipped the DGX Spark — a compact, Mac-mini-sized unit that NVIDIA CEO Jensen Huang debuted as “Project DIGITS” at CES 2025 under the GB10 Grace Blackwell label. The DGX Spark runs DGX OS, a customized Ubuntu 24.04 environment, and targets AI researchers, developers, and data scientists who need a local inference machine with a full CUDA software stack. Its starting price is $4,699.
RTX Spark uses essentially the same N1X processor silicon. The meaningful differences are the platform: RTX Spark devices ship with Windows, run inside slim laptop chassis or consumer-grade mini-desktops, are designed for all-day battery life, include G-SYNC OLED displays, and target a much broader audience. Not all RTX Spark SKUs will have all cores enabled — device manufacturers will offer configurations across a range, and lower-tier SKUs may ship with fewer CPU or GPU cores active, per The Register.
- Underlying chipRTX: N1X Grace Blackwell (same silicon) · DGX: GB10 Grace Blackwell (same silicon)
- OSRTX: Windows 11 (full CUDA stack) · DGX: DGX OS (Ubuntu 24.04, customized)
- Form factorRTX: 14–16-inch laptops + mini desktops · DGX: Standalone compact desktop unit
- Target buyerRTX: Creators, knowledge workers, gamers · DGX: AI researchers, developers
- PriceRTX: Undisclosed — fall 2026 · DGX: $4,699 (currently available)
- GamingRTX: Yes — RTX support, 1,000+ titles · DGX: No
“The PC is being reinvented. For forty years, you launched apps. Click. Type. With RTX Spark and Microsoft Windows, you ask — and the PC does the work.”
Jensen Huang, NVIDIA CEO — Computex 2026 keynote, Taipei, June 1, 2026
Six OEM partners confirmed for fall 2026. Microsoft’s Surface is among them. Acer and GIGABYTE to follow.
NVIDIA confirmed six OEM partners for the initial fall 2026 wave: ASUS, Dell, HP, Lenovo, Microsoft, and MSI. Acer and GIGABYTE are expected in a subsequent wave. The presence of Microsoft’s Surface line — in the form of the Surface Laptop Ultra — is notable: Microsoft has historically aligned Surface hardware closely with its platform strategy, and committing the Surface brand to a non-Intel, non-AMD, non-Qualcomm processor is a meaningful endorsement.
Confirmed device names from partner announcements include: the ASUS ProArt P16 (H7607) and ProArt P14 (H7407) laptops plus a ProArt Mini PC; the Dell XPS 16 Creator Edition; the HP OmniBook Ultra 16 and OmniBook X 14; the Lenovo Yoga Pro 9n; the Microsoft Surface Laptop Ultra; and the MSI Prestige N16 Flip AI+. None of these devices have disclosed pricing as of the June 1, 2026 announcement date.
- ASUSProArt P16 (H7607) · ProArt P14 (H7407) · ProArt Mini PC — ASUS Pressroom
- DellXPS 16 Creator Edition — NVIDIA Newsroom · Windows Blog
- HPOmniBook Ultra 16 · OmniBook X 14 — Windows Blog · NVIDIA product page
- LenovoYoga Pro 9n — Windows Blog · NVIDIA product page
- MicrosoftSurface Laptop Ultra — NVIDIA Newsroom · Windows Blog
- MSIPrestige N16 Flip AI+ — NVIDIA Newsroom · NVIDIA product page
- Acer / GIGABYTEModels unannounced — to follow initial wave — Tom's Hardware · CNBC
It beats the base M5 in a CPU benchmark but lags the M5 Max in memory bandwidth. The real test is inference throughput — and that data does not exist yet.
The most relevant competitive comparison is Apple’s M5 chip family — the current gold standard for unified-memory AI inference in a slim laptop. Pre-release Clang compile benchmarks (CPU-only, single and multi-threaded) show the RTX Spark scoring 43,149, ahead of Apple’s base M5 at 27,996 — a 54% CPU performance advantage. However, the RTX Spark is marginally slower than the 15-core M5 Pro and 21% slower than the 18-core M5 Pro in the same test. These are CPU benchmarks only and do not reflect GPU-accelerated AI inference performance.
The memory bandwidth picture is more nuanced. RTX Spark’s 300 GB/s LPDDR5X bandwidth (CPU-to-memory path) is roughly half the Apple M5 Max’s 546 GB/s. For large-model inference, memory bandwidth — not raw FLOP count — tends to be the bottleneck: the faster memory can feed the GPU, the higher the token-per-second throughput. NVIDIA’s NVLink-C2C interconnect runs at 600 GB/s between CPU and GPU within the package, which is a different path than the main memory bus and reflects internal data movement rather than the CPU-DRAM bandwidth that governs model loading.
AppleInsider published analysis citing a year-old N1X Geekbench 6 result matching M3 Max performance (released late 2023), characterizing the RTX Spark as roughly two generations behind Apple Silicon. The authors acknowledged that Geekbench CPU scores have limited relevance for AI workloads, and the benchmark pre-dates the announced 4.1 GHz Cortex-X925 clock speeds. No independent production benchmarks on shipping hardware exist as of June 1, 2026; devices are not yet in reviewers’ hands.
On the market reaction front: NVIDIA’s stock rose approximately 5% on announcement day while Qualcomm fell roughly 7%, according to 247 Wall St. That asymmetry reflects a specific fear in the Qualcomm camp: NVIDIA’s entry into Windows on Arm directly competes with Qualcomm’s Snapdragon X Elite platform, which has been the primary Windows Arm chip since Qualcomm’s exclusive deal with Microsoft. That exclusivity period recently expired, clearing the path for both NVIDIA and others to enter.
Microsoft re-architected Windows for the chip. Adobe rearchitected Photoshop. NVIDIA is betting the hardware alone is not the product.
Hardware announcements often run ahead of the software that justifies them. NVIDIA and Microsoft have made a visible effort to avoid that trap at launch. Microsoft announced a series of Windows-level changes specifically tuned for the RTX Spark’s 20-core heterogeneous CPU: workload profile scheduling to separate AI agent tasks from foreground work, integration with the Microsoft Power and Thermal Framework, DirectX 12 enhancements with neural rendering support, and native TensorRT GPU acceleration via Windows ML.
Microsoft has also invested in x86 emulation: Enhanced Prism, the Windows Arm compatibility layer, has been updated to reduce performance penalties for legacy applications. The installed base of x86 Windows software is the single largest practical obstacle to any Windows Arm platform; how well RTX Spark laptops run existing productivity software will matter more to mainstream buyers than the FP4 petaFLOP number.
Adobe committed to rearchitecting Photoshop and Premiere Pro for the platform, with NVIDIA claiming 2× performance improvements. Shantanu Narayen, Adobe CEO, appeared at the announcement alongside Jensen Huang and Microsoft CEO Satya Nadella. NVIDIA also states that over 100 Windows software providers are adopting the platform. That claim has not been independently verified and reflects NVIDIA’s own characterization of partner commitments.
“Our goal is to deliver unmetered intelligence to every home and every desk with Windows.”
Satya Nadella, Microsoft CEO — joint announcement with NVIDIA, June 1, 2026
“1 petaFLOP” is a real number — but it measures one narrow workload. It is not a verdict on whether RTX Spark out-muscles a high-end gaming desktop. It does not.
The headline figure — 1 petaFLOP — is precise, but it carries two asterisks that matter. The first is precision. That petaFLOP is an FP4 number: four-bit floating point, NVIDIA’s NVFP4 format, a data type built specifically for running AI models, not for general computing or gaming. Games and most desktop software run at FP32 (32-bit) or higher; scientific and engineering work often needs FP64. FP4 stores each number in one-eighth the bits of FP32, trading numerical precision for raw throughput — a technique called quantization. It is excellent for serving a large language model and useless for rendering a frame in Battlefield 6.
The second asterisk is sparsity. NVIDIA’s Blackwell Tensor Cores can exploit 2:4 structured sparsity — when at least two of every four weights in a model are zero, the hardware skips them and roughly doubles its effective throughput. The full “1 petaFLOP” assumes that best case. Strip the sparsity assumption out and the dense FP4 figure is about half — the ~500 TFLOPS this page cites in its spec table. Neither number is fabricated; both describe the same chip under different, clearly labeled assumptions. The marketing simply leads with the larger one.
This is why a cross-precision FLOP comparison is apples-to-oranges. On the same silicon, throughput roughly doubles each time you halve the precision — FP32 to FP16 to FP8 to FP4 — so an FP4 figure is, order-of-magnitude, several times larger than the same chip’s FP32 rate. Quoting the FP4-with-sparsity petaFLOP and implying it describes general or gaming horsepower is a category error. For a sense of scale, a flagship discrete gaming GPU tells the other half of the story: NVIDIA’s own GeForce RTX 5090 carries 32 GB of GDDR7 on a 512-bit bus delivering roughly 1.79 TB/s of memory bandwidth — about six times the RTX Spark’s 300 GB/s LPDDR5X — paired with a high-end Core i9 desktop CPU. For high-frame-rate gaming and full-precision FP32/FP64 work, that conventional tower is dramatically more capable.
- Headline number1 petaFLOP is FP4 (4-bit) with sparsity assumed — an AI-inference precision, not the FP32/FP64 precision games and general computing use. Dense FP4 is ~half (~500 TFLOPS).
- Why it inflatesHalving precision roughly doubles throughput (FP32→FP16→FP8→FP4), and 2:4 sparsity roughly doubles it again. The FP4-sparse figure is several times the same chip's FP32 rate — not comparable to a gaming/compute number.
- Memory bandwidthRTX Spark: 300 GB/s LPDDR5X. NVIDIA RTX 5090 desktop GPU: ~1.79 TB/s GDDR7 — roughly 6× more. Bandwidth, not FLOP count, governs both gaming frame rates and inference token speed.
- What each is forA high-end Windows desktop (RTX 5090 + Core i9) wins decisively on FP32 gaming and full-precision compute. RTX Spark wins at one thing a 24–32 GB gaming GPU physically cannot do: hold a large quantized model in 128 GB of unified memory and run it locally.
None of this makes the petaFLOP a lie. FP4 plus 128 GB of unified memory is genuinely useful for exactly one real thing: running large quantized AI models locally that a 24–32 GB gaming card simply cannot fit in its VRAM. That is the chip’s actual purpose, and on that task it is a different and capable tool. The honest framing is not “more powerful” or “less powerful” — it is a different tool for a different job. The petaFLOP is a benchmark chosen to flatter one narrow workload, not a measure of general performance, and readers should treat it as such.
No price. No exact ship date. No production benchmarks. The three questions that will determine whether RTX Spark is a premium niche or a platform shift.
The launch announcement left several commercially critical questions unanswered. Pricing is the most obvious gap. NVIDIA did not disclose any price or price band for RTX Spark devices. The closest reference point — the DGX Spark workstation at $4,699 — implies premium positioning, but a laptop at DGX Spark prices would occupy a narrow segment. The Register noted that “top end variants will not be inexpensive” without providing a number. No OEM partner released pricing on announcement day.
The timeline is “fall 2026.” That is not a ship date. It is a season. The gap between Computex announcement and retail availability for new laptop platforms has historically run three to six months, with premium SKUs shipping before volume configurations. Given that Computex was June 1, “fall 2026” means September at the earliest and could stretch toward the end of November. Holiday availability appears to be the target.
Production benchmarks on retail hardware do not exist. The pre-release Clang score is a single CPU benchmark on a sample that may differ from shipping configurations. The 1-petaFLOP FP4 AI figure, the 120B-parameter claim, and the 1440p gaming performance claims all come from NVIDIA marketing materials and have not been independently replicated. The memory bandwidth disadvantage versus Apple M5 Max (300 GB/s vs. 546 GB/s) will be quantifiable once devices ship — and it will matter most to users trying to push the largest models at the fastest token rates. Reviewers at Tom’s Hardware, The Verge, and others will run those tests when hardware arrives.
- Confirmed6,144-core Blackwell RTX GPU; 20-core Grace CPU (co-designed with MediaTek, Cortex-X925/A725)
- ConfirmedUp to 128 GB LPDDR5X unified memory; 300 GB/s bandwidth; NVLink-C2C at 600 GB/s
- ConfirmedTSMC 3nm process; 70 billion transistors
- Confirmed1 petaFLOP FP4 with sparsity (~500 TFLOPS dense) — NVIDIA claim, not independently replicated
- ConfirmedOEM partners: ASUS, Dell, HP, Lenovo, Microsoft, MSI (fall 2026); Acer, GIGABYTE to follow
- ConfirmedSpecific device names: ASUS ProArt P16/P14/Mini PC, Dell XPS 16 Creator, HP OmniBook Ultra 16/X 14, Lenovo Yoga Pro 9n, Microsoft Surface Laptop Ultra, MSI Prestige N16 Flip AI+
- Not disclosedPricing — no price or price band from NVIDIA or any OEM partner as of June 1
- Not disclosedSpecific ship date — 'fall 2026' only; no calendar date confirmed
- Not yet availableIndependent production benchmarks — CPU, GPU, AI inference, gaming, power consumption
- Marketing claim only120B-parameter local LLM capacity; 1 million token context; 1440p 100+ FPS gaming; 2× Adobe performance
Qualcomm dropped 7% on announcement day. The market read the NVIDIA entry as a platform-level threat to the Windows Arm incumbent.
The immediate market reaction was asymmetric: NVIDIA stock rose approximately 5%, while Qualcomm fell approximately 7%, according to 247 Wall St. That spread reflects the competitive reality. Qualcomm has been the dominant Windows Arm chip vendor since its exclusivity arrangement with Microsoft — and that exclusivity has now expired. NVIDIA’s entry with a chip that shares the GB10’s AI credentials, adds full CUDA gaming support, and comes with Microsoft as a partner is a more credible threat than anything Intel or AMD has put into the Windows Arm ecosystem.
Intel and AMD are the other incumbents under pressure. Both companies have dominated x86 laptop chips for decades. RTX Spark is not x86 — it is Windows on Arm — which means x86 application compatibility depends on Microsoft’s Enhanced Prism emulation layer. Whether that layer is performant enough to handle enterprise productivity software, legacy games, and specialized tooling is a question reviewers will answer when devices ship.
Introducing NVIDIA RTX Spark — the world's first Windows PC purpose-built for personal AI agents. 1 petaFLOP of AI, 128GB unified memory, full CUDA stack. Coming this fall from ASUS, Dell, HP, Lenovo, Microsoft, MSI and more. #Computex2026
NVIDIA's RTX Spark is here: Grace Blackwell SoC, 20-core Arm CPU (co-built with MediaTek), 6,144 Blackwell CUDA cores, up to 128GB LPDDR5X, 1 petaFLOP FP4. The chip formerly known as N1X is official. Qualcomm's Windows Arm exclusivity ended — and NVIDIA just walked through the door.
