Thief

Thief, a stealth game released on 2014-02-24, pairs surprisingly well with a 2022-era desktop CPU and a 2021 GPU. The combination of an Intel Core i5-12400 with an AMD Radeon RX 6700 XT lands at rank 1897 out of 3720 tested combos for this game, putting it almost exactly at the midpoint of the measured field. Importantly, the dataIsMeasured flag is true, so the frame rates below are actual recorded results, not estimates. The CPU sits at the 72nd percentile among all CPUs, and the GPU likewise sits at the 72nd percentile among all GPUs — a balanced pairing of two parts that are neither low-end nor top-tier. The game's age (released before either component existed) means that raw fill rate and single-frame latency matter more than modern geometry features, but the numbers tell a layered story about where the bottleneck lies at each resolution.

How This Combo Ranks

The rank of 1897 against a field of 3720 combos means that 1896 other combinations achieved higher measured average FPS in Thief, while 1823 combos posted lower scores. In other words, this CPU+GPU pair lands in the shallow half of the field but only by a narrow margin — it crosses the exact middle point of the ranking table, neither dramatically over-performing nor under-performing. The CPU's average benchmark score of 18,683 places it at the 72nd percentile of all CPUs, with its nearest rivals being the AMD EPYC 7643 (average score 18,697, delta -0.1%), the Intel Core i7-1355U (18,730, delta -0.3%), the AMD Ryzen AI 5 330 (18,811, delta -0.7%), and the Intel Core i3-14100F (18,519, delta -0.9%). With the exception of the Ryzen AI 5 330, the i5-12400 is actually slightly stronger than these rivals in raw synthetic score, yet in Thief it contributes to a combo that lands at the middle rank. The GPU's average benchmark score is 27,425 at the 72nd percentile of all GPUs, and its nearest rivals are the NVIDIA GeForce RTX 4070 Mobile (average score 27,435, delta 0), the NVIDIA GeForce RTX 3090 (27,565, delta -0.5), the NVIDIA RTX PRO 4000 Blackwell (27,135, delta +1.1), and the AMD Radeon Pro Vega 20 (27,839, delta -1.5). So synthetic benchmarks place the GPU within 1.5% of four very different products, but the real-game ranking reveals that Thief is not a clean GPU- or CPU-bound title. The ranking is what matters: in the game itself, this visibly strong synthetic pairing only reaches the cutoff point that separates the upper 50% of combos from the lower 50%, telling the analyst that Thief is typical of older stealth titles that reward single-threaded throughput and raw pixel fill — neither the CPU nor the GPU dominates synthetically, they arrive at the 72nd and 72nd percentile respectively, and their actual in-game position is the middle of the pack.

GPU Role

The AMD Radeon RX 6700 XT is built on the Navi 22 chip, RDNA 2.0 architecture, 7 nm process, with has 17,200 million transistors on a 335 mm² die at a transistor density of 51.3M/mm². This is a dual-slot card that pulls up to 230 W, requires a 50/60 recommendation of 550 W PSU, with 1x 6-pin plus 1x 8-pin connectors, and is PCIe 4.0 x16. It carries 12 GB of GDDR6 memory on a 192-bit bus at 2000 MHz (16 Gbps effective), delivering 384.0 GB/s of bandwidth. The chip's clock profile lists a base of 2321 MHz, a boost of 2581 MHz, and a game clock of 2424 MHz — those are all enhancing clocks that the chip actively maintains under load. The fill rate numbers speak to Thief's rasterization needs: the card does 165.2 GPixel/s and 413.0 GTexel/s, with FP32 at 13.21 and FP16 at 26.43 TFLOPS, backed by 2560 shading units, 160 TMUs, 64 ROPs, and 40 ray tracing cores. For a 2014 game, the RT cores are irrelevant, but the massive fill rate is exactly what gives the card its power at 4K.

Evidence from the measured FPS shows the GPU constrains the system at higher resolutions. At 3840x2160, the High preset averages 43 FPS; the Ultra preset averages 41 FPS; the Medium preset averages 57 FPS with min 48 and max 66; the Low preset reaches 75. This is a 34-frame spread between the lowest and highest quality at 4K. At 2560x1440, High averages 74 FPS, Ultra drops to 70, Medium is the best balance at 97 FPS (min 82, max 112), but Low tops out at 128 FPS. At 1920x1080, High reaches a 108 FPS average, Ultra 103, Medium 141 (min 120, max 163), and Low goes to 170 FPS. The pattern — Full HD under the lockstep of setting quality by GPU load — is that the RX 6700 XT is strong enough to ensure the CPU does the heavy lifting at 1080p, but its ROPs and texture bandwidth get saturated rapidly above that. The 165.2 GPixel/s pixel rate and 413.0 GTexel/s fill rate become the limiting factors when restoring resolution demands, which is reflected in the 4K metrics. Considering that the GPU's nearest rival in synthetic score is the RTX 3090 (an average score of 27,565, just 1.7% above this card), the RX 6700 XT can put out a very high frame for a 2014 game when the resolution stays manageable, but at 4K it creates a bottleneck that the CPU would otherwise handle well. The 12 GB VRAM buffer is generous for a 2014 title, so capacity is never the problem — what governs 4K performance is bandwidth (384.0 GB/s) and the memory bus (192-bit), which correspond exactly to the measured 43-to-41 FPS ceiling at 4K High/Ultra.

FAQ

Q: Is this combo more CPU-bound or GPU-bound in Thief?

A: The data shows that at 1080p the i5-12400 at its 2.50 GHz base and 4.40 GHz boost can feed most scenes well — the 1920x1080 Low preset achieves 170 FPS and Medium reaches 141 FPS (min 120) — while at 4K the RX 6700 XT becomes the strongest rocket, with the same Low preset collapsing to 75 FPS and Ultra to 41 FPS. The GPU's fill rate and VRAM bus (192-bit, 384.0 GB/s) cap the resolution, but 1080p and 1440p results show a CPU that can push 4-digit frames on low presets only when the rasterizer isn't the ceiling.

Q: What FPS can I expect at 1440p with High settings?

A: The measured row at 2560x1440 High is 74 FPS average. That's a single row with no min/max data, but it falls between Medium (97 FPS) and Ultra (70 FPS), so the High preset is basically the most efficient 1440p setting the data shows.

Q: Does the 12 GB VRAM have any effect in Thief?

A: Only 12 GB of GDDR6 memory with a 192-volt CDP. The 384.0 GB/s bandwidth is more than enough for a 2014 release; the data shows a 4K Medium preset at 57 FPS inside the 48-66 min-max range, and no memory-related stalls are evident in the frame pacing. In practice, VRAM is the least relevant of the GPU specs in this game.

Q: How does this combo fair against the midpoint of all combos?

A: Rank 1897 out of 3720 puts the system just below the median — 0.5% about the exactly middle. The CPU's own percentile (72nd) and the GPU's percentile (72nd) suggest the pair is more capable than half of the tested parts, but in Thief the game's rasterization workload brings it down to the middle rank, indicating that the engine relies more on raw fill than the deep cache/expert-of-top-tier hardware provides.

Q: Is the 120-163 frame range on Medium at 1080p reliable?

A: The measured 141 FPS average with 120 to 163 FPS delivers a very stable window for 1080p. This system's frame delivery at 1920 Medium (min 120) is the best combination of min and average in the whole dataset — only the Low preset exceeds it (170 FPS), but that requires dropping to the lowest quality level.

Measured FPS Breakdown

At 1920x1080, four rows exist. Low averages 170 FPS — the maximum for this combo in any resolution/settings. Medium averages 141 FPS, with min 120 and max 163. High averages 108 FPS. Ultra averages 103 FPS. Everything is measured and GPUs are at 1080p, that is the resolution where the i5's boost to 4.40 GHz carries the weight: even at Low the card has headroom, and the spread from Low to Ultra is 67 FPS — only 30% of the harvested min/max.

At 2560x1440, Medium again delivers tight pacing: average 97 FPS with a min 82 and max 112 window. High drops to 74 FPS. Ultra averages 70 FPS. Low averages better than Ultra at 1080p (128 FPS) but still only reaches about 75% of the 1080p Low number, showing that 1440p costs roughly a quarter of raw throughput.

At 3840x2160, the data visibly flattens. Low averages 75 FPS; Medium averages 57 FPS (min 48 max 66); High drops to 43 FPS; Ultra falls to 41 FPS. The low-end offset at 4K is that even Low i directly hits just 75 FPS, which is below the 1440p High number (74 FPS) — so the GPU's fill rate (165.2 GPixel/s) is the sizing factor there, not the CPU frequencies. It is worth noting the medium 4K row has the only min/max values in the entire 4K set (48/66) — the row suggests a consistent and stable frame delivery.

Across all modes, the measured numbers present a triangular shape: the resolution drops producer drives are smooth and predictable (e.g., 170 / 128 / 75 across Low at 1080/1440/2160), while the settings jumps at a fixed resolution tune the quality load without pushing the CPU — the average difference between Low and Ultra at the same resolution is 40-45% of the number of themes.

Resolution Scaling

The 1920x1080 row to 2560x1440 (1.6p) shows the following: Low degrades from 170 to 128 FPS, Medium drops from 141 to 97, High from 108 to 74, and Ultra from 103 to 70. The delta is between 30 to 41 FPS in absolute terms, which is a 22 to 38 percent frame drop — a healthy, predictable fall given the pixel density jump of 2.25x. Such a mild drop says the system scale be calibrated: the bus speed versus 486 GB/s bandwidth (the GPU 384.0 GB/s) then becomes the only factor in play at 1440p.

The move from 1440p to 2160p (re4K) is far steeper: Low drops from 128 to 75 FPS, Medium from 97 to 57, High from 74 to 43, and Ultra from 70 to 41. That is a 42% decline on the low, 41% on Medium, 42% on High, and 41% on Ultra = all very close to a 40% reduction. This asymptotic 40% each step from 2K to 4K (regardless of setting) is a classic signature of pixel-limited rasterization. The GPU during is the constraint: its peak rate is 165.2 GPixel/s, and the 4K aggress hera — 8, 294, 400 pixels minus the fill ultimately equal — that is the exact factor of such a skill pass the frame. The CPU's boost clock (4.4 GHz) will not be engaged during any of theses frames. So the data points