Log in Sign up
News 128GB Sells Out in Minutes While Benchmarks Say Macs Win: Nvidia's First PC Chip Finds Its Buyers Anyway News 50 Executives, One Saturday Seminar: Sam Altman Visits Samsung's Yongin Campus as Memory Becomes AI's Scarcest Input Productivity A 0.0089-Dollar Lead and 6,400 GitHub Stars: Jev Now Screens People for Your Recruiting Agent, 85% Cheaper Than Clay Business Baseten Seeks $2 Billion at Up to $30 Billion, and the Inference Layer Is Getting Lab-Sized Checks Gemini Four TPUs, One Switched On at a Time: Google's Orbital Data Center Prototype Is Online and Writing Haiku AI Agents AgiBot Tops Seven Robotics Benchmarks With Two New Models, and Wants Worken to Be the Token of Physical AI ChatGPT From December, ChatGPT and Codex Will Share One Usage Pool on Plus, Pro and Business News Anthropic Claims Its AI Also Disproved the Exponential Time Hypothesis, and the Field Is Calling It a 'Giant Bomb' News 128GB Sells Out in Minutes While Benchmarks Say Macs Win: Nvidia's First PC Chip Finds Its Buyers Anyway News 50 Executives, One Saturday Seminar: Sam Altman Visits Samsung's Yongin Campus as Memory Becomes AI's Scarcest Input Productivity A 0.0089-Dollar Lead and 6,400 GitHub Stars: Jev Now Screens People for Your Recruiting Agent, 85% Cheaper Than Clay Business Baseten Seeks $2 Billion at Up to $30 Billion, and the Inference Layer Is Getting Lab-Sized Checks Gemini Four TPUs, One Switched On at a Time: Google's Orbital Data Center Prototype Is Online and Writing Haiku AI Agents AgiBot Tops Seven Robotics Benchmarks With Two New Models, and Wants Worken to Be the Token of Physical AI ChatGPT From December, ChatGPT and Codex Will Share One Usage Pool on Plus, Pro and Business News Anthropic Claims Its AI Also Disproved the Exponential Time Hypothesis, and the Field Is Calling It a 'Giant Bomb'

128GB Sells Out in Minutes While Benchmarks Say Macs Win: Nvidia's First PC Chip Finds Its Buyers Anyway

News · By: AIReport Admin · 1 hour ago · 👁 9
128GB Sells Out in Minutes While Benchmarks Say Macs Win: Nvidia's First PC Chip Finds Its Buyers Anyway

The first laptops built on Nvidia's RTX Spark N1X chip went on sale in October, and the 128GB unified-memory configurations sold out within hours in China despite review scores that trail Apple's silicon. The pitch is not raw speed but native CUDA plus 128GB of memory at $2,599 and up - and Nvidia's own math claims 61 million tokens a month costs $1,200 via API versus about $20 in electricity.

Nvidia's first PC processor is finally in stores, and the market's verdict is split between the checkout page and the benchmark chart. The first machines built on the RTX Spark N1X - a superchip pairing a Grace-class CPU with Blackwell graphics, co-developed with MediaTek - went on sale through October, led by Lenovo's YOGA Pro 15 on October 16, with ASUS, MSI, Dell, HP and Microsoft machines following.

The top configurations vanished first. Lenovo's 128GB-plus-2TB YOGA at 42,999 yuan sold out within 24 hours of launch in China, and ASUS's ProArt创16 at 37,999 yuan for 128GB-plus-1TB cleared out in minutes on JD.com, according to Chinese retail reports. US pricing starts at $2,599. Meanwhile Nvidia's own desktop DGX Spark box has crept from $3,999 to $4,699 official pricing amid memory costs, with resellers asking as much as $6,950.

Raw performance reviews are less flattering. Trusted Reviews measured the ASUS ProArt P14 at 3,003 single-core and 19,171 multi-core in Geekbench 6 - behind the Snapdragon X2 Elite Extreme's 3,711 and 21,940 - while the Chinese tech channel Geekerwan found the chip loses to Apple's M5 Max on CPU and loses to mainstream RTX 5060 laptops in games. As a gaming machine, reviewers broadly agree, it is mediocre.

That is not the product Nvidia is selling. The GPU side holds up - 10,131 in 3DMark Time Spy, double the Adreno graphics in the Snapdragon flagship and ahead of AMD's Strix Halo - and the real payload is the memory: up to 128GB of unified memory at roughly 300GB/s, addressable by a 6,144-CUDA-core Blackwell GPU. That is enough to load 120-billion-parameter models locally, on a machine that weighs under two kilograms.

The economic argument is where Nvidia's own numbers get bold: per its briefing, running 61 million output tokens a month through Claude Sonnet 5.5 Medium costs about $1,200 in API fees, while serving the same volume with an open Qwen model on an RTX Spark laptop costs roughly $20 in electricity. A heavy agent user could amortize the machine in months - if the open models are good enough for the workload.

Sellouts at 43,000 yuan say something about demand for private, unmetered inference that benchmark tables keep missing. The more honest summary sits in one popular Chinese comment: it is the cheapest 128GB Blackwell card you can buy, strapped to a CPU that loses to a MacBook. For security-conscious firms, agent developers and privacy-driven professionals, that trade currently has more buyers than supply.

Comments (0)

Log in to join the discussion

Log In

No comments yet