AMD Radeon 660M vs NVIDIA GeForce GTX 670 Comparison

NVIDIA GeForce GTX 670 and AMD Radeon 660M are separated by a decade of GPU architecture, yet they land surprisingly close in aggregate benchmark scores. The GTX 670, a 2012 desktop discrete card, posts an average benchmark score of 12,773, while the 660M, a 2022 integrated graphics processor, averages 13,812. That places the 660M about 8.1% higher overall, and it also holds a better percentile ranking (55th vs. 52nd). However, the head-to-head results tell a different story, with the GTX 670 winning both shared tests. This is a clash between raw legacy silicon and modern efficiency, and the right choice depends entirely on the workload and platform constraints.

Where Each One Wins

The GTX 670 wins in raw compute throughput in the two benchmark suites where both GPUs have results. In Geekbench OpenCL, the GTX 670 scores 15,341 against the 660M's 12,876, a 19.1% advantage. In Geekbench Vulkan, the GTX 670 scores 15,553 versus 14,748, a 5.5% lead. That makes the GTX 670 the pick for any application that leans on OpenCL compute or Vulkan rendering, where its dedicated memory bandwidth and higher shading unit count pay off. The 660M, by contrast, wins on aggregate average score because it also has a Geekbench Metal result, but that test is not shared with the GTX 670, so it cannot be compared directly. The 660M's higher average is driven by its 12,876 OpenCL and 14,748 Vulkan scores plus a Metal score, while the GTX 670's average is pulled down by its lower OpenCL and Vulkan numbers relative to its own peak. For gaming or compute tasks that use these APIs, the GTX 670 is the stronger performer.

The 660M wins in platform flexibility and power efficiency. It is an integrated GPU with a 40 W TDP, versus the GTX 670's 170 W TDP. The 660M requires no power connectors and fits in an IGP slot width, while the GTX 670 needs a dual-slot cooler and two 6-pin power connectors. The 660M also supports PCIe 4.0 x8, whereas the GTX 670 uses PCIe 3.0 x16. For a thin-and-light laptop or a compact system without a discrete GPU slot, the 660M is the only viable option. Its system-shared memory means bandwidth is system-dependent, but that also removes the 2 GB VRAM ceiling that the GTX 670 has. The 660M's 6 nm process node and 13,100 million transistors on a 208 mm² die give it a transistor density of 63.0M per mm², dwarfing the GTX 670's 12.0M per mm² on 28 nm. That density advantage translates to modern features like ray tracing cores, which the GTX 670 lacks entirely.

Architecture Differences

The GTX 670 uses the GK104 chip on NVIDIA's Kepler architecture, built on a 28 nm process at TSMC. It packs 3,540 million transistors into a 294 mm² die, with a transistor density of 12.0M per mm². The 660M uses the Rembrandt chip on AMD's RDNA 2.0 architecture, built on a 6 nm process at TSMC. It has 13,100 million transistors in a 208 mm² die, yielding a density of 63.0M per mm². That is a massive density advantage, reflecting a decade of process node shrinks.

Core counts differ sharply. The GTX 670 has 1,344 shading units, 112 texture mapping units, and 32 ROPs. The 660M has just 384 shading units, 24 TMUs, and 16 ROPs. Despite having fewer cores, the 660M hits higher clock speeds: a base of 1,500 MHz and boost of 1,900 MHz, versus the GTX 670's 915 MHz base and 980 MHz boost. The 660M also has 6 ray tracing cores, while the GTX 670 has none. The 660M's FP32 performance is 1,459.2 GFLOPS and its FP16 is 2.918 TFLOPS (2:1), while the GTX 670's FP32 is 2.634 TFLOPS with no FP16 data listed. The 660M's pixel rate is higher at 30.40 GPixel/s versus 27.44 GPixel/s, but its texture rate is far lower at 45.60 GTexel/s against 109.8 GTexel/s.

Memory architecture is a fundamental split. The GTX 670 has 2 GB of GDDR5 on a 256-bit bus, delivering 192.3 GB/s of dedicated bandwidth. The 660M uses system-shared memory, with type, bus width, and bandwidth all listed as "System Shared" or "System Dependent." That means the 660M's effective memory performance varies with the host system's RAM, while the GTX 670 has fixed, predictable bandwidth. The GTX 670's memory clock is listed at 1,502 MHz with 6 Gbps effective; the 660M's memory clock is "System Shared." The GTX 670 supports DirectX 12 (11_0), OpenGL 4.6, and Vulkan 1.2.175. The 660M supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4, making it the more modern API set, particularly with DirectX 12 Ultimate features.

Head-to-Head Benchmarks

In Geekbench OpenCL, the GTX 670 scores 15,341 against the 660M's 12,876. That is a 19.1% delta, the largest gap in any shared test. The GTX 670's advantage here comes from its 1,344 shading units and dedicated 192.3 GB/s memory bandwidth, which feed compute workloads far better than the 660M's 384 shading units and system-shared memory. For users running OpenCL-accelerated tasks like video encoding or physics simulation, the GTX 670 is clearly ahead.

In Geekbench Vulkan, the GTX 670 scores 15,553 versus 14,748, a 5.5% delta. This is a closer margin, suggesting that the 660M's higher clocks and modern architecture narrow the gap when the API is more efficient. The 660M's 1,900 MHz boost clock and 30.40 GPixel/s pixel rate help it stay competitive, but the GTX 670's 109.8 GTexel/s texture rate and 2.634 TFLOPS FP32 still give it the edge. Notably, the 660M does not have a Geekbench Metal result, while the GTX 670 scores 7,424 in that test. That Metal score is not part of the head-to-head comparison, but it contributes to the 660M's higher average benchmark score, since the 660M's average of 13,812 benefits from having three scores (including Metal) while the GTX 670's average is based on its three scores as well. The 660M's average is 8.1% higher, but in the two tests where they meet directly, the GTX 670 wins both.

FAQ

Q: Which GPU has the higher average benchmark score?

A: The AMD Radeon 660M has a higher average benchmark score of 13,812, compared to the NVIDIA GeForce GTX 670's 12,773. That is an 8.1% difference in the 660M's favor.

Q: Does the GTX 670 beat the 660M in any shared benchmark?

A: Yes. The GTX 670 wins both head-to-head tests. It scores 15,341 in Geekbench OpenCL (19.1% ahead of the 660M's 12,876) and 15,553 in Geekbench Vulkan (5.5% ahead of the 660M's 14,748).

Q: Why does the 660M have a higher percentile ranking if it loses head-to-head?

A: The 660M ranks in the 55th percentile of all GPUs, while the GTX 670 ranks in the 52nd percentile. The 660M's higher average score of 13,812 versus 12,773, plus its additional Metal benchmark result, pushes it up the overall rankings despite losing the two shared tests.

Q: What is the most significant architectural difference between them?

A: The GTX 670 uses a 28 nm Kepler chip with 1,344 shading units and 2 GB of dedicated GDDR5 memory on a 256-bit bus. The 660M uses a 6 nm RDNA 2.0 chip with 384 shading units, 6 ray tracing cores, and system-shared memory. The 660M also has a much higher transistor density (63.0M per mm² vs. 12.0M per mm²).

Q: Which GPU requires more power and cooling?

A: The GTX 670 has a 170 W TDP, requires a dual-slot cooler, and needs two 6-pin power connectors. The 660M has a 40 W TDP, is an IGP with no power connectors, and has no separate slot width requirement.

Q: Does the GTX 670 have a launch MSRP?

A: Yes, the NVIDIA GeForce GTX 670 had a launch MSRP of 399 USD. The AMD Radeon 660M has no listed launch MSRP.

The Verdict

Choose the NVIDIA GeForce GTX 670 if you need maximum compute performance in OpenCL or Vulkan and have a desktop chassis that can accommodate a 170 W, dual-slot card with two 6-pin connectors. The data shows it leads the 660M by 19.1% in OpenCL and 5.5% in Vulkan. Its dedicated 192.3 GB/s memory bandwidth and 1,344 shading units make it the stronger raw performer for any workload that isn't limited by its 2 GB VRAM. It also has a fixed memory subsystem, so performance is not dependent on system RAM quality.

Choose the AMD Radeon 660M if you prioritize efficiency, modern API support, and platform integration. It consumes 40 W versus 170 W, needs no power connectors, and fits as an IGP. Its 6 nm RDNA 2.0 architecture brings ray tracing cores and DirectX 12 Ultimate support, which the GTX 670 lacks entirely. Its higher base and boost clocks (1,500 MHz and 1,900 MHz vs. 915 MHz and 980 MHz) and higher pixel rate (30.40 GPixel/s vs. 27.44 GPixel/s) make it competitive in lighter workloads, even if it loses the shared benchmarks. Its system-shared memory is a double-edged sword: it removes VRAM limits but makes bandwidth dependent on the host system. For a laptop or compact build where a discrete GPU is impractical, the 660M is the only sensible pick. For a desktop where power and space are available, the GTX 670 delivers more compute per benchmark point. The 660M's higher average score and percentile ranking suggest it is the better all-rounder in the broader GPU landscape, but the GTX 670's head-to-head wins show it still has teeth in specific APIs.

Specification Differences

| Field | NVIDIA GeForce GTX 670 | AMD Radeon 660M |

|-------|------------------------|------------------|

| Architecture | Kepler | RDNA 2.0 |

| Process Node | 28 nm | 6 nm |

| Transistors | 3,540 million | 13,100 million |

| Die Size | 294 mm² | 208 mm² |

| Transistor Density | 12.0M / mm² | 63.0M / mm² |

| Base Clock | 915 MHz | 1500 MHz |

| Boost Clock | 980 MHz | 1900 MHz |

| Memory Size | 2 GB | System Shared |

| Memory Type | GDDR5 | System Shared |

| Memory Bus Width | 256 bit | System Shared |

| Memory Bandwidth | 192.3 GB/s | System Dependent |

| Shading Units | 1344 | 384 |

| TMUs | 112 | 24 |

| ROPs | 32 | 16 |

| Ray Tracing Cores | None | 6 |

| Pixel Rate | 27.44 GPixel/s | 30.40 GPixel/s |

| Texture Rate | 109.8 GTexel/s | 45.60 GTexel/s |

| FP32 Performance | 2.634 TFLOPS | 1,459.2 GFLOPS |

| FP16 Performance | Not listed | 2.918 TFLOPS (2:1) |

| TDP | 170 W | 40 W |

| Slot Width | Dual-slot | IGP |

| Power Connectors | 2x 6-pin | None |

| Suggested PSU | 450 W | Not listed |

| Bus Interface | PCIe 3.0 x16 | PCIe 4.0 x8 |

| DirectX Support | 12 (11_0) | 12 Ultimate (12_2) |

| Vulkan Support | 1.2.175 | 1.4 |

| Display Outputs | 2x DVI, 1x HDMI 1.4a, 1x DisplayPort 1.2 | Portable Device Dependent |

| Release Date | 2012-05-09 | 2022-01-03 |

| Production Status | End-of-life | End-of-life |

| Launch MSRP | 399 USD | Not listed |