NVIDIA GeForce GTX 660 Ti vs NVIDIA GeForce GTX 670 Comparison

Head-to-Head Benchmarks

The recorded head-to-head data covers two compute workloads: Geekbench OpenCL and Geekbench Vulkan. The NVIDIA GeForce GTX 670 wins both, but the margins are modest. In OpenCL, the GTX 670 scores 15,341 against the GTX 660 Ti's 15,113, a delta of 1.5% in favor of the GTX 670. In Vulkan, the gap widens slightly: the GTX 670 posts 15,553 while the GTX 660 Ti scores 15,012, putting the GTX 670 ahead by 3.5%.

These are not dramatic victories. The GTX 670's advantage in Vulkan is roughly double its OpenCL edge, which suggests the extra memory bandwidth and larger ROP count matter more in the Vulkan workload than in pure compute. The GTX 660 Ti, despite having the same GPU die, the same clock speeds, and the same shading unit count, falls behind in both tests. The average benchmark score across all recorded tests tells a similar story: the GTX 660 Ti averages 15,063, while the GTX 670 averages 12,773. That average is dragged down for the GTX 670 because it also carries a Geekbench Metal score of 7,424, a test the GTX 660 Ti has no entry for. If you isolate the overlapping benchmarks, the GTX 670 is consistently ahead, but the overall average can mislead if you do not account for the Metal result.

Looking at the percentile rankings, the GTX 660 Ti sits at the 57th percentile of all GPUs in the database, while the GTX 670 sits at the 52nd percentile. That is a curious inversion: the GTX 670 wins every shared benchmark, yet its percentile is lower because the Metal score is included in its average. A builder comparing averages alone would see the GTX 660 Ti as the higher-ranked card, but the direct head-to-head data says otherwise.

Architecture Differences

Both cards are built on the same GK104 chip, the same Kepler architecture, and the same 28 nm process at TSMC. Transistor count is identical at 3,540 million, and the die size is exactly 294 mm², giving both a transistor density of 12.0M per mm². Base clock is 915 MHz and boost clock is 980 MHz on both. Memory type is GDDR5 with the same 1502 MHz memory clock, translating to 6 Gbps effective on both cards. Shading units match at 1,344, and TMUs match at 112. The FP32 throughput is identical at 2.634 TFLOPS, and the texture rate is identical at 109.8 GTexel/s. Pixel rate also matches at 27.44 GPixel/s.

The differences live in memory configuration and the backend. The GTX 660 Ti has a 192-bit memory bus with 2 GB of VRAM, yielding 144.2 GB/s of bandwidth. The GTX 670 has a 256-bit bus, also with 2 GB of VRAM, but bandwidth rises to 192.3 GB/s. That is a 33% bandwidth advantage for the GTX 670, which explains its lead in memory-heavy workloads. The ROP count also differs: the GTX 660 Ti has 24 ROPs, while the GTX 670 has 32, a 33% increase. Despite the higher ROP count, the pixel rate is identical because the clock is the same; the extra ROPs give the GTX 670 more headroom in fill-rate-bound scenarios that do not scale purely with pixel throughput.

Power consumption differs as well. The GTX 660 Ti is rated at 150 W TDP, while the GTX 670 is rated at 170 W. Both use dual-slot coolers and require two 6-pin power connectors. The suggested power supply is 450 W for both. Physical dimensions are the same length at 241 mm (9.5 inches), but the GTX 670 adds recorded height (111 mm, 4.4 inches) and width (38 mm, 1.5 inches) to its spec sheet; the GTX 660 Ti does not have those dimensions recorded. Display outputs are identical: 2x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2. Both support PCIe 3.0 x16, DirectX 12 (11_0), OpenGL 4.6, and Vulkan 1.2.175. Neither card has ray tracing or tensor cores.

Where Each One Wins

The GTX 670 wins every shared benchmark in the database. The OpenCL margin is small, 1.5%, which indicates that pure compute throughput is nearly identical between the two cards. That makes sense because the core config is the same: same clocks, same shaders, same TMUs, same FP32 rate. The Vulkan margin is larger at 3.5%, and that is where the GTX 670's 256-bit bus and 32 ROPs start to earn their keep. Vulkan workloads tend to stress memory bandwidth and geometry throughput more than raw ALU count, and the GTX 670 has a clear advantage in both of those areas.

The GTX 660 Ti does not have a single recorded win in the head-to-head tests. However, it does have a higher percentile rank (57 vs. 52) and a higher average benchmark score (15,063 vs. 12,773) because its benchmark set does not include a Metal score. If you are looking at the database average, the GTX 660 Ti looks like the stronger card. If you look at the tests that both cards actually ran, the GTX 670 is the winner. The Metal test is the only reason the GTX 670's average drops, and the GTX 660 Ti has no Metal result to compare.

For a builder, the practical split is this: the GTX 670 is the better choice for any workload that leans on memory bandwidth, such as high-resolution textures, anti-aliasing, or Vulkan-based titles. The GTX 660 Ti is not slower in compute, but it lacks the bandwidth and ROP headroom, so it will trail in scenarios that push data movement rather than shader math. Neither card supports modern ray tracing, so that is not a factor.

FAQ

Q: Which card wins in Geekbench OpenCL?

A: The NVIDIA GeForce GTX 670 wins with a score of 15,341 versus the GTX 660 Ti's 15,113, a margin of 1.5%.

Q: How much faster is the GTX 670 in Vulkan?

A: The GTX 670 scores 15,553 in Geekbench Vulkan, while the GTX 660 Ti scores 15,012. That is a 3.5% lead for the GTX 670.

Q: Do the two cards have the same GPU core?

A: Yes. Both use the GK104 chip with the same transistor count (3,540 million), die size (294 mm²), base clock (915 MHz), boost clock (980 MHz), shading units (1,344), and TMUs (112).

Q: Why does the GTX 660 Ti have a higher average benchmark score?

A: The GTX 660 Ti averages 15,063 across its recorded tests, while the GTX 670 averages 12,773. The GTX 670's average includes a Geekbench Metal score of 7,424, which the GTX 660 Ti does not have, dragging the average down.

Q: What is the memory bandwidth difference?

A: The GTX 660 Ti has a 192-bit bus with 144.2 GB/s of bandwidth. The GTX 670 has a 256-bit bus with 192.3 GB/s of bandwidth, a 33% increase.

Q: Are the power requirements different?

A: Yes. The GTX 660 Ti is rated at 150 W TDP, while the GTX 670 is rated at 170 W. Both use two 6-pin power connectors and recommend a 450 W power supply.

Specification Differences

| Specification | NVIDIA GeForce GTX 660 Ti | NVIDIA GeForce GTX 670 |

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

| Memory Bus Width | 192 bit | 256 bit |

| Memory Bandwidth | 144.2 GB/s | 192.3 GB/s |

| ROPs | 24 | 32 |

| TDP | 150 W | 170 W |

| Dimensions (Height) | Not recorded | 111 mm (4.4 inches) |

| Dimensions (Width) | Not recorded | 38 mm (1.5 inches) |

| Release Date | 2012-08-15 | 2012-05-09 |

| Launch MSRP | 299 USD | 399 USD |

All other core specifications are identical: 28 nm process, GK104 chip, 3,540 million transistors, 294 mm² die, 915 MHz base clock, 980 MHz boost clock, 2 GB GDDR5, 1,344 shading units, 112 TMUs, 2.634 TFLOPS FP32, 109.8 GTexel/s texture rate, 27.44 GPixel/s pixel rate, PCIe 3.0 x16, dual-slot cooler, 2x 6-pin power connectors, 450 W suggested PSU, and the same display outputs and API support.

The Verdict

The data points to the GTX 670 as the faster card in every test both cards share. The OpenCL win is narrow, but the Vulkan win is more decisive, and both results trace back to the same architectural difference: a 256-bit memory bus and 32 ROPs versus a 192-bit bus and 24 ROPs. The GTX 670 also launched earlier, on 2012-05-09, compared to the GTX 660 Ti's 2012-08-15, and carried a higher launch MSRP of 399 USD versus 299 USD. That price gap is the only reason a builder would consider the GTX 660 Ti, and the benchmark data does not justify it on performance grounds alone.

The GTX 660 Ti is not a bad card. Its compute throughput matches the GTX 670 exactly, and its percentile rank in the database is higher because its average is not weighted down by a Metal score. But in a direct comparison, the GTX 670 is ahead. If you are choosing between these two for a build that will run Vulkan titles or memory-intensive workloads, the GTX 670 is the better pick. If you are looking at the database average and want the higher percentile, the GTX 660 Ti will show a better number, but that number is an artifact of an incomplete benchmark set, not a real performance advantage.

For a builder, the practical recommendation is straightforward: the GTX 670 delivers more bandwidth, more ROPs, and wins both overlapping benchmarks. The GTX 660 Ti is the cheaper launch option, but the recorded data gives no scenario where it beats the GTX 670 in a shared test. Choose the GTX 670 unless the 100 USD launch MSRP difference is an absolute constraint, and even then, know that you are giving up measurable performance for that saving.