NVIDIA GeForce GTX 670 vs NVIDIA Tesla K10 Comparison

The Tesla K10 and the GeForce GTX 670 are two faces of the same silicon. Both are built on NVIDIA's Kepler architecture around the GK104 chip, both rolled out within days of each other in the spring of 2012, and both are now end-of-life parts. Yet they were aimed at completely different buyers: one is a passive-display compute accelerator with a five-thousand-dollar launch MSRP, the other a consumer graphics card that shipped with a full set of display outputs. The recorded data shows how differently the same design philosophy plays out when tuned for compute versus gaming.

Where Each One Wins

This comparison is lopsided in an interesting way: the GeForce GTX 670 wins the only head-to-head benchmark the database contains for this pairing, beating the Tesla K10 by 8.6 percent in Geekbench OpenCL, 15341 to 14029. So on raw measured performance in the one test where they overlap, the consumer card is clearly ahead.

That single result does not tell the whole story, though. The K10's strengths sit in areas the benchmark tables do not fully capture: it carries 4 GB of GDDR5 memory against the GTX 670's 2 GB, doubling the capacity available for large working sets, and it fields more shading hardware, 1536 shading units and 128 TMUs versus 1344 and 112. For workloads that scale with raw ALU count and memory footprint, the K10 holds the hardware advantage on paper even though the measured OpenCL result favors the GeForce card.

The GTX 670, by contrast, wins everywhere a gamer cares about. It posts higher pixel rate (27.44 GPixel/s against 23.84), higher texture rate (109.8 GTexel/s against 95.36), higher memory bandwidth (192.3 GB/s against 160.0), and higher FP32 throughput (2.634 TFLOPS against 2.289). It also has display outputs, which the K10 lacks entirely. In practical terms, the GTX 670 is the only one of the two that can actually drive a monitor.

Architecture Differences

On paper these are near twins. Both use the GK104 chip, Kepler architecture, TSMC's 28 nm process, 3,540 million transistors, a 294 mm² die, and identical transistor density of 12.0M per mm². Both connect over PCIe 3.0 x16 and support the same API levels: DirectX 12 (11_0), OpenGL 4.6, and Vulkan 1.2.175. Neither has RT cores or tensor cores, which makes sense given their age.

The differences are in configuration and tuning. The Tesla K10 runs its memory at 1250 MHz (5 Gbps effective) on a 256-bit bus for 160.0 GB/s of bandwidth, while the GTX 670 pushes memory faster, 1502 MHz (6 Gbps effective) on the same bus width, reaching 192.3 GB/s. Neither card's core clock strategy is fully recorded in the database, but the K10's listed base and boost clocks are absent while the GTX 670 carries 915 MHz base and 980 MHz boost.

The K10 compensates with more functional units: those 1536 shading units and 128 TMUs against 1344 and 112. Yet because the GTX 670 clocks its fixed-function and shading hardware higher, it ends up with higher pixel, texture, and FP32 rates despite having fewer units. This is a classic case of clock count versus unit count, and the measured data says the clock advantage wins for graphics-style workloads.

Power delivery reflects their different roles. The K10 is a 225 W board needing one 6-pin plus one 8-pin connector and a suggested 550 W PSU. The GTX 670 is a 170 W card with two 6-pin connectors and a 450 W PSU suggestion. The K10 is also physically longer at 272 mm versus 241 mm, and its recorded dimensions stop there, whereas the GTX 670 has full measurements on file: 241 mm long, 111 mm tall, 38 mm wide. Both are dual-slot cards.

The display situation is the sharpest divide. The GTX 670 offers two DVI outputs, one HDMI 1.4a, and one DisplayPort 1.2. The Tesla K10 has no outputs at all, which confirms its role as a dedicated compute accelerator rather than a graphics card.

Head-to-Head Benchmarks

The database holds exactly one overlapping test between these two, and it is decisive.

In Geekbench OpenCL, the GeForce GTX 670 scores 15341 and the Tesla K10 scores 14029. The GTX 670 wins by 8.6 percent. Given that both cards share the same die, the same process, and the same architecture, that gap comes down to configuration: the GTX 670's higher memory bandwidth (192.3 GB/s versus 160.0) and higher FP32 throughput (2.634 TFLOPS versus 2.289) translate directly into a faster OpenCL result, more than offsetting the K10's larger shading unit count.

The percentile rankings reinforce the picture. The GTX 670 sits at the 52nd percentile against all GPUs in the database, the K10 at the 55th, so both are mid-pack parts by modern standards, separated by a hair. Their rival clusters differ, though. The K10's average score of 14029 puts it within striking distance of the GeForce GTX 680, which scores 14150 and edges it out by under a percent. It also sits just above the AMD Radeon RX 570X (13871), the RTX A2000 Mobile (13821), and the Radeon 660M (13812). The GTX 670's average of 12773 places it in a tighter knot around the GeForce GTX 590 (12830), the Radeon Pro 455 (12831), the Radeon RX 7600M XT (12710), and the FirePro W5100 (12847), all within about half a percent of each other.

The GTX 670 also carries additional Geekbench results the K10 does not: 7424 in Geekbench Metal and 15553 in Geekbench Vulkan. That Vulkan score is the highest single number recorded for either card in this dataset.

The bottom line from the benchmark table is simple: the consumer card swept the head-to-head, one test, one win. But the K10's cluster placement shows it performing closer to a GTX 680 in average terms than its compute-branded sibling does.

FAQ

Q: Which card is faster in benchmarks?

A: The GeForce GTX 670. It won the only direct comparison, Geekbench OpenCL, by 8.6 percent, 15341 versus 14029, and it also leads in pixel rate, texture rate, memory bandwidth, and FP32 throughput.

Q: Does the Tesla K10 have any hardware advantages?

A: Yes. It has more shading units (1536 versus 1344), more TMUs (128 versus 112), and twice the memory capacity, 4 GB versus 2 GB. These favor large compute workloads even though the measured OpenCL result went the other way.

Q: Can the Tesla K10 be used for gaming with a monitor attached?

A: Not directly. The database lists no display outputs for the K10, while the GTX 670 provides two DVI, one HDMI 1.4a, and one DisplayPort 1.2 connectors.

Q: How do their power requirements differ?

A: The K10 is rated at 225 W with one 6-pin and one 8-pin connector and a suggested 550 W PSU. The GTX 670 is rated at 170 W with two 6-pin connectors and a suggested 450 W PSU.

Q: Are they built on the same chip?

A: Yes. Both use the GK104 chip on Kepler architecture, TSMC 28 nm, with 3,540 million transistors and a 294 mm² die.

Q: How do they compare against other GPUs in the database?

A: The K10 ranks in the 55th percentile, the GTX 670 in the 52nd. Both land mid-pack, with the K10 averaging 14029 and the GTX 670 averaging 12773 across their recorded tests.

The Verdict

For anyone choosing between these two today, the data points firmly at the GeForce GTX 670. It is faster in the one measured head-to-head, it wins every rate metric that matters for graphics (pixels, texels, bandwidth, FP32), it draws less power at 170 W versus 225 W, it needs a smaller PSU, and it is the only one of the pair with display outputs. It even recorded extra benchmark results, including a Vulkan score of 15553.

The Tesla K10 only makes sense in a narrow scenario: a compute workload that needs more than 2 GB of memory and benefits from its larger shading unit count, running headless with no monitor attached. Its OpenCL average of 14029 and its placement near the GTX 680 show it is no slouch, but against its own consumer sibling in the one test they share, it loses. For gaming, display work, and general use, the GTX 670 is the pick. For headless compute with a memory capacity constraint, the K10's 4 GB is the lone practical argument in its favor.

Specification Differences

• Memory size: Tesla K10: 4 GB GDDR5, GTX 670: 2 GB GDDR5
• Memory speed: K10 at 1250 MHz (5 Gbps effective), GTX 670 at 1502 MHz (6 Gbps effective)
• Memory bandwidth: K10: 160.0 GB/s, GTX 670: 192.3 GB/s (both 256-bit bus)
• Shading units: K10: 1536, GTX 670: 1344
• TMUs: K10: 128, GTX 670: 112
• ROPs: 32 on both (no difference)
• Core clocks: K10 base and boost not recorded, GTX 670: 915 MHz base, 980 MHz boost
• Pixel rate: K10: 23.84 GPixel/s, GTX 670: 27.44 GPixel/s
• Texture rate: K10: 95.36 GTexel/s, GTX 670: 109.8 GTexel/s
• FP32: K10: 2.289 TFLOPS, GTX 670: 2.634 TFLOPS
• TDP: K10: 225 W, GTX 670: 170 W
• Power connectors: K10: 1x 6-pin + 1x 8-pin, GTX 670: 2x 6-pin
• Suggested PSU: K10: 550 W, GTX 670: 450 W
• Display outputs: K10: none, GTX 670: 2x DVI, 1x HDMI 1.4a, 1x DisplayPort 1.2
• Length: K10: 272 mm, GTX 670: 241 mm
• Launch MSRP: K10: 5,099 USD, GTX 670: 399 USD
• Release date: K10: April 2012, GTX 670: May 2012
• Generation lineage: K10: Tesla Kepler (Kxx), predecessor Tesla Fermi, successor Tesla Maxwell. GTX 670: GeForce 600, predecessor GeForce 500, successor GeForce 700