Qualcomm's Adreno X2 Claims 2.3x Gains, Omits the Baseline

Qualcomm claims its Adreno X2 GPU is 2.3 times as power-efficient as its predecessor. Testing this month by Chips and Cheese on a shipping Asus Zenbook A16 found real gains in some workloads and clear regressions in others that Qualcomm's own comparison did not disclose.

The Adreno X2 ships inside Qualcomm's second-generation Snapdragon X2 Elite chips, the Arm processors Qualcomm sells to PC makers as an alternative to Intel and AMD. Chips and Cheese tested the GPU on Sept. 19, 2026, in an Asus Zenbook A16 supplied by Asus, running a 2880x1800 OLED display. A separate architectural teardown by SemiAccurate, published Nov. 26, 2025, laid out the silicon changes and flagged what Qualcomm's own performance slides leave out.

Qualcomm's 2.3x claim, and what it omits

Qualcomm's marketing puts three numbers on the Adreno X2: a 2.3x gain in performance per watt over the first-generation X Elite, a 70 percent lead over Adreno X1 in 3DMark Time Spy at matched power, and a 125 percent gain in performance per watt in the same test. SemiAccurate's Charlie Demerjian reported those figures from Qualcomm's own comparison slides, then flagged what was missing from them: no disclosure of the reference system's RAM configuration or storage speed, no disclosed thermal design power for either the Adreno X2 configuration tested (20 watts, by Demerjian's account) or the 100-watt-plus competing systems it was measured against, and an incomplete build sheet for the AMD Ryzen AI 9 HX 370 machine used as the rival. Demerjian called the comparison methodology "borderline dishonest" and wrote: "We don't doubt it is much better but we won't repeat marketing claims."

Qualcomm's claim	Comparison basis	Figure
GPU performance per watt	vs. first-generation Adreno X1	2.3x
3DMark Time Spy	vs. Adreno X1, matched power	+70%
Performance per watt, 3DMark	vs. Adreno X1	+125%
CPU performance	vs. prior generation	up to 39%

What changed in the silicon

Adreno X2 moves from three slices of six shader processors on the original Adreno X1 to four slices of eight shader processors, each pair grouped into what Qualcomm calls a micro shader processor texture processor. Clock speed rose from 1.5 GHz to 1.85 GHz. The design trades the X1's 128-wide instruction wave for a narrower 64-wide wave with dual-issue capability, and the register file per texture processor shrank from 192 KB to 128 KB. Local high-performance memory grew to 21 MB, split into 5.25 MB per slice, alongside a doubled 2 MB shared L2 cache and an 8 MB system-level cache, feeding a 192-bit LPDDR5X-9523 memory interface. Each shader processor pushes 2,048 32-bit arithmetic units, moving 128 texels and setting up four triangles per clock, with 64 pixels and 128 fragments emerging from eight back-ends per cycle.

Where independent testing backs the vendor

Chips and Cheese confirmed real graphics gains in workloads Qualcomm's own slides highlight: rasterization performance in 3DMark Wild Life Extreme, ray tracing throughput in the lighter scenes of 3DMark Solar Bay, and a doubling of frame rates in Cyberpunk 2077 compared with Adreno X1. Integer math also improved. INT32 additions now run at full rate, a change Chips and Cheese called an alignment with rival designs from AMD, Intel and Nvidia rather than a regression.

Where it falls apart

The same testing found the Adreno X2 regresses on general-purpose compute, the kind of workload that treats a GPU as a parallel processor rather than a graphics pipeline. On FluidX3D, a fluid-dynamics benchmark, the X2 measured worse than its own predecessor. On the FAHBench nav test, Nvidia's GTX 1050 3GB, a budget desktop card several architecture generations old, ran 36 percent faster than the Adreno X2. Chips and Cheese traced part of the shortfall to the memory system: of the 21 MB of high-performance local memory on the chip, only 1 MB can be allocated to compute workloads, and cache-miss latency climbs as memory footprints grow past that limit. The design also still lacks hardware fused multiply-add, the single-instruction operation most general-purpose compute code depends on, a limitation carried over unchanged from Adreno X1.

Chips and Cheese's own conclusion draws the line plainly: "It feels like Qualcomm took a specific set of graphics workloads and optimized for them, while AMD, Intel, and Nvidia set out to build more general purpose designs with robust performance characteristics across a broader range of use cases."

A pattern this site keeps finding in vendor benchmarks

Adreno X2 is not an isolated case. OpenAI's own comparison for its Jalapeño accelerator chip drew similar criticism for testing conditions that favored the newer part, and PrismML's Ternary Bonsai 2 27B shipped a headline score that did not hold up once outside testers ran the same benchmark independently. In each case, the vendor's number was not fabricated. It was measured under conditions the vendor chose and did not fully disclose, which is a different problem with the same fix: publish the reference configuration, the thermal limits and the software versions alongside the result.

What would change this read

If Qualcomm publishes the full reference-system configuration behind its 2.3x figure, including RAM, storage and thermal design power on both sides of the comparison, and a second lab reproduces something close to that number under matched conditions, the vendor claim gets more credible. Until that happens, the confirmed gains are the ones Chips and Cheese measured directly on hardware anyone can buy: roughly double the Cyberpunk 2077 frame rate of Adreno X1, and a real lead in light ray-tracing scenes. The efficiency multiplier on Qualcomm's slide remains a number only Qualcomm has tested.