NVIDIA GeForce GTX 670 vs NVIDIA Tesla K20Xm Comparison

The NVIDIA GeForce GTX 670 and the NVIDIA Tesla K20Xm are both Kepler-based cards from the same era, but they were built for different corners of the market. The GTX 670 is a consumer gaming card, while the K20Xm is a compute-focused accelerator. The benchmark data shows a clear, if not overwhelming, advantage for the Tesla part, but the story is more nuanced than just raw scores. The GTX 670 averages 12,773 points across its benchmark suite, while the K20Xm averages 12,625 points, making the GTX 670 the faster card on average by about 1.2%. However, this average is pulled down by the K20Xm lacking a Vulkan score, and the head-to-head results tell a different tale.

Head-to-Head Benchmarks

The two cards were tested against each other in only two shared workloads, and the Tesla K20Xm won both. In the Geekbench Metal test, the K20Xm scored 8,035 points against the GTX 670’s 7,424 points, a 7.6% lead. That is a solid margin, indicating the Tesla’s larger compute core count provides a tangible advantage in this API. The gap widens significantly in Geekbench OpenCL, where the K20Xm posts 17,215 points versus the GTX 670’s 15,341 points, a 10.9% difference. This is the Tesla’s home turf; OpenCL is a compute-oriented API, and the K20Xm’s 2,688 shading units simply overpower the GTX 670’s 1,344.

Looking at the broader benchmark landscape, the GTX 670’s average score of 12,773 is built on three tests: Metal, OpenCL, and Vulkan. The Vulkan score of 15,553 is its highest, and it is a workload where the Tesla cannot compete because it has no Vulkan result. The K20Xm’s average of 12,625 comes from just two tests, and it lacks the Vulkan data point entirely. This means the GTX 670’s average is artificially boosted by a strong Vulkan performance that the Tesla never demonstrates. In the tests where both cards appear, the Tesla wins decisively. The nearest rival data confirms the positioning: the GTX 670 is 0.5% ahead of the AMD Radeon RX 7600M XT, while the K20Xm is 0.7% behind that same AMD card. The delta between the two Nvidia cards is small in aggregate, but the per-test deltas favor the Tesla by a consistent 7-11%.

FAQ

Q: Which card has the higher average benchmark score?

A: The GTX 670, with an average of 12,773 points, is 1.2% higher than the K20Xm’s 12,625 points. This is partly because the GTX 670 has a Vulkan score (15,553) that the Tesla lacks.

Q: Which card wins the shared Metal and OpenCL benchmarks?

A: The Tesla K20Xm wins both. It scores 8,035 in Metal (7.6% ahead) and 17,215 in OpenCL (10.9% ahead) compared to the GTX 670’s 7,424 and 15,341, respectively.

Q: How do these cards compare to their nearest rivals?

A: The GTX 670 is 0.5% faster than the AMD Radeon RX 7600M XT and 0.4% slower than the NVIDIA GeForce GTX 590. The K20Xm is 0.7% slower than the RX 7600M XT and 1.2% slower than the GTX 670 itself.

Q: Do both cards support the same modern APIs?

A: Yes, both support DirectX 12 (11_0), OpenGL 4.6, and Vulkan 1.2.175. However, the Tesla K20Xm has no benchmark scores recorded for Vulkan, while the GTX 670 does.

Q: What is the memory configuration difference?

A: The GTX 670 has 2 GB of GDDR5 on a 256-bit bus, delivering 192.3 GB/s. The K20Xm has 6 GB of GDDR5 on a 384-bit bus, delivering 249.6 GB/s.

Q: Which card has more shading units and texture units?

A: The K20Xm has exactly double: 2,688 shading units and 224 TMUs, versus the GTX 670’s 1,344 shading units and 112 TMUs. The Tesla also has more ROPs: 48 versus 32.

Architecture Differences

Both cards are built on the Kepler architecture using a 28 nm process at TSMC, but they use different physical chips. The GTX 670 is based on the GK104 chip, which has 3,540 million transistors on a 294 mm² die, yielding a transistor density of 12.0 million per mm². The K20Xm uses the larger GK110 chip, with 7,080 million transistors on a 561 mm² die, giving it a density of 12.6 million per mm². The GK110 is physically more than twice the size of the GK104 and packs twice the transistor count. This directly translates into the core config: the K20Xm has 2,688 shading units, 224 TMUs, and 48 ROPs, while the GTX 670 has 1,344 shading units, 112 TMUs, and 32 ROPs. The Tesla’s compute throughput is correspondingly higher, with 3.935 TFLOPS of FP32 performance versus 2.634 TFLOPS for the GTX 670.

Memory architecture also diverges. The GTX 670 uses a 256-bit bus with 2 GB of GDDR5 running at 6 Gbps effective, achieving 192.3 GB/s. The K20Xm uses a wider 384-bit bus with 6 GB of GDDR5 at 5.2 Gbps effective, achieving 249.6 GB/s. The Tesla has three times the memory capacity and 30% more bandwidth, which is crucial for large compute datasets. The pixel and texture rates reflect this: the K20Xm manages 40.99 GPixel/s and 164.0 GTexel/s, while the GTX 670 manages 27.44 GPixel/s and 109.8 GTexel/s. The GTX 670 does have a higher memory clock in terms of effective data rate, but the Tesla’s wider bus compensates with more total bandwidth.

Specification Differences

The key differences between the two cards are stark. The GTX 670 has a base clock of 915 MHz and a boost clock of 980 MHz, while the K20Xm has no listed base or boost clocks—it is not a boost-oriented card. The GTX 670’s memory runs at 1502 MHz (6 Gbps effective), while the K20Xm’s runs at 1300 MHz (5.2 Gbps effective). The GTX 670 has 2 GB of memory, the K20Xm has 6 GB. The bus width is 256-bit versus 384-bit, and bandwidth is 192.3 GB/s versus 249.6 GB/s.

Power and physical specs differ as well. The GTX 670 has a TDP of 170 W and requires a 450 W power supply, while the K20Xm has a TDP of 235 W and requires a 550 W supply. The GTX 670 uses two 6-pin power connectors; the K20Xm lists no power connectors. The GTX 670 is 241 mm long and 111 mm tall, while the K20Xm is 267 mm long with no listed height or width. Both are dual-slot cards. The display outputs are a major split: the GTX 670 has 2x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2, while the K20Xm has no outputs at all—it is a compute-only card.

The launch MSRP for the GTX 670 was 399 USD, while the K20Xm launched at 7,699 USD. The GTX 670 released on 2012-05-09, and the K20Xm on 2012-11-11. Both are end-of-life products, with the GTX 670 succeeding the GeForce 500 series and preceding the GeForce 700 series, and the K20Xm succeeding Tesla Fermi and preceding Tesla Maxwell.

Where Each One Wins

The Tesla K20Xm wins in raw compute throughput. Its FP32 performance of 3.935 TFLOPS is 49% higher than the GTX 670’s 2.634 TFLOPS. It also wins in memory capacity and bandwidth, offering 6 GB versus 2 GB and 249.6 GB/s versus 192.3 GB/s. In the two shared benchmarks, Metal and OpenCL, the K20Xm is ahead by 7.6% and 10.9%, respectively. This makes it the better choice for compute-heavy workloads like OpenCL processing, where its higher shading unit count and wider memory bus can be fully utilized.

The GTX 670 wins in the aggregate benchmark average, scoring 12,773 versus 12,625, a 1.2% edge. It also wins in the Vulkan benchmark with a score of 15,553, a test the K20Xm does not even participate in. The GTX 670 has a lower TDP (170 W versus 235 W) and a lower suggested PSU requirement (450 W versus 550 W). It also has display outputs, making it a functional card for a desktop workstation that needs to drive a monitor. The GTX 670’s higher memory clock (6 Gbps effective versus 5.2 Gbps) does not translate to a bandwidth win, but it does indicate a different memory tuning philosophy.

The Verdict

If the workload is purely about compute performance in OpenCL or Metal, the NVIDIA Tesla K20Xm is the better card. The data shows it is 7.6% faster in Metal and 10.9% faster in OpenCL than the GTX 670, and it carries double the shading units, double the TMUs, and three times the memory. Its 6 GB frame buffer and 249.6 GB/s bandwidth are clearly aimed at large data sets. The lack of display outputs is irrelevant if the card is sitting in a server or a compute node.

If the workload includes Vulkan, or if the card must also drive a display, the NVIDIA GeForce GTX 670 is the better pick. It is the only one of the two with a Vulkan score, and it posts a higher average benchmark score overall. Its lower power draw and smaller physical footprint make it easier to integrate into a standard desktop build. The GTX 670 is also the more balanced card, with a 52nd percentile ranking against all GPUs, the same as the K20Xm, but with a higher raw average.

The choice comes down to specialization. The K20Xm is a dedicated compute accelerator that wins every test it shares with the GTX 670. The GTX 670 is a general-purpose card that happens to be competitive in compute while also offering display outputs and Vulkan support. For a builder focused on compute throughput, the K20Xm’s 10.9% OpenCL lead is decisive. For anyone who needs a card that can also render frames, the GTX 670’s versatility and Vulkan capability make it the practical option. Both cards are end-of-life, but the data suggests they served distinct purposes well.