GPU Comparison

The NVIDIA GeForce 920MX and the NVIDIA GeForce GT 640 represent two distinct approaches to mobile and desktop graphics, separated by four years of architectural evolution. The data reveals a fascinating split: the 920MX, a Maxwell-based mobile part, wins decisively in OpenCL compute workloads, while the older Kepler-based GT 640 dominates in Vulkan graphics performance with a commanding 21.6% margin. This head-to-head matchup is not a simple generational victory, but a nuanced tale of how architecture, memory bandwidth, and intended use case shape real-world benchmark outcomes.

Where Each One Wins

The benchmark results draw a clear line between compute-oriented tasks and graphics-heavy applications. The NVIDIA GeForce 920MX takes the crown in the Geekbench OpenCL test, scoring 4068 against the GT 640's 3736, an 8.9% advantage. This suggests that for workloads leveraging general-purpose GPU computing, such as video encoding, image processing, or scientific calculations, the newer Maxwell architecture provides a tangible benefit despite its smaller memory bus and lower overall transistor count.

Conversely, the NVIDIA GeForce GT 640 is the clear winner in the Geekbench Vulkan test, posting a score of 3809 compared to the 920MX's 2987. The 21.6% deficit for the 920MX is substantial, indicating that for gaming or any Vulkan-accelerated graphics workload, the older desktop card holds a significant edge. This is likely attributable to the GT 640's wider 128-bit memory interface, which delivers 28.51 GB/s of bandwidth versus the 920MX's 14.40 GB/s, a nearly 2x difference that directly impacts texture streaming and frame buffer performance in graphics APIs.

In terms of raw compute throughput, the GT 640 also holds the theoretical advantage with 692.7 GFLOPS of FP32 performance versus the 920MX's 508.4 GFLOPS. However, real-world OpenCL results invert this expectation, suggesting that Maxwell's architecture is more efficient at extracting performance from its shading units. The wins are split evenly at one apiece, but the magnitude of the Vulkan loss for the 920MX is more than double the scale of its OpenCL gain.

Architecture Differences

The two GPUs are built on fundamentally different architectures from different eras. The 920MX uses the GM108S chip based on Maxwell architecture, manufactured on a 28 nm process at TSMC. It packs 1,020 million transistors into a die size of just 77 mm², yielding a transistor density of 13.2M per mm². This is a compact, power-efficient design aimed at thin-and-light laptops, reflected in its 16 W TDP and MXM module slot width.

The GT 640, in contrast, utilizes the GK107 chip based on the older Kepler architecture, also on a 28 nm process but with a larger 118 mm² die. It houses 1,270 million transistors, giving it a lower transistor density of 10.8M per mm². This is a desktop-oriented part with a 65 W TDP, a single-slot form factor, and a 250 W suggested power supply, highlighting its more demanding power profile. The 920MX is part of the GeForce 900M generation, released in 2016, while the GT 640 belongs to the GeForce 600 series, launched in 2012.

Memory subsystems differ markedly. The 920MX features 2 GB of DDR3 on a 64-bit bus, while the GT 640 also has 2 GB of DDR3 but on a 128-bit bus. This doubles the GT 640's memory bandwidth to 28.51 GB/s. The core configurations also diverge: the 920MX has 256 shading units, 24 TMUs, and 8 ROPs, whereas the GT 640 offers 384 shading units, 32 TMUs, and 16 ROPs. These differences explain the GT 640's higher theoretical pixel rate (7.216 GPixel/s vs 7.944 GPixel/s, actually the 920MX is higher here) and texture rate (28.86 GTexel/s vs 23.83 GTexel/s, the GT 640 wins this).

Head-to-Head Benchmarks

The Geekbench OpenCL result is a surprising victory for the 920MX. Scoring 4068 against the GT 640's 3736, the 920MX outperforms its rival by 8.9% despite having fewer shading units (256 vs 384) and lower FP32 peak (508.4 GFLOPS vs 692.7 GFLOPS). This suggests Maxwell's scheduler and execution efficiency are superior for compute workloads, allowing it to extract more useful work per clock cycle. The result is even more notable given the 920MX's severely constrained memory bandwidth, indicating that the OpenCL benchmark is not bandwidth-limited but rather compute or latency-limited.

The Vulkan benchmark tells the opposite story. The GT 640 dominates with a score of 3809, leaving the 920MX trailing at 2987, a 21.6% gap. This is a massive margin that points to the GT 640's superior memory bandwidth (28.51 GB/s vs 14.40 GB/s) and its higher texture fill rate (28.86 GTexel/s vs 23.83 GTexel/s). Vulkan workloads often stress geometry and texture throughput, where the GT 640's 32 TMUs and 16 ROPs provide a clear advantage over the 920MX's 24 TMUs and 8 ROPs. The GT 640's higher pixel rate (7.216 GPixel/s vs 7.944 GPixel/s, wait, the 920MX is higher here) does not save it, suggesting that memory bandwidth is the limiting factor in this graphics API.

When looking at average benchmark scores, the 920MX leads with 3528 versus the GT 640's 3210, a difference of roughly 9.9%. This aggregate figure, however, masks the divergent strengths revealed by individual tests. The 920MX's nearest rivals include the NVIDIA GeForce GT 545 (deltaPct -1.8%), RTX 5000 Mobile Ada Generation (-1.9%), and GT 735M (-2.4%), while it sits 2.7% ahead of the Quadro 2000M. The GT 640, meanwhile, is 0.9% ahead of the Intel Arc Pro B60, 1.5% ahead of the Quadro P1000, but trails the GeForce 920M by 2.3% and the GT 730M by 3.2%.

Specification Differences

The two GPUs diverge on several key specification fields. The most striking differences are in their power and form factor: the 920MX draws only 16 W and uses an MXM Module slot, while the GT 640 consumes 65 W and occupies a Single-slot with a 145 mm length (5.7 inches). The GT 640 also lists a suggested PSU of 250 W, whereas the 920MX has no such requirement. Both use no power connectors, but the GT 640's desktop nature means it draws power from the PCIe slot.

Memory architecture differs significantly: the 920MX has a 64-bit bus with 14.40 GB/s bandwidth, while the GT 640 uses a 128-bit bus with 28.51 GB/s. Core counts also vary: 256 shading units, 24 TMUs, and 8 ROPs for the 920MX versus 384 shading units, 32 TMUs, and 16 ROPs for the GT 640. The 920MX has a base clock of 965 MHz and boost clock of 993 MHz, while the GT 640's core clocks are not listed. Memory clocks are close: 900 MHz (1800 Mbps effective) for the 920MX versus 891 MHz (1782 Mbps effective) for the GT 640.

The bus interface differs: the 920MX uses PCIe 3.0 x8, while the GT 640 uses PCIe 3.0 x16, potentially affecting data transfer rates in bandwidth-sensitive scenarios. Display outputs also differ: the 920MX is "Portable Device Dependent" (no fixed outputs), while the GT 640 offers 1x DVI, 1x HDMI 1.4a, and 1x DisplayPort 1.2. The GT 640 supports Vulkan 1.2.175, while the 920MX supports Vulkan 1.4. Both support DirectX 12 (11_0) and OpenGL 4.6. The launch MSRP of the GT 640 was 99 USD; the 920MX has no listed launch MSRP.

FAQ

Q: Which GPU has higher raw compute performance?

A: The GT 640 has a higher theoretical FP32 peak at 692.7 GFLOPS, compared to the 920MX's 508.4 GFLOPS. However, in the Geekbench OpenCL test, the 920MX scores 4068 versus the GT 640's 3736, an 8.9% advantage, indicating better real-world compute efficiency.

Q: Why does the GT 640 dominate in Vulkan benchmarks?

A: The GT 640's 128-bit memory bus provides 28.51 GB/s of bandwidth, nearly double the 920MX's 14.40 GB/s. Its 32 TMUs and 16 ROPs also exceed the 920MX's 24 TMUs and 8 ROPs, which is critical for graphics workloads. The Vulkan score reflects this: 3809 for GT 640 versus 2987 for 920MX.

Q: What are the power consumption differences?

A: The 920MX is rated at 16 W TDP, making it suitable for portable devices, while the GT 640 draws 65 W and requires a 250 W suggested power supply. The 920MX uses an MXM Module slot, whereas the GT 640 is a single-slot desktop card measuring 145 mm (5.7 inches).

Q: How do their memory subsystems compare?

A: Both have 2 GB of DDR3 memory, but the 920MX uses a 64-bit bus with 14.40 GB/s bandwidth, while the GT 640 offers a 128-bit bus with 28.51 GB/s. The memory clocks are similar: 900 MHz (1800 Mbps effective) for the 920MX versus 891 MHz (1782 Mbps effective) for the GT 640.

Q: Which GPU has better driver and API support?

A: The 920MX supports Vulkan 1.4, while the GT 640 supports Vulkan 1.2.175. Both support DirectX 12 (11_0) and OpenGL 4.6. The newer 920MX (released 2016) may have more recent driver optimizations, but the GT 640 (released 2012) still receives support for its listed APIs.

Q: How do these GPUs rank against other cards?

A: The 920MX has an average benchmark score of 3528 and sits at the 21st percentile of all GPUs, while the GT 640 scores 3210 on average at the 20th percentile. The 920MX's nearest rival is the GT 545 (deltaPct -1.8%), while the GT 640's closest competitor is the Intel Arc Pro B60 (deltaPct 0.9%).

The Verdict

The data supports a clear recommendation based on use case. For users prioritizing compute performance, OpenCL workloads, or power efficiency, the NVIDIA GeForce 920MX is the superior choice. Its 8.9% OpenCL lead over the GT 640, combined with its 16 W TDP versus 65 W, makes it an excellent fit for mobile devices where battery life and thermals matter. The 920MX also has a higher pixel rate (7.944 GPixel/s vs 7.216 GPixel/s) and a newer architecture with Vulkan 1.4 support.

For gamers or users running Vulkan-accelerated graphics applications, the NVIDIA GeForce GT 640 is unequivocally better. Its 21.6% Vulkan advantage is decisive, driven by double the memory bandwidth and higher texture and ROP counts. The GT 640's 28.51 GB/s bandwidth and 28.86 GTexel/s texture rate provide a substantial edge in rendering scenarios, making it the better choice for desktop systems where power consumption is less of a concern.

The split is nearly even, one win each, but the magnitudes differ. The 920MX's OpenCL victory is modest at 8.9%, while the GT 640's Vulkan triumph is a landslide at 21.6%. This asymmetry suggests that for most users, the GT 640 offers the more compelling overall package for graphics-centric tasks, while the 920MX excels in compute and efficiency. The choice ultimately hinges on whether the workload is compute-bound or graphics-bound, with the GT 640's memory bandwidth making it the safer bet for general-purpose desktop use.