NVIDIA GeForce GTX 1650 vs NVIDIA GeForce GTX 660 Comparison

FAQ

Q: Which GPU has a higher average benchmark score?

A: The NVIDIA GeForce GTX 660 has a higher average benchmark score of 9022, while the NVIDIA GeForce GTX 1650 averages 7472. Despite this, the GTX 1650 wins both direct head-to-head tests, which suggests the average score is skewed by the mix of tests each card was run through.

Q: How does the GTX 1650 compare to the GTX 660 in OpenCL performance?

A: The GTX 1650 scores 29629 in Geekbench OpenCL versus 11347 for the GTX 660, a delta of -61.7% relative to the 1650. This means the GTX 1650 is roughly 2.6 times faster in that specific workload.

Q: What is the difference in memory capacity between the two cards?

A: The GTX 660 has 2 GB of GDDR5 memory on a 192-bit bus, while the GTX 1650 has 4 GB of GDDR5 on a 128-bit bus. The 660 has higher memory bandwidth at 144.2 GB/s, but the 1650 offers double the capacity.

Q: Which card has a smaller process node?

A: The GTX 1650 uses a 12 nm process, while the GTX 660 uses a 28 nm process. The 1650 also packs more transistors (4,700 million) into a slightly smaller die (200 mm²), giving it a much higher transistor density of 23.5M per mm² versus 11.5M per mm².

Q: Do both cards support DirectX 12?

A: Yes, but with a difference: the GTX 660 supports DirectX 12 (11_0), meaning it is limited to feature level 11_0, while the GTX 1650 supports DirectX 12 (12_1), the full feature level. Both support OpenGL 4.6 and Vulkan, though the 1650 lists Vulkan 1.4 compared to 1.2.175 for the 660.

Q: What is the launch MSRP of each card?

A: The GTX 660 launched at an MSRP of 229 USD, while the GTX 1650 launched at 149 USD.

Architecture Differences

The two cards belong to entirely different architectural generations. The GTX 660 is built on Kepler (chip GK106, GeForce 600 series) and released in the GeForce 600 generation. The GTX 1650 is Turing (chip TU117, GeForce 16 series). The 660 is manufactured on TSMC's 28 nm process with 2,540 million transistors on a 221 mm² die. The 1650 moves to TSMC's 12 nm node with 4,700 million transistors on a 200 mm² die. That is a 85% increase in transistor count while shrinking the die area by 21 mm², leading to a transistor density jump from 11.5M per mm² to 23.5M per mm².

Clock speeds differ notably. The GTX 660 has a base clock of 980 MHz and a boost of 1032 MHz. The GTX 1650 runs at 1485 MHz base and 1665 MHz boost. Even at those higher clocks, the 1650's TDP is 75 W versus 140 W for the 660, a dramatic efficiency improvement. The 660 requires a 6-pin power connector and a 300 W suggested PSU; the 1650 has no additional power connectors and only a 250 W suggested PSU.

The shading units are close: 960 for the 660 versus 896 for the 1650. However, the 660 has 80 texture mapping units and 24 ROPs, while the 1650 has 56 TMUs and 32 ROPs. This shift means the 1650 favors pixel throughput over texture throughput. Pixel rate for the 1650 is 53.28 GPixel/s, more than double the 660's 20.64 GPixel/s. Texture rate is 93.24 GTexel/s versus 82.56 GTexel/s, a smaller margin. FP32 compute is 2.984 TFLOPS for the 1650 versus 1.981 TFLOPS for the 660, a 50% lead for the newer card. The 1650 also lists FP16 performance at 5.967 TFLOPS (2:1 ratio), a capability the 660 does not report.

Memory architecture diverges as well. The 660 has 2 GB GDDR5 on a 192-bit bus with 144.2 GB/s bandwidth. The 1650 has 4 GB GDDR5 on a 128-bit bus but only 128.1 GB/s bandwidth. The 660's wider bus and higher bandwidth are offset by the 1650's doubled capacity. Effective memory speed is 6 Gbps for the 660 versus 8 Gbps for the 1650, which helps the 1650 compensate for its narrower bus.

API support also differs. The 660 supports DirectX 12 (11_0) and Vulkan 1.2.175. The 1650 supports DirectX 12 (12_1) and Vulkan 1.4. Both support OpenGL 4.6. Display outputs differ: the 660 has 2x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2; the 1650 has 1x DVI, 1x HDMI 2.0, and 1x DisplayPort 1.4a.

The Verdict

The data points to a clear generational split. The GTX 1650 wins both direct head-to-head benchmark comparisons by a wide margin: 61.7% in OpenCL and 65.5% in Vulkan. It also offers double the memory capacity (4 GB versus 2 GB), a full DirectX 12_1 feature set, and far lower power requirements (75 W versus 140 W, no power connector versus 1x 6-pin).

The GTX 660, however, holds a higher average benchmark score (9022 versus 7472) and a higher percentile rank (45th versus 40th). This is likely because the two GPUs were measured on different benchmark suites; the 660's recorded tests (Geekbench Metal, OpenCL, Vulkan) are not the same as the 1650's broader set (which includes Passmark DX9 through DX12, G2D, G3D, and compute tests). In the shared tests, the 1650 dominates.

Who should pick which? For any modern workload that uses Vulkan or OpenCL, the GTX 1650 is the stronger choice by a massive margin. Its lower TDP and lack of external power connectors also make it far easier to install in a wide range of systems. The GTX 660 might still appeal to someone strictly targeting older DirectX 11-era applications where its wider memory bus and higher bandwidth could matter, but the 1650's raw compute advantage and modern feature set make it the better pick for nearly all cases in the recorded data.

Specification Differences

| Specification | NVIDIA GeForce GTX 660 | NVIDIA GeForce GTX 1650 |

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

| Architecture | Kepler | Turing |

| Process Node | 28 nm | 12 nm |

| Transistors | 2,540 million | 4,700 million |

| Die Size | 221 mm² | 200 mm² |

| Base Clock | 980 MHz | 1485 MHz |

| Boost Clock | 1032 MHz | 1665 MHz |

| Memory Size | 2 GB | 4 GB |

| Memory Bus | 192 bit | 128 bit |

| Memory Bandwidth | 144.2 GB/s | 128.1 GB/s |

| Shading Units | 960 | 896 |

| TMUs | 80 | 56 |

| ROPs | 24 | 32 |

| Pixel Rate | 20.64 GPixel/s | 53.28 GPixel/s |

| Texture Rate | 82.56 GTexel/s | 93.24 GTexel/s |

| FP32 | 1.981 TFLOPS | 2.984 TFLOPS |

| TDP | 140 W | 75 W |

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

| Suggested PSU | 300 W | 250 W |

| DirectX | 12 (11_0) | 12 (12_1) |

| Vulkan | 1.2.175 | 1.4 |

| Launch MSRP | 229 USD | 149 USD |

Head-to-Head Benchmarks

The two recorded head-to-head tests both show decisive wins for the GTX 1650.

In Geekbench OpenCL, the GTX 1650 scores 29629 against the GTX 660's 11347. The delta is -61.7% from the 1650's perspective, meaning the 660 is only about 38% as fast. This is a massive gap that reflects the 1650's higher clock speeds, more efficient architecture, and doubled memory capacity, even though the 660 has a wider memory bus and higher bandwidth.

In Geekbench Vulkan, the gap is even larger. The GTX 1650 scores 33042, while the GTX 660 scores 11415, a delta of -65.5%. The 1650 is roughly 2.9 times faster in Vulkan. This suggests the Turing architecture has a particularly strong advantage in modern low-level graphics APIs, and the 660's Kepler design is significantly less efficient at handling Vulkan's driver overhead and parallel workload dispatch.

It is importantly the 660 also has a Geekbench Metal score of 4305, but the 1650 has no Metal score recorded, so no direct comparison is possible there. Similarly, the 1650 has Passmark DX9 (124), DX10 (39), DX11 (58), DX12 (35), G2D (561), G3D (7880), and GPU compute (3048) scores, but the 660 has none of those. The only overlapping benchmarks are the two Geekbench tests, and the 1650 wins both decisively.

Where Each One Wins

The GTX 1650 wins in every directly comparable workload: OpenCL and Vulkan. It also wins on memory capacity (4 GB versus 2 GB), which matters for modern games and compute tasks that exceed 2 GB. Its pixel rate is more than double (53.28 versus 20.64 GPixel/s), suggesting a strong advantage in fill-rate-bound scenarios such as high-resolution rendering or heavy post-processing effects. Its FP32 throughput is 50% higher (2.984 versus 1.981 TFLOPS), which benefits general compute and shader-heavy workloads. The 1650 also has a large efficiency win: 75 W TDP with no additional power connector, versus 140 W with a 6-pin connector. This makes it suitable for compact or older systems with limited PSU headroom.

The GTX 660 retains a few specific advantages. Its memory bandwidth is higher at 144.2 GB/s versus 128.1 GB/s, which could benefit certain bandwidth-sensitive workloads that fit within its 2 GB frame buffer. It has more texture mapping units (80 versus 56), giving it a higher texture rate relative to its pixel rate, though the 1650 still beats it in absolute texture rate (93.24 versus 82.56 GTexel/s). The 660 also has a wider memory bus (192-bit versus 128-bit), which can reduce latency in some access patterns. In terms of display outputs, the 660 offers dual DVI ports, which may be useful for legacy monitors or multi-display setups with DVI-only panels.

In summary, the GTX 1650 is the superior card for modern graphics APIs, compute workloads, and efficiency. The GTX 660's only remaining edge is in raw memory bandwidth and legacy connectivity, but the 1650's dominance in the recorded benchmarks makes it the clear choice for anyone comparing the two today.