On September 22, 2026, Qualcomm simultaneously unveiled the 2nm "Snapdragon 8 Elite Extreme Gen 6" and "Snapdragon 8 Elite Gen 6" at its Snapdragon Summit in Maui. The company touted the Prime core's peak 5.0GHz as "the world's first 5GHz mobile CPU," but when placed alongside official documentation for the previous-generation Gen 5, the maximum clock speed increase amounts to only about 5.5%. The larger claimed improvements are actually in GPU performance and power efficiency, and many new features appear only in the Extreme variant's materials. Notably, the standard model's documentation has also dropped mention of APV recording support, which Gen 5 offered. What exactly changed between generations for these two Gen 6 chips—and what stayed the same?
From 4.74GHz to 5.0GHz: A 5.5% Clock Increase
Gen 5's official product brief listed the Prime core's maximum frequency at 4.74GHz. Both Gen 6 variants hit 5.0GHz, a difference of 0.26GHz. The Prime core's peak frequency rose approximately 5.5% from Gen 5's 4.74GHz to 5.0GHz. The calculation—(5.0−4.74)÷4.74×100—yields 5.49%, rounded to one decimal place.
The figure shifts slightly depending on the comparison point. Gen 5 also had a variant capped at 4.6GHz; using that as the baseline, the increase becomes approximately 8.7%. The six Performance cores jumped from 3.62GHz to 4.0GHz, a roughly 10.5% increase—a larger percentage gain than the flagship Prime core, despite receiving less attention.
Clock speed gains and performance gains are separate figures. Roughly speaking, CPU processing performance is determined by multiplying "instructions processed per clock cycle" by "clock cycles per second." If only frequency rises by 5.5% with an unchanged design, performance gains should also land around 5.5%.
Qualcomm's claimed CPU performance improvements over Gen 5 are 13% for Extreme and 10% for the standard model—both exceeding the clock speed increase. It's natural to assume the gap is filled by non-frequency factors, such as the cache restructuring discussed below. However, Qualcomm hasn't broken down exactly what accounts for the difference.
There's also an unexplained gap between the standard and Extreme variants. Despite identical core configurations and maximum frequencies, their claimed CPU performance differs by 3 percentage points (13% vs. 10%). Both figures are internal Qualcomm numbers annotated as "subject to change based on OEM implementation," with no independent verification yet available. Qualcomm's materials don't explain the 3-point discrepancy.
Compared to rivals, the 5.0GHz figure itself stands out. MediaTek's officially stated Dimensity 9600 Pro Prime core tops out at 4.55GHz, while Samsung's Exynos 2600 sits at 3.8GHz. The claim of "world's first 5GHz mobile CPU" should be read accurately as Qualcomm's own assertion. Higher clock speeds don't necessarily translate to superior performance rankings across CPUs with different architectures.
Total Cache Shrinks by a Third, Shifts to Shared Design
Behind the clock speed story, the CPU's memory architecture also changed. According to PC Watch (Kazuki Kasahara, September 23, 2026), Gen 5 gave each of the Prime and Performance clusters 12MB of cache, totaling 24MB. Gen 6 replaces this with a single 16MB "Oryon Flex Cache" shared across all 8 cores. Worth noting: Gen 5's 12MB×2 configuration wasn't listed in Qualcomm's official brief—it comes from media reporting.
Total CPU cache dropped from Gen 5's 24MB (12MB × 2 clusters) to 16MB—a one-third reduction. The math: (16−24)÷24 = −33.3%. In theory, a shared pool design allows the Prime core access to the entire 16MB. This possibility stems from the shared-cache architecture itself; no official Qualcomm documentation states that the Prime core can actually access the full 16MB in practice.
Cache is small, fast memory placed close to the CPU. The more frequently-used data that fits here, the fewer trips needed to slower, more distant main memory—reducing CPU wait time. With per-cluster dedicated cache, even if the Performance cluster is idle, the Prime core can't borrow its 12MB. A shared design lets capacity flow toward whichever core is currently under heavy load. In scenarios where load concentrates on a few cores—like launching apps or rendering webpages—this design uses capacity more efficiently.
The picture changes, however, when all 8 cores run heavy workloads simultaneously. Dividing 16MB evenly across 8 cores works out to 2MB per core. Qualcomm hasn't disclosed the details of its allocation scheme, so it's possible the Prime core doesn't always get access to the full 16MB. Neither generation's comparison includes the SLC (System Level Cache) shared across the entire chip.
How much this cache restructuring accounts for the gap between the 5.5% clock increase and the 13% CPU performance claim remains unclear. Still, framing Gen 6's CPU improvements purely in terms of clock speed overlooks this separate architectural change.
Lining Up Gen 5 and the Two Gen 6 Variants, Item by Item
Qualcomm issued separate product briefs for each of the two new chips, formatted almost identically to the Gen 5 brief. Pulling matching items from all three documents produces the following table. All improvement percentages are Qualcomm's official figures relative to Gen 5, annotated by the company as "subject to change based on OEM implementation."
| Item | Gen 5 | Gen 6 (Standard) | Gen 6 Extreme |
|---|---|---|---|
| Model number | SM8850-AC, etc. | SM8950 | SM8975 |
| Manufacturing process | 3nm | 2nm | 2nm |
| Prime/Performance max frequency | 4.74GHz/3.62GHz | 5.0GHz/4.0GHz | 5.0GHz/4.0GHz |
| Memory | LP-DDR5x, up to 5300MHz | LP-DDR5x/6, up to 5300MHz | LP-DDR5x/6, up to 5300MHz |
| Max video recording | 4K 120fps | 8K 30fps | 8K 60fps |
| Slow motion | 1080p 480fps | 4K 120fps, 1080p 960fps | 4K 240fps, 1080p 960fps |
| APV recording | Supported | Not listed | Supported |
| Dedicated GPU memory (Adreno HPM) | 18MB | Architecture mentioned only, no capacity listed | 18MB |
| GPU AI features | No mention of matrix cores | No mention of Neural Fusion | Adreno matrix cores, Neural Fusion |
| NPU shared memory | Unconfirmed | No increase mentioned | 50% increase |
| H.266 decoding | Unconfirmed | Not listed | Supported |
- Gen 6 Extreme
- Gen 6 Standard
データを表で見る
| Gen 6 Extreme (%) | Gen 6 Standard (%) | |
|---|---|---|
| CPU Performance | 13 | 10 |
| GPU Performance | 44 | 35 |
| NPU Performance | 35 | 14 |
| AI Performance/W | 33 | 20 |
| CPU Power Efficiency | 37 | 37 |
| GPU Power Efficiency | 40 | 40 |
The chart shows the largest gap between the two variants is in NPU performance: 35% for Extreme versus 14% for the standard model. Meanwhile, CPU and GPU power efficiency improvements are identical across both chips at 37% and 40% respectively—suggesting the benefits of the 2nm process reach the standard model equally. Note that "power efficiency improvement" means delivering the same performance using less power—it doesn't mean power consumption drops by 37%.
The table also reveals where improvements are concentrated. Comparing Gen 5, Gen 6 Standard, and Gen 6 Extreme item-by-item in the official briefs shows that Neural Fusion, 8K60 recording, and the 50% NPU shared memory increase appear only in the Extreme brief, while APV recording—which Gen 5 supported—is absent from the standard model's brief. Android Authority's Hadlee Simons called the standard model's lack of this recording feature "especially disappointing."
Some figures remained unchanged. Maximum memory speed stays at 5300MHz across all three chips—aside from adding LPDDR6 support, the official number hasn't moved. The Extreme's dedicated GPU memory also holds steady at 18MB, matching Gen 5; the improvement Qualcomm highlights instead is a 12% power reduction through a redesigned structure.
Reading these documents requires caution. The absence of an item in a brief doesn't confirm the feature's absence. The standard model's brief doesn't mention matrix cores under the GPU section, yet its AI Engine section states "Adreno GPU with matrix cores." The standard model's product page lists HPM capacity as "12 GB"—likely a typo for MB, though the actual capacity can't be confirmed. Whether the standard model includes a dedicated AI processing circuit within its GPU cannot currently be determined from official documentation.
Gaming Super-Resolution and Expanded NPU Memory: Two Features Exclusive to the Top-Tier Model
The Extreme brief explicitly lists "Adreno matrix cores" under GPU specifications—a dedicated AI processing circuit built into the GPU. This powers "Adreno Neural Fusion," which handles gaming super-resolution and frame generation. Qualcomm states that in games supporting Neural Fusion, rendering power consumption can be reduced by up to 40%.
The power savings stem from how GPU workload is reduced. In super-resolution, the game is first rendered at lower resolution, and AI estimates the missing pixels to produce a higher-resolution image. In frame generation, AI-estimated intermediate frames are inserted between actually-rendered frames. Both approaches reduce the amount of pixel-by-pixel calculation the GPU must perform, offloading that work to inference instead.
Delegating inference to dedicated matrix-multiplication circuits rather than general-purpose compute units allows the same workload to run on less power. It's worth noting that the 40% figure represents a maximum in supported games—not every game will see reduced power consumption.
NPU architecture also changed. Gen 5's Hexagon NPU combined 12 scalar units, 8 vector units, and 1 accelerator. Gen 6 keeps 12 scalar and 8 vector units, adds 1 tensor unit, and introduces a newly designed "Element Accelerator." Support for numerical precision ranging from INT2 to FP16 was already present in Gen 5. In other words, what's new in Gen 6's NPU is the combination of processing units, not the range of supported number formats.
The Extreme variant additionally increases the NPU's shared memory by 50%. This is said to enable running Mixture of Experts (MoE) models on-device with over 30B (30 billion) parameters. HotHardware reports that in Qualcomm's example, actual parameters used per token were approximately 3B.
MoE is a model architecture where, among many "expert" networks, only a subset is invoked for each input. Even though only about 3B parameters are used per calculation, which experts get called depends on the input, so all 30B parameters must remain accessible on-device. Memory capacity—not computational speed—determines the maximum model size a device can run. Given that the 50% shared memory increase is exclusive to Extreme, running large AI models on-device appears to be architected around the Extreme variant specifically.
HotHardware wrote about this 50% increase as though it were a shared specification across both chips. However, the official brief lists "50% Larger shared memory" only for the Extreme, and both 9to5Google and Notebookcheck also treat this as Extreme-exclusive.
Qualcomm frames this two-tier structure as a deliberate strategy. Chris Patrick, SVP and GM leading the mobile handsets business, stated that the multi-flagship strategy allows the company to "push the boundaries of what's possible in mobile while bringing the latest AI technology to a broader range of premium devices." The framework assigns Extreme the role of pushing upper limits, while the standard model delivers AI features to a wider price range. Xiaomi's William Lu has stated the company will differentiate use cases for the two chips across its flagship lineup.
Fourth to Market on 2nm: How Far Do Qualcomm's "Firsts" Actually Hold Up?
Gen 6 isn't the first smartphone SoC announcement built on a 2nm process. Lining up each company's announcement dates tells the story:
| Announcement Date | Product | Key Claimed Specs at Launch |
|---|---|---|
| December 19, 2025 | Samsung Exynos 2600 | Samsung Foundry 2nm GAA, Prime core max 3.8GHz |
| September 9, 2026 | Apple A20 Pro | Up to 40% GPU improvement, 50% increased memory bandwidth |
| September 15, 2026 | MediaTek Dimensity 9600 Pro | Prime core max 4.55GHz, native 4K240 output, up to 30B LLM support, LPDDR6 support |
| September 22, 2026 | Snapdragon 8 Elite Extreme Gen 6 / 8 Elite Gen 6 | Prime core max 5.0GHz, 8K60 recording (Extreme) |
2nm smartphone SoCs were announced in this order: Exynos 2600 (December 19, 2025), A20 Pro (September 9, 2026), and Dimensity 9600 Pro (September 15, 2026), with Snapdragon 8 Elite Gen 6 (September 22, 2026) arriving fourth. This reflects announcement order, not the order in which devices launch or mass production begins. Each company's "2nm" involves different fabs and processes, so transistor density and other characteristics aren't necessarily comparable. Notebookcheck's announcement article headline called it the "world's first 2nm chip"—a claim that doesn't hold up against this timeline.
What about Qualcomm's other claims of "firsts"? Regarding the 5GHz mobile CPU, the highest confirmed rival maximum frequency is the Dimensity 9600 Pro's 4.55GHz, so the claim holds as an official figure at time of announcement. As for 8K60fps video recording, Qualcomm's brief states it "enables the first commercial device capable of recording 8K60." Overseas outlets FoneArena and GSMArena reported that the Dimensity 9600 Pro also supports 8K60, though MediaTek's official press release text contains no mention of 8K. The 8K60 "first" claim is best understood as Qualcomm's own characterization.
The situation differs for 4K240fps slow motion. MediaTek's official press release explicitly states the Dimensity 9600 Pro can natively output 4K240. The vivo X500 Pro Max, which uses this chip, is scheduled to launch in China on September 24. While 4K240 is new for Snapdragon, it's not a first for smartphones overall. Given that Gen 6's most significant gains lie in GPU and NPU performance, the more meaningful competitive comparison isn't clock speed or video "firsts"—it's how far ahead Qualcomm can push AI and gaming performance-per-watt.
Undisclosed Specs and a 9-Company Adoption List: What to Verify Once Devices Arrive
Qualcomm's official press release and the two product briefs omit the manufacturing partner's name, memory bus width, NPU TOPS figures, and absolute shared memory capacity. The manufacturing process is described only as "2nm Process Technology." While speculation points to TSMC's 2nm (N2P), South Korea's Chosun Biz reported that Qualcomm is in discussions with Samsung about 2nm manufacturing. That report cited Samsung's yield rate at approximately 55%, though neither claim has been officially confirmed.
For memory, only "LP-DDR5x/6, up to 5300MHz" and a maximum capacity of 24GB are disclosed—no bus width, and no speed figures specific to LPDDR6 operation. For the NPU, only the "50% shared memory increase" and configuration details are given; neither TOPS performance metrics nor the post-increase absolute capacity are provided. HotHardware also reported that Qualcomm did not disclose the absolute shared memory capacity. The official press release itself contains no performance or specification figures at all, instead directing readers to product pages and briefs.
Even disclosed comparisons require careful reading of the conditions behind them. The Extreme brief lists the X105 5G modem (up to 14.8Gbps downlink) alongside the Wi-Fi 8-capable FastConnect 8800, describing the latter as "2x faster" with "3x the range at gigabit speeds." This comparison is an internal iperf measurement pitting the new product's 4x4 antenna and 320MHz bandwidth configuration against the previous generation's 2x2 antenna and 160MHz bandwidth setup. It's a comparison between configurations where both antenna count and bandwidth doubled—not a generational comparison under matched conditions.
The roster of adopting manufacturers has also shifted. The official press release names nine companies: HONOR, iQOO, Motorola, OnePlus, OPPO, REDMI, RedMagic, vivo, and Xiaomi. Absent are Samsung, Sony, POCO, realme, ROG, Nubia, and ZTE—all named in the Gen 5 release dated September 24, 2025. Absence from the list doesn't necessarily mean non-adoption, but for Japanese readers, the absence of Sony—maker of the Xperia line—stands out. As of now, no Japan launch plans for Gen 6-equipped devices have been confirmed.
Only a handful of specific device timelines have been announced. Vivo stated that the iQOO 16 and iQOO Pad Ultra tablet will be among the first devices to feature the Extreme chip. RedMagic will use it in the 12Pro+, and Motorola is adopting Extreme in its flagship lineup for the first time. Motorola's Signature 27 will launch in North America "in the coming months," backed by a promised seven years of OS updates.
Pricing is trending upward. Bloomberg reported on July 24, 2026, that Qualcomm notified customers via letter of double-digit percentage price increases for products shipping after September 1. Qualcomm itself has not disclosed specific figures. With 2nm manufacturing costs and rising memory prices converging, positioning the standard model for "a broader range of premium devices" may reflect the reality that only a limited number of device makers can absorb Extreme's pricing.
The figures that matter most will only be confirmed once devices actually reach the market. The first thing to verify is how closely third-party measurements match the claimed 37% CPU and 40% GPU power efficiency improvements. If the standard model can deliver comparable power efficiency at a lower price point than Extreme, it could end up being chosen for real-world battery life rather than headline clock speed numbers.