NVIDIA Tesla M10

NVIDIA’s Tesla M10 is an end-of-life compute accelerator built on the Maxwell architecture, using the GM107 chip fabricated on a 28 nm process at TSMC. It packs 1,870 million transistors on a 148 mm² die, yielding a transistor density of 12.6 million per square millimeter. The card ships with 640 shading units, 40 texture mapping units, and 16 ROPs. Its average benchmark score across the available tests is 9634, placing it at the 45th percentile of all GPUs in the database. This is a decidedly mid-pack result, and the data suggests the M10 sits in a crowded performance tier where tiny percentage swings separate it from several older consumer and workstation parts.

Benchmark Performance

The Tesla M10’s synthetic performance, as measured by the average of its Geekbench OpenCL and Vulkan scores, lands at 9634. Breaking down the individual tests, the card scores 10318 in Geekbench OpenCL and 8950 in Geekbench Vulkan. The gap between these two is notable: the OpenCL result is roughly 15% higher than the Vulkan result, which indicates that the Maxwell architecture’s compute paths are better optimized for OpenCL workloads in this specific benchmark suite. For a card with no display outputs, this compute-oriented skew is unsurprising.

Relative to its nearest rivals, the M10 is essentially in a dead heat, with deltas of less than half a percent in either direction. Against the NVIDIA GeForce GTX 650 Ti Boost, the M10 trails by a negligible 0.3%. Against the AMD Radeon Pro WX 2100, it is 0.4% behind. Versus the NVIDIA GeForce GTX 960M, it is also 0.4% down. The only rival it edges out is the NVIDIA GeForce GTX 465, where the M10 leads by 0.4%. These deltas are well within run-to-run variance for any benchmark, meaning the M10 is functionally performance-identical to all four of these cards in aggregate.

What does this mean in practice? The M10’s 1.672 TFLOPS of FP32 compute and 52.24 GTexel/s texture rate are the raw numbers driving these scores. The pixel rate of 20.90 GPixel/s is modest, consistent with the 16 ROPs. In real workloads, this card will not break any speed records—it sits firmly in the entry-level compute segment, roughly equivalent to a mid-range GPU from 2012 or a low-end mobile part from 2014. The 45th percentile ranking confirms this: more than half of all GPUs in the database outperform it, while a large chunk of older hardware sits just below.

Power and Cooling

The Tesla M10 carries a thermal design power of 225 W, which is substantial for a card with this level of performance. That TDP is roughly 2.3 times higher than what you would expect from a modern 100 W-class desktop GPU, and it reflects the Maxwell architecture’s older process node and the card’s compute-oriented power delivery. The board is dual-slot in width and measures 267 mm (10.5 inches) in length, so it will require a reasonably spacious chassis.

Power delivery is handled by a single 8-pin PCIe power connector. The suggested power supply rating is 550 W, which is a conservative figure given the 225 W TDP—it leaves ample headroom for a typical CPU and motherboard combination. The card interfaces with the system via PCIe 3.0 x16, so it will run in virtually any modern motherboard without issue. Notably, the card has no display outputs, meaning it is strictly a compute or rendering accelerator that must be paired with a separate GPU for any visual output. The 225 W TDP also means cooling is non-trivial; the dual-slot design is necessary to dissipate the heat generated under sustained load.

How It Compares

NVIDIA GeForce GTX 650 Ti Boost: The M10 trails this rival by 0.3% in average score, making them inseparable in real-world compute performance. Both are Maxwell-era parts, but the 650 Ti Boost is a consumer gaming card with display outputs. The M10’s edge is its 8 GB VRAM, which is double what most 650 Ti Boost cards shipped with, but raw compute throughput is statistically identical.

NVIDIA GeForce GTX 465: The M10 leads this card by 0.4%. The GTX 465 is a much older Fermi-based part, so the M10’s slight win is expected given the architectural generational leap from Fermi to Maxwell. However, the margin is so thin that any driver or benchmark variation could flip the result. The M10’s lower power draw per unit of compute is a more meaningful advantage than the score delta.

NVIDIA GeForce GTX 960M: The M10 is 0.4% behind this mobile GPU. The GTX 960M is a laptop part with significantly lower TDP and clock speeds, yet it matches the M10’s aggregate compute output. This highlights how inefficient the M10 is by modern standards—a mobile chip from the same era delivers the same performance at a fraction of the power. The M10’s sole advantage is its 8 GB VRAM versus the 960M’s typical 2-4 GB.

AMD Radeon Pro WX 2100: The M10 is 0.4% behind this workstation card. The WX 2100 is a newer, Polaris-based part with modern feature support, yet it barely edges out the older Maxwell design. This suggests that for raw compute throughput, the M10 is competitive with entry-level professional cards from several generations later. However, the WX 2100 offers display outputs and newer API support, making it a more flexible option despite the near-identical score.

Who Should Consider It

The benchmark data indicates the Tesla M10 is suited for specific, narrow use cases. With an average score of 9634 and a 45th percentile ranking, this card is not a candidate for high-refresh or high-resolution gaming. At 1080p with medium settings, it would struggle to maintain playable frame rates in modern titles, and its lack of display outputs means it cannot serve as a primary gaming GPU anyway.

The realistic target is compute acceleration in a server or workstation environment where the 8 GB VRAM is the primary selling point. For tasks like batch image processing, virtual desktop infrastructure, or light CUDA workloads that require more memory than typical consumer cards offer, the M10’s 8 GB frame buffer is its main attraction. However, the raw compute power is only on par with a GTX 650 Ti Boost, so any workload that is compute-bound rather than memory-bound will see lackluster performance. At 1440p or 4K resolutions, the M10’s 83.20 GB/s bandwidth and 1.672 TFLOPS will be bottlenecks—these resolutions require far more memory throughput and shading power than this card provides. The 45th percentile ranking reinforces that this is a budget compute card, not a performance part.

Ray Tracing and Feature Set

The Tesla M10 has no ray tracing cores and no tensor cores, as it is based on the Maxwell architecture that predates NVIDIA’s RTX line. DirectX 12 support is present but limited to the 11_0 feature level, which means it cannot handle the more advanced DirectX 12 Ultimate features like mesh shaders or variable rate shading. OpenGL 4.6 and Vulkan 1.4 are supported, providing reasonable compatibility with modern compute and graphics APIs, but the hardware lacks any dedicated acceleration for ray-traced workloads.

This is a pure rasterization and compute card. For anyone considering it for ray tracing, the data is clear: it will run software-based ray tracing at extremely low performance, if at all, and the absence of tensor cores means no AI-accelerated features like DLSS. The M10’s feature set is firmly rooted in the 2015-2016 era, and it offers no forward-looking capabilities. Its API support is adequate for legacy applications, but modern software expecting DirectX 12 Ultimate or hardware ray tracing will either fail to run or fall back to CPU-based processing.

Memory Subsystem

The Tesla M10 is equipped with 8 GB of GDDR5 memory on a 128-bit bus, yielding a memory bandwidth of 83.20 GB/s. The memory clock is 1300 MHz, which translates to 5.2 Gbps effective. This configuration is unusual: 8 GB on a 128-bit bus is a capacity-first design, prioritizing frame buffer size over bandwidth. For comparison, the rival GTX 650 Ti Boost and GTX 960M typically ship with 1-2 GB or 2-4 GB respectively, so the M10’s capacity advantage is significant.

However, the bandwidth is a severe limitation. At 83.20 GB/s, the M10 has less than a third of the bandwidth of a modern mid-range GPU. This means that while it can hold large datasets in VRAM, moving data in and out is slow. For high-resolution textures at 4K, the 8 GB capacity is sufficient, but the bandwidth will cause stuttering and low frame rates as the GPU waits for data. In compute workloads, the 128-bit bus will throttle any memory-intensive operation. The 20.90 GPixel/s pixel rate and 52.24 GTexel/s texture rate are consistent with this memory constraint—the card is balanced, but that balance is at a low performance level. The 8 GB capacity is the sole standout specification, making the M10 a niche option for workloads that need large VRAM pools but tolerate low bandwidth.