AMD fired an early shot across Nvidia’s bow this week. On October 5, 2026, the chipmaker published a set of generative-AI benchmarks for its Ryzen AI Max+ Pro 495 processor, code-named Gorgon Halo, just two days before Nvidia’s RTX Spark devices are expected to launch alongside a Microsoft event on October 7. The timing was not subtle. AMD wanted the first word in a fight over who controls the fast-growing “agentic PC” category, according to Tom’s Hardware, which first reported the benchmark release.

The numbers AMD shared are messy in places, and the company stumbled over its own sales pitch during a press briefing, first claiming to have shipped “10s of millions” of AI PCs before walking that back to “over half a million” agentic PCs specifically. Still, the episode says a lot about where PC hardware is heading in late 2026: big unified-memory chips built to run large language models locally, a three-way scramble between AMD, Nvidia, and Intel, and a marketing war running well ahead of independent testing.

What AMD Actually Announced

AMD’s Ryzen AI Max+ Pro 495, part of the Gorgon Halo lineup, is a refresh of the Strix Halo architecture that debuted last year, according to Tom’s Hardware’s reporting on the benchmark release. The first Gorgon Halo-powered systems reportedly went on sale just days before AMD’s benchmark drop, with some configurations priced above $7,099. That is workstation money, not consumer-laptop money, and it signals AMD is chasing developers and AI engineers rather than mainstream buyers.

For its benchmark comparison, AMD did not test against Nvidia’s unreleased RTX Spark at all. Instead, it measured its top-spec 192GB configuration of the Ryzen AI Max+ Pro 495 against Intel’s Core Ultra X9 388H (a Panther Lake chip) running in a system with 64GB of memory, using ComfyUI to measure generative-AI throughput across multiple model runs. The reported advantage ranged from 1.1x to as high as 32.2x, though Tom’s Hardware flagged that top figure as a clear outlier tied to a model it could not identify or verify on Hugging Face. Strip out that one result and the real-world gap looks a lot more modest, closer to the 1.1x-to-2x range typical of generational hardware differences.

That is an important caveat for anyone reading AMD’s press materials at face value. A single outlier result pulling an average up by 30x tells you more about one unoptimized AI model than it does about silicon. Readers comparing independent CPU benchmark data from UL will notice that generational jumps between competing chip families rarely exceed 2x in broad workloads, which makes AMD’s outlier figure worth treating with real skepticism until third-party labs run their own tests.

Why Nvidia’s RTX Spark Timing Matters

Nvidia has been teasing RTX Spark for months, and the chip at its center, officially named N1X, is already well documented thanks to earlier leaks and Nvidia’s own disclosures. As Tom’s Hardware previously reported, N1X pairs an Arm-based Grace CPU with a Blackwell RTX GPU and comes in two configurations: a higher-end version with a 20-core CPU and 6,144 CUDA cores, and a lower-end version with an 18-core CPU and 5,120 CUDA cores. Both top out at 128GB of unified memory, which Tom’s Hardware notes is identical to the GB10 chip used in Nvidia’s own DGX Spark, since the two chips are nearly the same silicon aimed at different form factors.

Leaked Geekbench 7 results cited in reporting this week put the N1X’s single-core scores at 2,541 and 2,570 across the two configurations, with multi-core scores of 21,776 and 23,126. For context, a competing high-end processor cited in the same comparison data scored 3,207 single-core and 22,197 multi-core, suggesting N1X’s strength lies far more in parallel, GPU-adjacent workloads than in raw single-thread speed. That tracks with Nvidia’s own positioning: N1X isn’t built to win spreadsheet benchmarks, it’s built to run CUDA, TensorRT, and DLSS workloads locally without a cloud GPU rental.

The expected October 7 launch window lines up with a Microsoft event, hinting that Windows-on-Arm positioning will be part of Nvidia’s pitch. That is a notable shift. Readers who tracked how Nvidia split its DGX Spark lineup into 64GB and 128GB tiers earlier this year, or how RTX Spark N1X first shipped with 6,144 CUDA cores, will recognize the pattern: Nvidia segments its AI-PC silicon the same way it segments its data-center GPUs, by memory tier and core count, and lets partners set the final price.

The Memory Capacity Argument

Strip away the marketing noise and the Gorgon Halo versus RTX Spark fight reduces to one real engineering trade-off: memory capacity versus raw throughput. AMD’s top Ryzen AI Max+ Pro 495 configuration supports up to 192GB of unified memory. Nvidia’s RTX Spark platform caps out at 128GB. That 64GB gap matters enormously for anyone trying to run large open-weight language models locally, since memory capacity, not compute, is usually the hard ceiling on which model sizes fit on a single machine without offloading to disk or network storage.

Running a bigger model locally on more memory generally means a lower tokens-per-second rate than a smaller model squeezed into faster, more optimized silicon. AMD is betting developers care more about fitting a 70-billion-parameter model entirely in memory than about shaving milliseconds off inference latency. Nvidia is betting the opposite: that its CUDA software stack, DLSS, and Blackwell GPU architecture will win on raw speed and on the sheer size of its existing developer ecosystem, even with a smaller memory ceiling.

It’s worth noting both companies are explicitly chasing a market that barely existed two years ago. AMD’s own framing, calling these “agentic PCs” rather than AI PCs, points to a shift in messaging: the sales pitch is no longer about chatbots and image generators running locally, it’s about autonomous software agents that need local memory and compute to operate without constant cloud round-trips.

AMD’s Shipping Numbers Don’t Add Up Cleanly

The most telling moment from AMD’s press briefing wasn’t a benchmark at all. According to Tom’s Hardware’s account, AMD representatives initially told press that the company had shipped “10s of millions” of AI PCs, only to clarify moments later that the real figure, specific to Strix Halo and Gorgon Halo agentic-PC silicon, was “over half a million” units. That is a difference of roughly two orders of magnitude between the first claim and the walked-back correction.

The discrepancy likely reflects two different categories getting conflated in the room: AMD almost certainly has shipped tens of millions of Ryzen chips broadly classified as “AI PCs” under Microsoft’s Copilot+ PC branding umbrella, which covers any processor with a qualifying neural processing unit. But Gorgon Halo and Strix Halo specifically target the higher-memory, higher-wattage “agentic PC” tier that competes directly with RTX Spark and Apple’s larger M-series chips, and that is a much smaller, newer segment measured in the hundreds of thousands, not tens of millions.

For a company trying to convince developers and enterprise buyers that its hardware has real market traction, getting caught overstating a number by roughly 20-to-40x in the same briefing is an unforced error. It also hands Nvidia an easy talking point heading into its own launch window: if AMD can’t keep its own numbers straight, why trust its benchmark comparisons against Intel?

Competitive Landscape: AMD, Nvidia, and Intel

The table below lines up the three chips currently fighting over the agentic-PC category, using the specs each company has disclosed or that have leaked through credible benchmark databases.

ChipMakerCPU ConfigGPU / ComputeMax Unified MemoryStatus (Oct. 6, 2026)
Ryzen AI Max+ Pro 495 (Gorgon Halo)AMDx86, Strix Halo-derivedIntegrated RDNA GPU192GBShipping, systems from $7,099
RTX Spark N1X (high config)Nvidia20-core Grace (Arm)6,144 CUDA cores (Blackwell)128GBExpected launch Oct. 7, 2026
RTX Spark N1X (low config)Nvidia18-core Grace (Arm)5,120 CUDA cores (Blackwell)32GBExpected launch Oct. 7, 2026
Core Ultra X9 388H (Panther Lake)Intelx86, Panther LakeIntegrated Arc GPU128GB (platform max)Shipping, used as AMD’s benchmark baseline

Notice that AMD didn’t actually benchmark against Nvidia at all, since RTX Spark hardware wasn’t publicly available to test yet. That’s a legitimate constraint, but it also means every performance claim AMD published this week is an AMD-vs-Intel comparison dressed up in anti-Nvidia messaging. Readers should wait for independent third-party reviews, ideally from outlets that test all three platforms side by side on identical workloads, before drawing conclusions about which chip actually wins the agentic-PC race.

Geekbench Leaks: What the Early Numbers Show

While AMD’s benchmarks focus on generative-AI throughput, leaked Geekbench 7 results give a different lens on RTX Spark’s raw CPU performance. The table below compares the leaked N1X scores against the Intel chip AMD used as its own baseline, plus a high-end x86 desktop part for scale.

ProcessorGeekbench 7 Single-CoreGeekbench 7 Multi-CoreList Price
RTX Spark N1X (high config, leaked)2,57023,126Not yet announced
RTX Spark N1X (low config, leaked)2,54121,776Not yet announced
Intel Core Ultra 9 Processor 285K3,207 (cited comparison)22,197 (cited comparison)$589
AMD Ryzen Threadripper PRO 9995WXn/a (workstation-class)29,265n/a

The gap in single-core performance is the headline here. N1X’s Arm-based Grace CPU trails a top desktop Intel chip by roughly 20% in single-thread Geekbench scores, while staying competitive in multi-core workloads. That is consistent with how Arm-based designs are typically tuned: wide and efficient rather than clocked for peak single-thread speed. For agentic workloads that lean on parallel inference rather than single-thread bursts, that trade-off may not matter much in practice, but it will show up in everyday tasks like app launches or single-threaded scripting.

Historical Context: From Strix Halo to Gorgon Halo

Gorgon Halo didn’t appear out of nowhere. It’s a direct successor to AMD’s Strix Halo platform, which established the “big APU with huge unified memory” formula that AMD is now defending against Nvidia’s incoming hardware. The category itself is barely two years old, and it emerged specifically because running large language models locally requires memory bandwidth and capacity that traditional discrete-GPU laptops couldn’t offer without eye-watering VRAM costs.

Intel’s entry into this fight came later, through Panther Lake and the Core Ultra X9 lineup, which AMD used as its benchmark punching bag this week rather than engaging Nvidia directly. Intel has leaned on its Core Ultra platform and Microsoft’s Copilot+ PC certification program to stay relevant in the conversation, even without a chip that matches AMD’s or Nvidia’s memory ceilings.

Apple sits adjacent to this fight rather than squarely inside it. Tom’s Hardware’s reporting specifically frames Gorgon Halo as targeting RTX Spark primarily, and Apple’s larger M-series chips only secondarily, since Apple’s own M6 and M5 silicon runs a closed hardware-software stack that doesn’t compete on the same open Windows-PC terms. Still, any laptop buyer shopping for a machine that can run large models locally now has to weigh AMD, Nvidia, Intel, and Apple options against each other, something that wasn’t really possible before 2025.

Market Impact: Why This Fight Matters Beyond Specs

The agentic-PC category is small today. AMD’s own corrected figure puts it at roughly half a million units shipped, but it sits at the intersection of two much larger trends: the push to run AI workloads outside expensive cloud GPU rentals, and the broader PC industry’s search for a reason to drive upgrade cycles after years of relatively flat demand. Every major chipmaker has a financial incentive to own this category early, even if the current volumes are modest.

For AMD, Gorgon Halo sits alongside its broader data-center push. The company has been expanding its AI silicon ambitions across the stack, from client chips like Gorgon Halo down to server parts, a strategy visible in how AMD’s EPYC-based Helios platform landed a $1.2 billion HPE and Vultr deal earlier this year. A strong local-AI-PC story reinforces that broader enterprise pitch: buy AMD silicon from the laptop to the data center.

For Nvidia, RTX Spark represents an attempt to bring its data-center CUDA dominance down to the desktop and laptop, something it has struggled to do as effectively as AMD and Intel historically, given Nvidia’s lack of an x86 or long-standing client-CPU business. Partnering with Arm for the Grace CPU side of N1X is Nvidia’s workaround, and if it succeeds, it opens a new front against both AMD and Intel in a market Nvidia has never directly competed in before: general-purpose PC processors.

Pricing will ultimately decide who wins more than benchmarks will. Gorgon Halo systems starting above $7,099 put AMD’s flagship configuration firmly in professional workstation territory, pricing out most consumers. If Nvidia prices RTX Spark systems meaningfully lower, even with a smaller 128GB memory ceiling, it could win the volume argument even while losing the memory-capacity argument.

What Independent Testing Will Need to Settle

Three questions remain open until outlets can test real retail hardware side by side. First, does AMD’s 1.1x-to-32.2x ComfyUI advantage over Intel hold up once the outlier model is excluded and more workloads are tested? Second, how does RTX Spark’s actual shipping performance compare to the leaked Geekbench figures, which are unverified third-party leaks rather than Nvidia-published numbers? Third, and most practically, what will finished RTX Spark systems actually cost at retail, since Nvidia has not announced consumer pricing as of October 6, 2026?

Until those answers arrive, buyers evaluating a Snapdragon X2, Core Ultra, or AMD Ryzen AI machine for local AI work are stuck comparing marketing claims rather than apples-to-apples results. That’s a familiar pattern in chip launches, but it’s worth remembering that AMD’s own numbers this week needed a correction within the same briefing.

5 Predictions for the Agentic-PC Market

  • Nvidia will emphasize software and CUDA ecosystem lock-in rather than raw spec sheets when RTX Spark formally launches, since its memory ceiling trails AMD’s by 64GB.
  • Independent benchmarks published in the weeks after RTX Spark’s launch will likely show AMD’s 32.2x outlier claim was not representative of typical workloads, settling closer to a 1.5x-to-2x average advantage over Intel’s Panther Lake chips.
  • Intel will respond with aggressive pricing on Core Ultra X9 systems rather than a new chip, since Panther Lake is already shipping and a true answer to 192GB unified memory configurations isn’t imminent.
  • AMD will continue blurring the line between broad “Copilot+ PC” shipment figures and narrower “agentic PC” figures in future marketing, making third-party unit-tracking firms more important for accurate market sizing.
  • Expect at least one more memory-capacity leapfrog move within two product cycles, as all three companies chase the ability to run ever-larger open-weight models entirely in local memory.

Frequently Asked Questions

What is AMD Gorgon Halo?
Gorgon Halo is the code name for AMD’s current generation of high-memory APUs aimed at local AI workloads, with the Ryzen AI Max+ Pro 495 as its flagship chip. It’s a refresh of the earlier Strix Halo platform, supporting up to 192GB of unified memory.

When does Nvidia’s RTX Spark launch?
RTX Spark is expected to launch around October 7, 2026, timed to a Microsoft event, according to reporting cited by Tom’s Hardware. Nvidia had not confirmed final consumer pricing as of October 6, 2026.

What chip powers RTX Spark?
RTX Spark devices use Nvidia’s N1X chip, which pairs an Arm-based Grace CPU with a Blackwell RTX GPU. It ships in two configurations: a 20-core CPU with 6,144 CUDA cores, and an 18-core CPU with 5,120 CUDA cores.

How much memory can Gorgon Halo and RTX Spark support?
AMD’s Ryzen AI Max+ Pro 495 supports up to 192GB of unified memory in its top configuration. Nvidia’s RTX Spark tops out at 128GB, the same ceiling as Nvidia’s own DGX Spark GB10 chip.

Did AMD actually benchmark against Nvidia’s RTX Spark?
No. AMD compared its Ryzen AI Max+ Pro 495 against Intel’s Core Ultra X9 388H, since RTX Spark hardware was not yet publicly available to test. The comparison used ComfyUI to measure generative-AI throughput.

How many agentic PCs has AMD actually shipped?
AMD corrected an initial claim of “10s of millions” of AI PCs down to “over half a million” agentic PCs specifically, referring to its Strix Halo and Gorgon Halo lineup, according to Tom’s Hardware’s account of the briefing.

What does “agentic PC” mean?
It’s industry shorthand for PCs with enough local memory and compute to run AI agents and large language models directly on the device, without relying on constant cloud access. AMD, Nvidia, and Intel are all now marketing hardware specifically for this category.

How much do Gorgon Halo systems cost?
The first Gorgon Halo-powered systems launched days before AMD’s benchmark release, with some configurations priced above $7,099, putting them in professional workstation territory rather than mainstream consumer pricing.