NVIDIA GeForce GTX 670 vs NVIDIA Tesla K20c Comparison

NVIDIA’s Kepler generation produced two very different interpretations of the same architecture. The Tesla K20c is a compute-oriented accelerator with no display outputs, while the GeForce GTX 670 is a conventional graphics card built for gaming and general use. Benchmark data shows a clear split: the GTX 670 wins the only shared test decisively, but the K20c counters with a far larger memory pool and higher raw compute throughput. The choice between them depends entirely on whether the workload is graphics-bound or compute-bound.

Where Each One Wins

The GeForce GTX 670 wins the head-to-head benchmark competition outright. In the only test both cards share — Geekbench OpenCL — the GTX 670 scores 15,341 against the K20c’s 11,479, a delta of -25.2% from the K20c’s perspective. This is not a marginal victory; the GTX 670 outperforms the Tesla by roughly a quarter in this metric. The GTX 670 also holds a higher percentile rank among all GPUs at 52, versus 51 for the K20c, and a higher average benchmark score of 12,773 versus 11,479.

The Tesla K20c’s wins are not visible in the head-to-head numbers, but they appear in its specifications. It carries 5 GB of GDDR5 memory, more than double the GTX 670’s 2 GB, and its 320-bit memory bus delivers 208.0 GB/s of bandwidth versus 192.3 GB/s for the GTX 670. Its shading unit count of 2,496 is nearly double the GTX 670’s 1,344, and its FP32 throughput of 3.524 TFLOPS exceeds the GTX 670’s 2.634 TFLOPS by a substantial margin. The K20c also has more texture mapping units (208 vs 112) and more ROPs (40 vs 32), translating to higher pixel and texture rates.

The practical split is clear: the GTX 670 wins on OpenCL execution speed and overall benchmark standing, while the K20c wins on memory capacity, memory bandwidth, shading resources, and theoretical compute throughput. For workloads that saturate FP32 compute or require large working sets in memory, the K20c holds the advantage. For anything that relies on the OpenCL path measured by Geekbench, the GTX 670 is faster.

Architecture Differences

Both cards are built on the 28 nm process at TSMC, but they use different chips. The Tesla K20c uses the GK110 chip, while the GeForce GTX 670 uses the GK104. This is the core architectural divide: GK110 is NVIDIA’s flagship compute silicon, built with 7,080 million transistors on a 561 mm² die, yielding a transistor density of 12.6 million per square millimeter. The GK104 is a smaller, graphics-focused chip with 3,540 million transistors on a 294 mm² die, at a density of 12.0 million per square millimeter.

The transistor count difference is enormous — the K20c packs exactly double the transistors of the GTX 670. That extra silicon translates directly into more execution resources: 2,496 shading units versus 1,344, 208 TMUs versus 112, and 40 ROPs versus 32. The K20c’s larger chip also supports a wider 320-bit memory bus versus 256-bit, which is why it achieves higher bandwidth despite lower memory clock speeds.

Memory clock differences are notable. The K20c runs its GDDR5 at 1300 MHz (5.2 Gbps effective), while the GTX 670 runs at 1502 MHz (6 Gbps effective). The GTX 670’s faster memory clock helps it narrow the bandwidth gap, but the K20c’s wider bus still wins overall. The GTX 670 does have a base clock of 915 MHz and a boost clock of 980 MHz; the K20c lists no base or boost clock in the data, only its memory clock.

Power and interface differences further separate the two. The K20c has a TDP of 225 W and requires a 550 W suggested PSU, while the GTX 670 draws 170 W and suggests a 450 W PSU. The K20c uses a 1x 6-pin + 1x 8-pin power connector configuration; the GTX 670 uses 2x 6-pin. The bus interface also differs: the K20c runs PCIe 2.0 x16, while the GTX 670 runs PCIe 3.0 x16, giving the latter a faster host connection.

Display outputs are a fundamental differentiator. The K20c has no outputs at all — it is a compute card, not a graphics card. The GTX 670 provides 2x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2, making it a fully functional desktop GPU. The K20c’s dimensions are 267 mm long, while the GTX 670 is 241 mm long, 111 mm tall, and 38 mm wide. Both are dual-slot cards.

Head-to-Head Benchmarks

The sole benchmark comparison available is Geekbench OpenCL. The GeForce GTX 670 scores 15,341, and the Tesla K20c scores 11,479. The delta is -25.2%, meaning the GTX 670 is 25.2% faster than the K20c in this test. This is the only head-to-head data point, but it is unambiguous: the GTX 670 wins.

Context from the nearest rivals clarifies where each card stands. The K20c’s nearest rivals by average score include the AMD Radeon Pro 5500M at 11,528 (-0.4%), the AMD Radeon RX 7800 XT at 11,627 (-1.3%), the NVIDIA GeForce GTX 1660 at 11,680 (-1.7%), and the NVIDIA GeForce GTX 780M at 11,261 (+1.9%). The K20c’s average score of 11,479 places it in the middle of this group, slightly behind the RX 7800 XT and GTX 1660, slightly ahead of the GTX 780M, and essentially tied with the Pro 5500M.

The GTX 670’s nearest rivals are the NVIDIA GeForce GTX 590 at 12,830 (-0.4%), the AMD Radeon Pro 455 at 12,831 (-0.4%), the AMD Radeon RX 7600M XT at 12,710 (+0.5%), and the AMD FirePro W5100 at 12,847 (-0.6%). The GTX 670’s average score of 12,773 puts it ahead of the RX 7600M XT and behind the GTX 590, Pro 455, and FirePro W5100 by less than 1%. The GTX 670 also has additional benchmark results beyond OpenCL: 7,424 in Geekbench Metal and 15,553 in Geekbench Vulkan, tests the K20c does not appear in.

The data shows the GTX 670 is the stronger OpenCL performer by a wide margin over the K20c, and it also holds a slightly higher overall percentile (52 vs 51). The K20c’s theoretical compute advantages do not translate into OpenCL success in this measurement.

Specification Differences

| Specification | NVIDIA Tesla K20c | NVIDIA GeForce GTX 670 |

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

| Chip | GK110 | GK104 |

| Generation | Tesla Kepler (Kxx) | GeForce 600 |

| Transistors | 7,080 million | 3,540 million |

| Die Size | 561 mm² | 294 mm² |

| Transistor Density | 12.6M / mm² | 12.0M / mm² |

| Base Clock | Not listed | 915 MHz |

| Boost Clock | Not listed | 980 MHz |

| Memory Clock | 1300 MHz (5.2 Gbps) | 1502 MHz (6 Gbps) |

| Memory Size | 5 GB | 2 GB |

| Memory Bus | 320 bit | 256 bit |

| Memory Bandwidth | 208.0 GB/s | 192.3 GB/s |

| Shading Units | 2,496 | 1,344 |

| TMUs | 208 | 112 |

| ROPs | 40 | 32 |

| Pixel Rate | 36.71 GPixel/s | 27.44 GPixel/s |

| Texture Rate | 146.8 GTexel/s | 109.8 GTexel/s |

| FP32 | 3.524 TFLOPS | 2.634 TFLOPS |

| TDP | 225 W | 170 W |

| Power Connectors | 1x 6-pin + 1x 8-pin | 2x 6-pin |

| Suggested PSU | 550 W | 450 W |

| Bus Interface | PCIe 2.0 x16 | PCIe 3.0 x16 |

| Display Outputs | No outputs | 2x DVI, 1x HDMI 1.4a, 1x DisplayPort 1.2 |

| Length | 267 mm | 241 mm |

| Release Date | 2012-11-11 | 2012-05-09 |

| Predecessor | Tesla Fermi | GeForce 500 |

| Successor | Tesla Maxwell | GeForce 700 |

| Launch MSRP | 3,199 USD | 399 USD |

FAQ

Q: Which card is faster in OpenCL?

A: The GeForce GTX 670 is faster. It scores 15,341 in Geekbench OpenCL, while the Tesla K20c scores 11,479, a delta of -25.2% in favor of the GTX 670.

Q: How much memory does each card have?

A: The Tesla K20c has 5 GB of GDDR5, while the GeForce GTX 670 has 2 GB of GDDR5. The K20c also has a wider 320-bit memory bus versus the GTX 670’s 256-bit bus.

Q: Can the Tesla K20c connect to a display?

A: No. The Tesla K20c has no display outputs. The GeForce GTX 670 provides 2x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2.

Q: Which card has more shading units?

A: The Tesla K20c has 2,496 shading units, nearly double the GeForce GTX 670’s 1,344. The K20c also has 208 TMUs versus 112, and 40 ROPs versus 32.

Q: What is the power consumption difference?

A: The Tesla K20c has a TDP of 225 W and requires a 550 W suggested PSU. The GeForce GTX 670 has a TDP of 170 W and requires a 450 W suggested PSU.

Q: How do their average benchmark scores compare?

A: The GeForce GTX 670 has an average benchmark score of 12,773, while the Tesla K20c has an average of 11,479. The GTX 670 also holds a 52nd percentile rank versus the K20c’s 51st.

The Verdict

The GeForce GTX 670 is the better choice for anyone who needs a functional graphics card. It wins the only head-to-head test by 25.2%, has a higher average benchmark score (12,773 vs 11,479), and provides full display outputs. Its PCIe 3.0 interface, lower 170 W TDP, and smaller 241 mm length make it more practical for desktop integration. The GTX 670 also supports Vulkan and Metal benchmarks, with scores of 15,553 and 7,424 respectively, demonstrating broader API coverage.

The Tesla K20c is the better choice for compute-heavy workloads that can use its resources. Its 5 GB memory capacity is 2.5 times larger than the GTX 670’s 2 GB, and its 208.0 GB/s bandwidth exceeds the GTX 670’s 192.3 GB/s. Its FP32 throughput of 3.524 TFLOPS is 33.8% higher than the GTX 670’s 2.634 TFLOPS, and its pixel rate of 36.71 GPixel/s versus 27.44 GPixel/s shows higher fill capability. The K20c’s 7,080 million transistors on a 561 mm² die represent a fundamentally more powerful compute chip.

The data does not support the K20c for general-purpose use. It has no display outputs, a higher power draw, and loses the only available benchmark. The GTX 670 is simply faster in measured OpenCL performance and more versatile as a hardware product. However, the K20c’s memory and compute specifications suggest it was designed for a different class of workload — one that the benchmark data here does not capture. For users whose applications can leverage 5 GB of memory and 3.5 TFLOPS of FP32, the K20c has advantages the GTX 670 cannot match. For everyone else, the GTX 670 is the data-backed winner. The launch MSRP difference is stark — 3,199 USD for the K20c versus 399 USD for the GTX 670 — and the performance data does not justify that gap in consumer terms.