NVIDIA GeForce GTX 660 vs NVIDIA GeForce GTX 950A Comparison

Head-to-Head Benchmarks

The only directly comparable benchmark in the database is Geekbench OpenCL. In that test, the NVIDIA GeForce GTX 660 scores 11,347 while the NVIDIA GeForce GTX 950A scores 10,273. That gives the GTX 660 a 9.5% lead, which is a meaningful gap in raw compute throughput. The GTX 950A trails by roughly a tenth of a full doubling in performance, so it is not a near tie, but it is also not a generational stomping.

Looking at the broader database context, the GTX 950A sits at the 48th percentile of all GPUs, while the GTX 660 sits at the 45th percentile. Despite losing the head-to-head test, the 950A has a slightly higher overall standing in the database. That is explained by the way average scores are calculated: the GTX 660 has three recorded benchmarks (OpenCL at 11,347, Vulkan at 11,415, and Metal at 4,305), which pull its average down to 9,022. The GTX 950A has only one benchmark, so its average is identical to its OpenCL score of 10,273.

The closest rivals for the GTX 950A are all within a small window. The AMD Radeon RX 6500M averages 10,362, which is 0.9% higher, and the NVIDIA Tesla C2075 averages 10,400, which is 1.2% higher. On the lower side, the AMD Radeon R9 M375 averages 10,070, which is 2% lower, and the AMD Radeon RX 550X averages 10,481, which is 2% higher. The GTX 660's nearest rivals include the NVIDIA TITAN V CEO Edition at 9,037 (0.2% higher), the NVIDIA GeForce GTX 560 at 9,058 (0.4% higher), the AMD Radeon 550X at 8,918 (1.2% lower), and the AMD Radeon Pro WX 5100 at 8,863 (1.8% lower).

The key takeaway is that the GTX 660 wins the single shared benchmark, but the 950A has a better percentile ranking and a higher raw average score. That apparent contradiction is a result of the 660 being dragged down by its weak Metal result. If you focus purely on OpenCL, the older card is faster.

Architecture Differences

The GTX 950A uses the GM107 chip built on the Maxwell architecture. The GTX 660 uses the GK106 chip built on Kepler. Both are manufactured on a 28 nm process at TSMC, so the process node is identical. The transistor counts differ substantially: the GM107 has 1,870 million transistors on a die size of 148 mm², giving a density of 12.6 million transistors per square millimeter. The GK106 has 2,540 million transistors on a 221 mm² die, for a density of 11.5 million per square millimeter. The 660's chip is physically larger and packs more transistors, but the 950A has a slightly higher transistor density, reflecting Maxwell's more efficient layout.

The shading unit counts are very different. The GTX 660 has 960 shading units, 80 texture mapping units, and 24 ROPs. The GTX 950A has 640 shading units, 40 TMUs, and 16 ROPs. That is a 50% advantage for the 660 in shading units and a 100% advantage in both TMUs and ROPs. The 660's pixel rate is 20.64 GPixel/s versus 17.98 GPixel/s for the 950A, and its texture rate is 82.56 GTexel/s versus 44.96 GTexel/s. The 660 also produces 1.981 TFLOPS of FP32 compute versus 1,438.7 GFLOPS for the 950A, a gap of roughly 37%.

The memory subsystems are also distinct. The 950A uses 2 GB of DDR3 on a 128-bit bus, yielding 32.03 GB/s of bandwidth. The 660 uses 2 GB of GDDR5 on a 192-bit bus, yielding 144.2 GB/s. That is a 4.5x bandwidth advantage for the 660, which is enormous and directly explains why the 660 wins the OpenCL test despite being older. Effective memory speed is listed as 2 Gbps for the 950A versus 6 Gbps for the 660.

The feature sets differ in API support. Both support DirectX 12 (11_0) and OpenGL 4.6, but the Vulkan versions are different: the 950A supports Vulkan 1.4, while the 660 supports Vulkan 1.2.175. The 660 also has a Vulkan benchmark recorded in the database, which the 950A lacks.

Power draw and physical format are major separators. The 950A has a TDP of 75 W, uses no power connectors, and is an MXM module with an MXM-B (3.0) interface. The 660 has a TDP of 140 W, requires a single 6-pin power connector, and is a dual-slot PCIe 3.0 x16 card with a length of 241 mm (9.5 inches). The 660 also has a suggested PSU of 300 W. Display outputs differ too: the 950A is listed as portable device dependent, while the 660 has 2x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2.

The Verdict

The data points to a clear split. If you need raw compute and memory bandwidth for OpenCL workloads, the GTX 660 is the better pick. It wins the head-to-head benchmark by 9.5%, has 4.5x the memory bandwidth, and offers 37% more FP32 performance. It also has more shading units, TMUs, and ROPs. For any task that stresses texture throughput or fill rate, the 660 is decisively ahead.

If you are constrained by power, physical space, or the need for a mobile form factor, the GTX 950A is the practical choice. It draws 75 W versus 140 W, needs no auxiliary power connector, and exists as an MXM module. Its Vulkan support is newer (1.4 versus 1.2.175), and its percentile ranking is slightly higher (48th versus 45th). But the 950A does not win any recorded benchmark against the 660, and its memory bandwidth is a severe bottleneck.

The GTX 660 is the stronger desktop card for gaming and general compute, assuming you can accommodate its power draw and physical footprint. The GTX 950A is a low-power mobile part that trades performance for efficiency. For a desktop builder, the 660 is the obvious choice. For a laptop or compact MXM system where power is limited, the 950A is the only sensible option.

FAQ

Q: Which GPU is faster in the Geekbench OpenCL test?

A: The NVIDIA GeForce GTX 660 scores 11,347 versus 10,273 for the GTX 950A, a 9.5% advantage for the 660.

Q: Does the GTX 950A have any benchmark where it beats the GTX 660?

A: No. The only head-to-head benchmark recorded is Geekbench OpenCL, and the GTX 660 wins that one. The 950A has zero wins and the 660 has one win.

Q: Why does the GTX 950A have a higher percentile rank than the GTX 660?

A: The 950A sits at the 48th percentile of all GPUs, while the 660 is at the 45th percentile. This is because the 660's average score of 9,022 is dragged down by its Metal result of 4,305, despite its OpenCL score of 11,347 and Vulkan score of 11,415.

Q: What is the memory bandwidth difference?

A: The GTX 660 has 144.2 GB/s of bandwidth on a 192-bit bus with GDDR5 memory, while the GTX 950A has 32.03 GB/s on a 128-bit bus with DDR3. The 660 offers 4.5x more bandwidth.

Q: Can the GTX 950A be used in a desktop PC?

A: The 950A is an MXM module with an MXM-B (3.0) interface and no power connectors. It is designed for portable devices. The GTX 660 is a dual-slot PCIe 3.0 x16 card that fits a standard desktop slot and requires a 6-pin connector.

Q: Which card has better Vulkan support?

A: The GTX 950A supports Vulkan 1.4, while the GTX 660 supports Vulkan 1.2.175. The 950A has a newer Vulkan version, though the 660 has a recorded Vulkan benchmark and the 950A does not.

Where Each One Wins

The GTX 660 wins in compute-heavy scenarios. Its FP32 output of 1.981 TFLOPS versus 1,438.7 GFLOPS gives it a 37% edge in raw math throughput. Its texture rate of 82.56 GTexel/s versus 44.96 GTexel/s is nearly double, so texture-heavy workloads will favor the 660. Its pixel rate of 20.64 GPixel/s versus 17.98 GPixel/s is also higher, though the gap is smaller. The 660's memory bandwidth advantage is the largest differentiator: 144.2 GB/s versus 32.03 GB/s means any workload that streams large data sets will be dramatically faster on the 660.

The GTX 950A wins in efficiency and portability. It has a 75 W TDP versus 140 W, which means less heat and lower power draw. It requires no power connector, so it can be installed in systems without auxiliary power delivery. Its MXM form factor makes it suitable for laptops and compact modules, where the 660's 241 mm length and dual-slot width would never fit. The 950A also has newer Vulkan support (1.4 versus 1.2.175), which could matter for applications that depend on the latest Vulkan extensions.

For gaming, the 660 is the better bet based on its higher shading unit count (960 versus 640), higher ROP count (24 versus 16), and 4.5x memory bandwidth. For basic display output or low-power compute tasks, the 950A is sufficient, but it will struggle with anything that requires significant memory throughput.

Specification Differences

| Specification | NVIDIA GeForce GTX 950A | NVIDIA GeForce GTX 660 |

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

| Chip | GM107 | GK106 |

| Architecture | Maxwell | Kepler |

| Generation | GeForce 900A | GeForce 600 |

| Process Node | 28 nm | 28 nm |

| Transistors | 1,870 million | 2,540 million |

| Die Size | 148 mm² | 221 mm² |

| Transistor Density | 12.6M / mm² | 11.5M / mm² |

| Base Clock | 993 MHz | 980 MHz |

| Boost Clock | 1124 MHz | 1032 MHz |

| Memory Clock | 1001 MHz, 2 Gbps effective | 1502 MHz, 6 Gbps effective |

| Memory Size | 2 GB | 2 GB |

| Memory Type | DDR3 | GDDR5 |

| Memory Bus Width | 128 bit | 192 bit |

| Memory Bandwidth | 32.03 GB/s | 144.2 GB/s |

| Shading Units | 640 | 960 |

| TMUs | 40 | 80 |

| ROPs | 16 | 24 |

| Pixel Rate | 17.98 GPixel/s | 20.64 GPixel/s |

| Texture Rate | 44.96 GTexel/s | 82.56 GTexel/s |

| FP32 | 1,438.7 GFLOPS | 1.981 TFLOPS |

| TDP | 75 W | 140 W |

| Slot Width | MXM Module | Dual-slot |

| Power Connectors | None | 1x 6-pin |

| Suggested PSU | None listed | 300 W |

| Bus Interface | MXM-B (3.0) | PCIe 3.0 x16 |

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

| DirectX | 12 (11_0) | 12 (11_0) |

| OpenGL | 4.6 | 4.6 |

| Vulkan | 1.4 | 1.2.175 |

| Release Date | 2015-03-12 | 2012-09-05 |

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

| Predecessor | GeForce 800A | GeForce 500 |

| Successor | None listed | GeForce 700 |

| Launch MSRP | None listed | 229 USD |

| Benchmark (OpenCL) | 10,273 | 11,347 |

| Benchmark (Vulkan) | None recorded | 11,415 |

| Benchmark (Metal) | None recorded | 4,305 |

| Average Benchmark Score | 10,273 | 9,022 |

| Percentile vs All GPUs | 48 | 45 |