GPU Comparison

The NVIDIA Tesla M2090 and AMD FirePro W5100 represent two distinct eras of professional GPU design, separated by nearly three years of architectural evolution. The data shows that the Tesla M2090, built on Fermi 2.0, wins the only directly comparable benchmark, while the FirePro W5100 counters with modern features and a dramatically lower power envelope. In the Geekbench OpenCL test, the Tesla M2090 scores 13075 against the FirePro W5100’s 11888, a 10% advantage for the NVIDIA part. This head-to-head result places the M2090 in the 53rd percentile of all GPUs, while the W5100 sits just one point lower at the 52nd percentile, making the overall performance gap nearly imperceptible in a broader context.

Head-to-Head Benchmarks

The sole direct comparison available is the Geekbench OpenCL test, where the NVIDIA Tesla M2090 decisively outperforms the AMD FirePro W5100. The M2090 scores 13075, while the W5100 manages 11888, yielding a 10% delta in favor of NVIDIA. This is not a marginal victory; it is a clear, measurable lead that reflects the fundamental compute-oriented design of the Fermi architecture. The M2090’s score places it within a tight cluster of rivals: it is 0.7% ahead of the GeForce GTX 1660 SUPER (12986), 1% ahead of the RTX 3050 Ti Mobile (12940), and 1.1% ahead of the Radeon RX 580 (12928). Only the GTX 950 edges it out, with the M2090 trailing by 0.9% against that card’s 13189.

The FirePro W5100’s OpenCL score of 11888 is not directly listed among its nearest rivals, but its average benchmark score of 12847 across OpenCL and Vulkan tests tells a broader story. In the Vulkan test, the W5100 scores 13805, which is significantly higher than its OpenCL result, indicating that the GCN architecture benefits substantially from modern API overhead reduction. The W5100’s nearest rivals include the Radeon Pro 455 (12831, 0.1% behind), the GTX 590 (12830, 0.1% behind), and the GTX 670 (12773, 0.6% ahead of the W5100). The only rival that beats the W5100 in average score is the Radeon 740M at 12870, a 0.2% delta.

Interpreting these numbers, the M2090’s 10% lead in OpenCL is substantial for a card that is three years older in release date. However, the W5100’s Vulkan result suggests that in workloads leveraging that API, the AMD card pulls far ahead of its own OpenCL performance. The data does not provide a head-to-head Vulkan comparison, so the M2090’s Vulkan capability is unknown, as its API list shows no Vulkan support. This is a critical differentiator: the M2090 cannot run Vulkan workloads, while the W5100 not only supports Vulkan 1.2.170 but scores 13805 in that test, a figure that would likely surpass the M2090 if it were capable of running the benchmark.

Where Each One Wins

The NVIDIA Tesla M2090 wins in raw OpenCL compute throughput, leveraging its 512 shading units, 64 texture mapping units, and 48 render output units. Its FP32 performance of 1,332.2 GFLOPS is complemented by a texture rate of 41.66 GTexel/s and a pixel rate of 20.83 GPixel/s. These figures indicate a card built for heavy parallel floating-point workloads, where the 384-bit memory bus and 177.4 GB/s bandwidth prevent data starvation during intensive calculations. The M2090’s 6 GB of GDDR5 memory is also a significant advantage over the W5100’s 4 GB, allowing larger datasets to reside on-card without spilling to system memory.

The AMD FirePro W5100 wins in every other measurable category except raw OpenCL score. Its FP32 output is actually higher at 1,428.5 GFLOPS, a 7.2% advantage over the M2090, despite the M2090 winning the benchmark. The W5100 also has a higher texture rate at 44.64 GTexel/s, a 7.2% lead. However, its pixel rate is lower at 14.88 GPixel/s, a 28.6% deficit, due to having only 16 ROPs versus the M2090’s 48. The W5100’s Vulkan support is a decisive feature win, as is its PCIe 3.0 x16 interface, which doubles the bandwidth of the M2090’s PCIe 2.0 x16 slot. The W5100 also has 4x DisplayPort 1.2 outputs, whereas the M2090 has no display outputs at all, making the AMD card a functional workstation GPU for visual output.

In terms of power efficiency, the W5100 is in a different league. Its 50 W TDP is exactly one-fifth of the M2090’s 250 W, and it requires no external power connectors, drawing everything from the PCIe slot. The M2090 needs both a 6-pin and an 8-pin connector, with a suggested power supply of 600 W compared to the W5100’s 250 W. For multi-GPU systems or densely populated servers, this power differential is enormous, allowing the W5100 to fit into far more constrained chassis.

Architecture Differences

The architectural divide between these two cards is stark. The NVIDIA Tesla M2090 uses the GF110 chip on the Fermi 2.0 architecture, fabricated on a 40 nm process at TSMC. This chip contains 3,000 million transistors on a 520 mm² die, resulting in a transistor density of 5.8 million per square millimeter. The Fermi design prioritizes raw compute through 512 shading units arranged in a configuration that maximizes FP32 throughput, though it lacks native Vulkan support and only reaches DirectX 12 (11_0) API level.

The AMD FirePro W5100 uses the Bonaire chip on the GCN 2.0 architecture, also fabricated at TSMC but on a more advanced 28 nm process. This chip contains 2,080 million transistors on a 160 mm² die, yielding a transistor density of 13.0 million per square millimeter, more than double the M2090’s density. The GCN architecture uses 768 shading units, which are organized differently than Fermi’s, and supports DirectX 12 (12_0) as well as Vulkan 1.2.170. The W5100’s memory subsystem is narrower at 128 bit versus the M2090’s 384 bit, which explains its lower bandwidth of 96.00 GB/s versus 177.4 GB/s, despite the W5100’s faster 6 Gbps effective memory clock versus the M2090’s 3.7 Gbps.

The M2090’s memory clock is listed as 924 MHz, translating to 3.7 Gbps effective, while the W5100 runs at 1500 MHz, or 6 Gbps effective. Despite the W5100’s faster memory, the M2090’s three-times-wider bus gives it an 84.8% bandwidth advantage. The M2090’s die size is 520 mm², which is 3.25 times larger than the W5100’s 160 mm², reflecting the older process node and the brute-force approach to compute density. The W5100’s smaller die and newer process allow it to achieve higher FP32 performance (1,428.5 GFLOPS) with far fewer transistors per performance unit.

The production status for both is end-of-life, but their release dates differ significantly: the M2090 launched on July 24, 2011, while the W5100 arrived on March 30, 2014. The M2090’s predecessor is listed as Tesla, with a successor of Tesla Kepler, while the W5100 follows FirePro Terascale and precedes Radeon Pro Polaris. The M2090 is a dual-slot, 248 mm card with no display outputs, designed for compute servers, whereas the W5100 is a single-slot, 173 mm card with four DisplayPort 1.2 outputs, designed for professional workstations.

FAQ

Q: Which card wins in OpenCL performance?

A: The NVIDIA Tesla M2090 wins the Geekbench OpenCL test with a score of 13075, which is 10% higher than the AMD FirePro W5100’s 11888.

Q: Does the FirePro W5100 support Vulkan?

A: Yes, the W5100 supports Vulkan version 1.2.170 and scores 13805 in the Geekbench Vulkan test. The Tesla M2090 has no Vulkan support listed in its API specifications.

Q: How do their power requirements compare?

A: The FirePro W5100 has a 50 W TDP and requires no external power connectors, while the Tesla M2090 has a 250 W TDP and needs one 6-pin and one 8-pin power connector. The suggested PSU for the W5100 is 250 W, versus 600 W for the M2090.

Q: Which card has more memory bandwidth?

A: The Tesla M2090 has 177.4 GB/s of bandwidth from its 384-bit bus, which is 84.8% higher than the FirePro W5100’s 96.00 GB/s from its 128-bit bus.

Q: What is the difference in shading unit counts?

A: The FirePro W5100 has 768 shading units, while the Tesla M2090 has 512. Despite having 50% more shading units, the W5100 scores lower in OpenCL, indicating architectural efficiency differences.

Q: Can the Tesla M2090 drive displays?

A: No, the M2090 lists “No outputs” for display connections. The FirePro W5100 includes 4x DisplayPort 1.2 outputs.

The Verdict

The data supports a clear split based on workload requirements. For pure OpenCL compute tasks where maximum benchmark score is the sole criterion, the NVIDIA Tesla M2090 is the superior choice, delivering a 10% higher score than the FirePro W5100. Its 6 GB memory capacity and 177.4 GB/s bandwidth also make it better suited for large-scale data processing that fits within the frame buffer. The M2090’s 48 ROPs provide a pixel rate advantage of 20.83 GPixel/s versus the W5100’s 14.88 GPixel/s, which matters for certain compute workloads that rely on rasterization operations.

However, for a modern workstation environment, the AMD FirePro W5100 presents a more compelling package. The 28 nm process node enables 1,428.5 GFLOPS of FP32 performance, which is 7.2% higher than the M2090, and the card achieves this with an 80% lower TDP. The W5100’s Vulkan support, PCIe 3.0 interface, and four DisplayPort outputs make it a functional GPU for interactive graphics and compute, whereas the M2090 is a compute-only accelerator with no display capability. The W5100’s average benchmark score of 12847 across OpenCL and Vulkan is within 1.7% of the M2090’s OpenCL-only score, indicating that on modern APIs, the AMD card is at least competitive.

The percentile rankings reinforce this: the M2090 sits at the 53rd percentile versus the W5100’s 52nd, a negligible difference in real-world performance distribution. The M2090’s nearest rival, the GTX 1660 SUPER, is 0.7% behind, while the W5100’s closest competitor, the Radeon Pro 455, is 0.1% behind. Neither card leads its peer group by a meaningful margin. Ultimately, the choice hinges on whether the user needs a low-power, feature-rich workstation card with modern API support and display outputs, or a legacy compute accelerator with a memory bandwidth advantage and a higher OpenCL score. The data does not support the M2090 for general-purpose workstation use, given its lack of outputs and higher power draw, but it remains a valid choice for OpenCL-only compute clusters where its 10% benchmark lead matters more than efficiency.