A visual reference for understanding processor and GPU performance across phones, laptops, and desktops. Approximate benchmark scores for iconic chips plotted over time — with context on what the numbers actually mean, which benchmarks matter, and what the charts don't cover (tablets, consoles, servers, integrated GPUs, AI accelerators).
Benchmarks are standardized tests that stress specific parts of a chip. No single benchmark tells the whole story — each measures something different. Here's what each one used on this page actually tests, and what it means in practice.
Runs a gauntlet of real-world tasks — text compression, PDF rendering, photo editing, machine learning inference — on a single CPU core. Measures how fast one core can complete a task. This is the metric that determines how snappy your phone feels when opening apps, scrolling, or loading web pages.
Renders a complex 3D scene using the Cinema 4D engine across all available cores. Purely CPU-bound — no GPU involved. The multi-core score shows how well a chip handles parallelizable workloads like video encoding, code compilation, 3D rendering, and scientific simulation. Best for desktops where thermal constraints are minimal and sustained all-core loads are common.
Same 3D rendering test but restricted to a single core. Reveals raw per-core IPC (instructions per clock) and clock speed. Critical for tasks that can't be parallelized: game logic, application UI responsiveness, single-threaded build steps, web browsing, and everyday snappiness. Most laptop and desktop tasks depend heavily on this metric.
A DirectX 12 benchmark rendering complex game-like scenes with tessellation, volumetric lighting, and particle effects. The Graphics score isolates GPU performance from CPU bottlenecks. Correlates strongly with actual gaming frame rates at 1440p. The industry's de facto GPU comparison standard.
Benchmarks don't capture sustained performance under thermal limits, real-world battery drain, memory bandwidth effects, driver maturity, software optimization (Apple's tight integration advantage), or workload-specific behavior. A chip that benchmarks 20% higher might feel identical in daily use or worse in a thermally constrained chassis.
A desktop chip runs at 125–253W. A laptop chip runs the same architecture at 35–55W. A phone chip runs at 4–8W. Comparing across form factors is apples-to-oranges — the phone chip achieving even 40% of the desktop score at 3% of the power is an engineering marvel. Always compare within form factor.
Geekbench 6 single-core scores for flagship phone SoCs. Apple's A-series maintained a consistent ~30–50% lead in single-core IPC over Qualcomm for years, but 2025 marks a historic convergence: the Snapdragon 8 Elite Gen 5, Apple A19 Pro, and MediaTek Dimensity 9500 all cluster between 3,500–3,900 — the closest three-way race in mobile silicon history.
The A11 Bionic (2017) was a watershed — Apple's first neural engine and a massive IPC leap that opened a multi-year gap. Qualcomm responded with custom Oryon cores in the Snapdragon 8 Elite (2024), their most competitive single-core showing in years. By 2025, the race has fundamentally changed: Apple's A19 Pro (3,895), Qualcomm's SD 8 Elite Gen 5 (3,524), and MediaTek's Dimensity 9500 (3,635) are all within ~10% of each other in single-core. The era of Apple's unquestioned mobile CPU dominance may be ending — though Apple still leads in performance-per-watt by a significant margin.
Cinebench R23 single-core scores for notable laptop CPUs. Per-core speed determines how responsive a laptop feels — browsing, code editors, game logic, and application launch times all depend on it. By 2024–2025, all four laptop ecosystems (Intel, AMD, Apple, Qualcomm) sit within ~15% of each other, after a decade where Apple and Intel traded the lead. The widening gap in this chart isn't performance — it's the power each vendor burns to get there.
Intel's high clock speeds kept it per-core competitive even through the Skylake stagnation years (2015–2019, visible as the flat section). Apple's M1 (2020) was a landmark — ~1500 points from just 10W — though Intel's i7-11800H matched it at 4.5× the power draw. Lunar Lake (2024) flips the script for Intel: ~1800 points at just 17W makes it the most efficient x86 laptop core ever at 106 pts/W — nearly 3× Intel's own H-series. By 2025, all four platforms deliver similar per-core speed but differ dramatically in power consumption.
Cinebench R23 multi-core scores for flagship desktop CPUs. Desktop tells a story of AMD's comeback: trailing Intel significantly (2013–2016) before surging to leadership (2019–present). Intel's response with hybrid P+E core architectures in 2021 reignited competition. These chips run without meaningful thermal constraints, so scores represent near-peak silicon capability.
AMD's Zen architecture (Ryzen, 2017) broke Intel's stranglehold on desktop performance. The Ryzen 9 5950X (2020) was the first AMD chip to simultaneously lead in both single-core and multi-core. Intel's Alder Lake counter-attack (2021) with hybrid cores was genuinely innovative — using small efficient cores alongside big performance cores, a concept borrowed from mobile. By 2023, both companies routinely hit 35,000+ multi-core, up from ~5,500 a decade prior — a 6–7× improvement.
3DMark Time Spy Graphics scores for mobile GPUs. Laptop GPUs run at significantly reduced power compared to their desktop counterparts — typically 60–150W vs. 200–450W. NVIDIA dominates this space so completely that AMD mobile GPUs barely register in the market. Each generation brings roughly 20–30% improvement at the same power, with diminishing returns becoming visible in the RTX 40/50 series.
More than any other component, laptop GPU performance is hostage to the chassis. An RTX 4070 Laptop at 140W performs 40%+ better than the same chip at 80W in a thin machine. Manufacturers often use the same GPU name across wildly different power tiers. Always check the TGP (Total Graphics Power) spec — the model name alone is nearly meaningless for predicting real performance.
3DMark Time Spy Graphics scores for desktop GPUs, including both NVIDIA and AMD entries. Desktop GPUs are the most straightforward comparison — fixed power targets, consistent cooling, and well-understood driver stacks. The RTX 30-series (2020) was an inflection point with a massive generational leap. The RTX 5090 (2025) hit ~47,100 — the largest absolute score ever recorded in Time Spy — while AMD's RX 9070 XT offered remarkable value at ~30,000 for $550.
GamersNexus calls it "a mistake NVIDIA will never make again." At $699, the 1080 Ti offered near-Titan Xp performance — NVIDIA essentially cannibalized its own $1,200 card. When the RTX 2080 launched at $799 a year later with barely better rasterization, reviewers told everyone to just buy a 1080 Ti instead. It remained competitive for 5+ years, matching the RTX 3060 ($329) in 2022. No flagship GPU since has offered this combination: massive gen-on-gen leap (+35% over 1080), aggressive pricing ($50 over the 980 Ti's $649), 11GB VRAM for longevity, and a successor generation that validated rather than replaced it. The 1080 Ti moment is visible on this chart — it punched into a tier where it had no right to be at its price.
The RTX 5090 (2025) hit ~47,100 at $2,000. But AMD's RX 9070 XT delivers ~30,000 at $550 — matching the $1,000+ RTX 4080 at half the price. The RTX 5080 at ~33,000 essentially matches the 4090 for $1,000. For most gamers, the midrange sweet spot ($300–600) delivers 80% of the experience. The two charts below make this visible.
The per-category charts above each use different benchmarks, making cross-form-factor comparison impossible. Geekbench 6 runs natively on phones, laptops, and desktops — same workloads, same scoring — giving a single axis for all three. Time Spy runs on both laptop and desktop GPUs. Same benchmark, same scale, different devices.
The A19 Pro (3,895) now outscores every laptop Intel H-series chip (Ultra 9 285HX: ~3,000) and most desktop chips from 2023. Apple's M4 Pro (3,850) barely edges it out. Qualcomm's SD 8 Elite Gen 5 (3,524) beats every desktop AMD chip from 2023. The gap between form factors has collapsed in single-core — phones in 2025 deliver per-core IPC that would have been best-in-class on any platform just 2–3 years ago. The remaining desktop advantage is in sustained multi-core throughput, memory bandwidth, and the ability to run at full power indefinitely.
Read across: a 2025 flagship phone delivers per-core speed equivalent to a laptop from ~2024 and a desktop from ~2023–2024. Each row is roughly the same GB6 SC tier.
| GB6 SC Tier | 📱 Phone | 💻 Laptop | 🖥️ Desktop |
|---|---|---|---|
| ~1,200 | A11 · 2017 | i7-7700HQ · 2017 | R7 1800X · 2017 |
| ~1,700 | A15 · 2021 | i7-11800H · 2021 | i9-9900K · 2019 |
| ~2,500 | A17 Pro · 2023 | i7-12700H · 2022 | i9-12900K · 2022 |
| ~3,000 | SD 8 Elite · 2024 | Ultra 9 285HX · 2025 | i9-13900K · 2023 |
| ~3,500–3,900 | A19 Pro · 2025 | M4 Pro · 2024 | R9 9950X · 2024 |
A consistent pattern holds across GPU generations: a laptop midrange GPU roughly matches the previous generation's desktop midrange. The RTX 4070 Laptop (~13,400) matches the desktop RTX 3070 (~14,000). The RTX 3070 Laptop (~10,300) matches the desktop RTX 2070 (~10,800). And in 2025, the gap nearly vanishes: the RTX 5070 Laptop (~18,000) essentially equals the desktop RTX 4070 (~17,900). NVIDIA's improving power efficiency is compressing the mobile-to-desktop penalty from ~30% down to ~20% over five generations.
The A19 Pro at ~5W scores 779 GB6 pts/W. The R9 9950X at 170W scores 20 pts/W. That's a 39× efficiency gap — though this overstates the difference because desktop TDP reflects all-core power, not single-core draw (which is ~30–50W peak, making the real gap more like 10–15×). Even so, the phone achieving 115% of the desktop's single-core speed at a fraction of the power is remarkable. The M4 Pro at ~30W sits in the efficiency sweet spot: near-desktop single-core performance, near-phone efficiency, and enough thermal headroom for sustained workloads.
Absolute benchmarks show who's fastest — but not who gives you the most for your money, or who wastes the least power getting there. These paired charts normalize performance by launch price and power consumption across every category. Higher is better on all charts. Hover for chip names, prices, and wattages.
Apple delivered roughly 2× the benchmark score per watt of Snapdragon for a decade. That gap has nearly closed: the SD 8 Elite Gen 5 approaches Apple's efficiency while costing $300 less (Galaxy S26 vs iPhone 17 Pro). Per-dollar, Snapdragon has been the better value since 2021 because Samsung cut prices while Apple held at $999+. Caveat: phone price includes display, camera, memory — the SoC is only ~$40–80 of the BOM, so "per dollar" here reflects whole-device value, not chip value.
The per-watt chart tells a clear story: Lunar Lake at 17W delivers 106 pts/W, nearly 3× Intel's own H-series (42 pts/W) and comfortably ahead of Apple's M4 Pro (64 pts/W). Per-dollar is a closer fight — AMD's Ryzen H chips consistently edge out Intel on value thanks to competitive pricing at the $309–549 range. Apple is excluded from per-dollar because M-chips aren't sold standalone, but at the laptop level, a $2,000 MacBook Pro competes on total cost of ownership through battery life and longevity.
AMD's Zen architecture didn't just match Intel — it demolished the value equation. The Ryzen 7 5800X at $449 offered better multi-core per dollar than Intel's $589 i9-12900K. Per-watt, AMD Ryzen holds a consistent ~30% advantage because Intel's hybrid architecture (Alder Lake, Raptor Lake) pushes actual power consumption to 200–300W despite 125W PBP ratings. The 9800X3D at $479/105W is the current efficiency champion.
Laptop GPU efficiency improved steadily until the RTX 4060 Laptop (115W), which produces more compute per watt than any prior mobile GPU — a result of Ada Lovelace's efficiency-focused architecture. Per-dollar (using desktop MSRP as proxy since laptop GPUs aren't sold standalone), the midrange tracks close to desktop value. The RTX 50-series flagship laptops push TGP back above 150W, trading some of that efficiency gain for raw performance.
The 1080 Ti's per-dollar spike in 2017 is visible: 12.6 pts/$ when the flagship trend was declining. No flagship since has matched it. The midrange (x70) consistently offers 30–60% better value per dollar AND per watt than the flagship tier. The RTX 5090 at 575W delivers 82 pts/W — less efficient than the RTX 4060 at 115W delivering 91 pts/W, despite costing 7× more.
How much faster does each generation actually get? This table shows typical percentage improvements per generation for each category. Use this to calibrate your intuition — if someone says "30% faster," you'll know whether that's a normal generational bump or a genuine leap.
Typical: 10–20% per year
Big leaps: A11 (+44%), A17 Pro (+55%), SD 8 Elite (+39%)
2025: Convergence year — Apple, Qualcomm, MediaTek all within 10%
Typical: 8–15% per year
Big leaps: i7-11800H (+20%), i7-12700H (+21%), Alder Lake hybrid design
Stall years: 2017–2019 Intel (~2% per-core gains)
Typical: 15–30% per year
Big leaps: 12900K (+60% via hybrid), 13900K (+40%)
Stall years: 2013–2016 Intel (~5% per year, the "toothpaste era")
Typical: 30–50% per generation (2 yr cycles)
Big leaps: RTX 3070 (+48% over 2070), 5090 (+30% over 4090)
Value shift: RTX 5080 ≈ RTX 4090, RX 9070 XT ≈ RTX 4080
Benchmarks are only useful if you can map them to experience. Below is a rough translation from abstract scores to what you'd actually feel as a user. The bands are approximate — your specific software, settings, and expectations matter — but they'll anchor the numbers from the charts above.
For each benchmark, there's a threshold beyond which additional performance is nearly invisible to the user. In single-core, it's around 2,000 GB6 — roughly where a 2022 flagship phone or laptop sits. In GPU gaming, it's around Time Spy 14,000 for a 1440p 60fps gamer. Once you're above the threshold for your workload, you're paying for headroom and longevity, not perceptible speed. Knowing your threshold prevents overspending.
Performance charts show improvement, but they don't show what you paid for it. The "midrange" price point has shifted significantly over a decade — and the separation between "enthusiast" and "halo" tiers didn't exist before 2018. The GTX 1080 Ti story is partly a pricing story.
| Gen | x60 (Budget) | x70 (Midrange) | x80 (Enthusiast) | Halo (Ti/90) |
|---|---|---|---|---|
| Maxwell (2014) | GTX 960 · $199 | GTX 970 · $329 | GTX 980 · $549 | 980 Ti · $649 |
| Pascal (2016) | GTX 1060 · $249 | GTX 1070 · $379 | GTX 1080 · $549 | 1080 Ti · $699 |
| Turing (2018) | RTX 2060 · $349 | RTX 2070 · $499 | RTX 2080 · $699 | 2080 Ti · $1,199 |
| Ampere (2020) | RTX 3060 · $329 | RTX 3070 · $499 | RTX 3080 · $699 | 3090 · $1,499 |
| Ada (2022) | RTX 4060 · $299 | RTX 4070 · $599 | RTX 4080 · $1,199 | 4090 · $1,599 |
| Blackwell (2025) | RTX 5060 · $299 | RTX 5070 · $549 | RTX 5080 · $999 | 5090 · $1,999 |
| 10-yr change: Budget +50% ($199→$299) · Midrange +67% ($329→$549) · Enthusiast +82% ($549→$999) · Halo +208% ($649→$1,999) | ||||
Before 2018, NVIDIA's tiers were spaced predictably: roughly $200–$300 between each step. Turing changed the math. The RTX 2080 Ti at $1,199 created a new "halo" tier that had never existed in the consumer GPU market, and the x80 class jumped from $549–$699 to $1,199 in a single generation (RTX 4080). Meanwhile, the budget tier barely moved — the RTX 4060 at $299 costs roughly the same as a GTX 960 did a decade ago. The GTX 1080 Ti at $699 delivered what would become $1,199 performance in the next generation. That's why it's the GOAT: it sat at the old enthusiast price while delivering what became the new halo-tier performance.
Time Spy measures raw rendering power, but VRAM determines how long a GPU stays viable. A GPU with enough compute but not enough VRAM will stutter, drop textures, or force lower settings years before its shader performance becomes the limiting factor. This is the #1 reason real-world GPU lifespan diverges from benchmark predictions.
| Generation | x60 | x70 | x80 | Halo | AAA Demand |
|---|---|---|---|---|---|
| Maxwell (2014) | 2 GB | 4 GB* | 4 GB | 6 GB | 2 GB |
| Pascal (2016) | 6 GB | 8 GB | 8 GB | 11 GB | 3–4 GB |
| Turing (2018) | 6 GB | 8 GB | 8 GB | 11 GB | 4–6 GB |
| Ampere (2020) | 12 GB | 8 GB | 10 GB | 24 GB | 6–8 GB |
| Ada (2022) | 8 GB | 12 GB | 16 GB | 24 GB | 8–10 GB |
| Blackwell (2025) | 8 GB | 12 GB | 16 GB | 32 GB | 10–12 GB |
| * GTX 970 had 4 GB but only 3.5 GB at full bandwidth due to memory architecture split | |||||
The RTX 3070 launched with 8 GB of VRAM — the same amount as the GTX 1070 four years earlier. Game demand crossed the 8 GB threshold around 2022, which means the 3070 hit the VRAM wall only two years into its life, despite its shader performance still being competitive. For comparison, the GTX 1080 Ti shipped with 11 GB in 2017, more than the RTX 3070 that "replaced" its performance tier three years later. That extra VRAM is a major reason the 1080 Ti aged so well. The same risk exists today: the RTX 5070 at 12 GB launched into a world where AAA games already recommend 10–12 GB at high settings. Two years from now, 12 GB could be the new 8 GB.
When you see a sudden jump in performance-per-watt, it's almost always a process node shrink. When you see stagnation (Intel 2014–2020), it's being stuck on the same node. This chart tracks which manufacturing process each product category was using, explaining the "why" behind the efficiency curves above.
Apple reached 7nm in 2018 (A12), 5nm in 2020 (A14/M1), and 3nm in 2023 (A17 Pro). NVIDIA reached roughly 7nm-equivalent density in 2022 (RTX 4000 on TSMC 4N), a full four years after Apple. This isn't because NVIDIA engineers are slower — it's because GPUs are physically larger dies, which makes cutting-edge nodes more expensive per chip and lower-yielding. A phone SoC is ~100mm². An RTX 4090 die is ~609mm². Fabbing a large die on the newest node costs dramatically more and wastes more silicon on defects. NVIDIA optimizes for cost-per-transistor, not smallest-possible-node, which is why Ampere (RTX 3000) used Samsung 8nm when Apple was already on TSMC 5nm.
Benchmarks are useful directional tools but inherently incomplete. Here's what you should keep in mind when interpreting this data — and what entire categories of silicon this page deliberately omits.
Apple's performance partly comes from controlling both hardware and software. A benchmark-inferior chip with better OS integration can feel faster in practice. Android OEMs layer skins that add overhead.
Many chips boost to high frequencies for seconds, then thermally throttle. Benchmarks may capture peak performance, but a 30-minute video export tells a different story — especially in thin laptops.
Cinebench R23 loves many cores. Real apps like Photoshop, most games, and web browsers are largely single-threaded or use 4–6 threads at most. A 32-core monster might lose to a fast 8-core in daily use.
CPU and GPU benchmarks isolate compute, but real performance depends on RAM speed, SSD throughput, and system bandwidth. A fast CPU with slow storage will still feel sluggish.
NVIDIA's "RTX 4060" laptop chip shares a name but not specs with the desktop 4060. Intel's "i7" has meant wildly different things across 15 generations. Don't compare by name — compare by benchmark.
Traditional benchmarks don't measure on-device AI performance (Stable Diffusion, LLM inference, live translation). NPUs and tensor cores matter increasingly but aren't captured here. This is a growing blind spot.
iPad Pro uses the same M-series chips as MacBooks — an M4 iPad Pro benchmarks identically to an M4 MacBook Air. Android tablets use phone SoCs. Tablets aren't shown separately because they share silicon with other categories.
PS5 and Xbox Series X use custom AMD APUs (~2020 Zen 2 + RDNA 2). The Switch 2 uses a custom NVIDIA Tegra. Console chips are heavily optimized for their fixed software stack — direct benchmark comparison with PC hardware is misleading. A PS5's "equivalent" PC GPU varies from RTX 2070 to RTX 3060 depending on the game.
When comparing efficiency across form factors (phone vs. laptop vs. desktop), TDP ratings aren't apples-to-apples. A phone SoC's 5W is close to its actual peak draw. A desktop's 125W TDP is its all-core sustained rating — single-core tasks may draw only 30–50W. The raw "pts/W" gap overstates the efficiency difference by 2–3×.
AMD EPYC, Intel Xeon, AWS Graviton, and Apple M Ultra chips live in a different performance universe — 64–128 cores, 300W+ TDPs, and enterprise pricing. Not covered here because the benchmarks and buying criteria are fundamentally different.
Most laptops ship without discrete GPUs. AMD's Radeon 780M/890M, Intel's Arc 140V (Lunar Lake), and Apple's M-series GPU are the actual graphics most people use. Integrated GPU performance has improved dramatically — a 2024 iGPU can match a 2018 discrete GPU — but isn't charted here.