NVIDIA GeForce GTX 670

The NVIDIA GeForce GTX 670, built on the 28 nm Kepler architecture with the GK104 chip, occupies a distinct middle ground in the benchmark hierarchy. Its average benchmark score of 12,014 places it at the 50th percentile of all GPUs, meaning it performs better than half of the database's tracked graphics cards. The card integrates 1,344 shading units, 112 texture mapping units, and 32 ROPs, delivering 2.634 TFLOPS of FP32 compute and a texture rate of 109.8 GTexel/s. With a boost clock of 980 MHz and 2 GB of GDDR5 memory on a 256-bit bus, the GTX 670 achieves a memory bandwidth of 192.3 GB/s. These specifications define a card that was competitive at its 2012 launch and remains relevant in specific low-resolution scenarios, though its age is evident in modern API support and ray tracing capabilities.

How It Compares

Against the AMD Radeon Pro 5500M, the GTX 670 holds a negligible 0.2% advantage in average score (12,014 vs. 11,986). This near-parity is notable given the Radeon Pro 5500M is a mobile workstation GPU from a much later generation. The data suggests that in raw compute throughput, the older Kepler architecture still trades blows with a modern mobile part, likely due to the GTX 670's higher shader count and dedicated desktop power envelope. However, the Radeon Pro 5500M likely wins on efficiency and feature support, which are not captured in the raw score.

The closest rival is the NVIDIA GeForce GTX 960, which scores 12,045, a marginal 0.3% higher than the GTX 670. This 31-point difference is within statistical noise, indicating that the GTX 670 and GTX 960 deliver effectively identical average performance. This is a remarkable result for the older card, as the GTX 960 is a Maxwell-generation part with newer architecture optimizations. The GTX 670 compensates for its older design with a wider 256-bit memory bus and higher shader count (1,344 vs. the GTX 960's lower count), though the GTX 960 benefits from higher clock speeds and better compression.

The AMD Radeon RX 6500 XT trails the GTX 670 by 1.0% in average score (11,897 vs. 12,014). This is a surprising outcome because the RX 6500 XT is a modern RDNA2 card with significantly faster memory technology on paper. Yet the benchmark data shows the GTX 670's 192.3 GB/s bandwidth and 2 GB buffer are sufficient to edge out the RX 6500 XT's lower bus width and smaller L3 cache in these specific tests. The advantage is slim, but it demonstrates that the GTX 670's balanced design—32 ROPs and 112 TMUs—still holds up in synthetic Geekbench workloads.

The NVIDIA GeForce GTX 960A, a mobile variant of the GTX 960, scores 12,002, which is 0.1% lower than the GTX 670. This delta is essentially a tie, reinforcing the conclusion that the GTX 670's desktop-class specifications—particularly its 2 GB GDDR5 on a 256-bit bus—allow it to match or slightly exceed mobile derivatives of newer architectures. In practical terms, the GTX 670 and GTX 960A are interchangeable in performance, though the GTX 960A would consume less power in a laptop chassis.

Power and Cooling

The GTX 670 carries a thermal design power (TDP) of 170 W, which is modest by modern standards but substantial for a dual-slot card from 2012. The suggested power supply is 450 W, a figure that reflects the card's peak draw alongside a typical system's other components. The card requires two 6-pin power connectors, which is a standard configuration for mid-range GPUs of its era; users with modern power supplies will need to ensure these connectors are available, as the card does not use the newer 8-pin or 12VHPWR standards.

Cooling is handled by a dual-slot design with a length of 241 mm (9.5 inches), a height of 111 mm (4.4 inches), and a width of 38 mm (1.5 inches). This footprint means the card will fit in most mid-tower cases, but users should verify clearance against drive cages or cable management channels. The 28 nm process node from TSMC, with 3,540 million transistors on a 294 mm² die, produces a transistor density of 12.0 million per square millimeter. This density is low by today's standards, which contributes to the 170 W TDP being manageable with a single fan in most aftermarket designs, though reference models may run louder under sustained load. The PCIe 3.0 x16 bus interface is fully backward compatible, but the card will run at reduced bandwidth if installed in an older PCIe 2.0 slot, which may slightly impact performance in bandwidth-sensitive tests.

Who Should Consider It

Benchmark results indicate the GTX 670 is best suited for 1080p gaming at medium to low settings in modern titles, or high settings in older games from its 2012 release era. Its Geekbench scores—5,234 in Metal, 15,236 in OpenCL, and 15,572 in Vulkan—show a strong bias toward compute-heavy workloads over graphics-API-specific tests. The Vulkan score is 30% higher than the Metal score, suggesting the card handles driver-optimized APIs better than Apple's Metal abstraction layer. For gamers, this means the card will perform relatively better in DirectX 12 (11_0) and Vulkan titles that leverage its 1,344 shading units efficiently.

At 1440p, the 2 GB memory buffer becomes a limiting factor, as many modern games exceed 2 GB of VRAM usage at that resolution, causing texture thrashing and stutter. The 192.3 GB/s bandwidth helps mitigate this, but the card's 27.44 GPixel/s pixel fill rate and 109.8 GTexel/s texture rate are insufficient for high-detail 1440p rendering. Users with a 1440p monitor should restrict the GTX 670 to esports titles or older games with reduced texture quality. Conversely, at 1080p with low presets, the card's 50th percentile standing means it will comfortably handle games that are not VRAM-hungry, delivering playable framerates in most competitive shooters and indie titles. The card is also a candidate for HTPC builds, given its dual-slot profile and support for 2x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2 outputs, though its 170 W TDP requires adequate case airflow.

FAQ

Q: How does the GTX 670 compare to the GTX 960 in raw performance?

A: The GTX 670 scores 12,014 on average, while the GTX 960 scores 12,045, a difference of only 0.3%. This places them in a statistical tie, meaning the older card matches the newer Maxwell part in Geekbench workloads.

Q: What is the card's memory configuration and bandwidth?

A: The GTX 670 uses 2 GB of GDDR5 memory on a 256-bit bus, yielding a bandwidth of 192.3 GB/s. The memory clock is 1,502 MHz, which translates to 6 Gbps effective.

Q: Does the GTX 670 support modern APIs like Vulkan and DirectX 12?

A: Yes, it supports DirectX 12 (feature level 11_0), OpenGL 4.6, and Vulkan 1.2.175. However, the DirectX 12 support is at the 11_0 feature level, which means it lacks some newer features like mesh shaders.

Q: What is the required power supply wattage for this card?

A: The suggested power supply is 450 W, and the card requires two 6-pin power connectors. The TDP of the card is 170 W.

Q: What is the card's physical size?

A: It is a dual-slot card measuring 241 mm (9.5 inches) in length, 111 mm (4.4 inches) in height, and 38 mm (1.5 inches) in width.

Q: Is the GTX 670 still a viable option for gaming?

A: At its 50th percentile ranking, it is viable for 1080p gaming at low-to-medium settings in modern titles, but the 2 GB VRAM and lack of ray tracing hardware make it unsuitable for high-detail 1440p or 4K gaming.

Ray Tracing and Feature Set

The GTX 670 has no dedicated ray tracing cores or tensor cores, as these features were introduced in later NVIDIA architectures. The card's Kepler design, released in 2012, predates the RTX series by six years, and as such, it has no hardware acceleration for real-time ray tracing or DLSS-style AI upscaling. Benchmark data confirms this absence, as the card's Geekbench scores are limited to compute (OpenCL), graphics (Vulkan), and Apple's Metal API—none of which stress RT-specific hardware.

In terms of API support, the card lists DirectX 12 (11_0), OpenGL 4.6, and Vulkan 1.2.175. The DirectX 12 feature level of 11_0 is the minimum tier for DX12, meaning the card can run DX12 titles but cannot use features like conservative rasterization or rasterizer-ordered views. The Vulkan 1.2.175 support is more robust, allowing the card to access modern Vulkan extensions, though performance will still be bounded by the 2.634 TFLOPS FP32 throughput. The card's 32 ROPs deliver a pixel rate of 27.44 GPixel/s, which is the primary constraint for fill-rate-bound workloads at high resolutions. For users seeking ray-traced effects, the GTX 670 is not a viable option, and the data shows it should be paired with older, rasterization-based games or used for compute tasks via OpenCL, where its 15,236 score indicates reasonable parallel processing capability.