GPU Comparison

The AMD Radeon 660M and NVIDIA Tesla M2090 represent two very different approaches to GPU design, separated by over a decade of architectural evolution. The 660M is a modern integrated graphics processor built for mobile efficiency, while the M2090 is a legacy compute accelerator designed for high-throughput server workloads. Benchmark data shows the M2090 holds a narrow lead in the one shared test, but the 660M counters with a significantly more advanced feature set and dramatically lower power demands. This analysis breaks down the data to determine which card wins in specific scenarios.

FAQ

Q: Which GPU has the higher average benchmark score?

A: The AMD Radeon 660M posts an average benchmark score of 13812, while the NVIDIA Tesla M2090 scores 13075. The 660M sits in the 55th percentile of all GPUs, compared to the M2090's 53rd percentile.

Q: How do the two compare in the Geekbench OpenCL test?

A: The NVIDIA Tesla M2090 wins the head-to-head Geekbench OpenCL test with a score of 13075 versus the AMD Radeon 660M's 12876. This represents a 1.5% performance advantage for the M2090.

Q: What are the key architectural differences?

A: The 660M uses RDNA 2.0 architecture on a 6 nm process with 13,100 million transistors, while the M2090 uses Fermi 2.0 on a 40 nm process with only 3,000 million transistors. The 660M also features 6 ray tracing cores, which the M2090 lacks entirely.

Q: How does memory configuration differ?

A: The AMD Radeon 660M uses system-shared memory with bandwidth described as "System Dependent," whereas the NVIDIA Tesla M2090 has 6 GB of dedicated GDDR5 memory on a 384-bit bus, providing 177.4 GB/s of bandwidth.

Q: Which card has higher power requirements?

A: The NVIDIA Tesla M2090 has a 250 W TDP and requires a 600 W suggested power supply, while the AMD Radeon 660M has a 40 W TDP and uses no power connectors, being an integrated graphics processor.

Q: What API support does each card provide?

A: The AMD Radeon 660M supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4. The NVIDIA Tesla M2090 supports DirectX 12 (11_0) and OpenGL 4.6, but has no Vulkan support.

Architecture Differences

The AMD Radeon 660M is built on TSMC's 6 nm process node, packing 13,100 million transistors into a 208 mm² die. This yields a transistor density of 63.0 million transistors per square millimeter. In contrast, the NVIDIA Tesla M2090 uses TSMC's older 40 nm process, with 3,000 million transistors spread across a much larger 520 mm² die, resulting in a density of just 5.8 million transistors per square millimeter. The 660M's RDNA 2.0 architecture represents a modern graphics design with integrated ray tracing capabilities, featuring 6 dedicated RT cores. The M2090's Fermi 2.0 architecture, by comparison, has no ray tracing hardware whatsoever.

The 660M is an integrated graphics processor (IGP) from the Rembrandt mobile generation, meaning it shares memory with the system rather than having its own dedicated VRAM. Its memory bus width and type are both listed as "System Shared," with bandwidth dependent on the host system's memory configuration. The M2090, on the other hand, is a discrete dual-slot card with its own 6 GB of GDDR5 memory on a 384-bit interface, delivering a fixed 177.4 GB/s of bandwidth. The 660M's memory clock is also system-shared, while the M2090 runs its memory at 924 MHz (3.7 Gbps effective).

The transistor count difference is stark: the 660M has over four times as many transistors as the M2090 (13,100 million versus 3,000 million), yet it consumes far less power. This is primarily due to the manufacturing process advantage and architectural efficiency. The 660M's shading unit count is 384, compared to the M2090's 512, and the 660M has 24 texture mapping units and 16 render output units, while the M2090 has 64 and 48 respectively. Despite having fewer shading units, the 660M achieves a higher pixel rate of 30.40 GPixel/s versus the M2090's 20.83 GPixel/s, and a higher texture rate of 45.60 GTexel/s versus 41.66 GTexel/s.

Head-to-Head Benchmarks

The only direct benchmark comparison available between these two GPUs is the Geekbench OpenCL test. In this test, the NVIDIA Tesla M2090 scores 13075, edging out the AMD Radeon 660M's score of 12876. This gives the M2090 a 1.5% lead, represented by a delta percentage of -1.5 from the 660M's perspective. While this is a win for the older card, the margin is extremely narrow and falls well within typical run-to-run variance for benchmark testing.

Looking at the broader benchmark landscape, the 660M has two recorded benchmark scores: 12876 in Geekbench OpenCL and 14748 in Geekbench Vulkan. The M2090 only has a single Geekbench OpenCL score of 13075, with no Vulkan result. This means the 660M's average benchmark score of 13812 is bolstered by its strong Vulkan performance, which is not tested on the M2090. The M2090's average of 13075 is based solely on its OpenCL result.

The 660M's nearest rivals in benchmark data include the NVIDIA RTX A2000 Mobile at 13821 (-0.1% delta), the AMD Radeon RX 570X at 13871 (-0.4%), and the AMD Radeon RX 7900 XT at 13745 (+0.5%). The M2090's nearest rivals include the NVIDIA GeForce GTX 1660 SUPER at 12986 (+0.7%), the NVIDIA GeForce GTX 950 at 13189 (-0.9%), and the NVIDIA GeForce RTX 3050 Ti Mobile at 12940 (+1%). These comparisons show both cards performing in a similar mid-range tier, with the 660M's percentile ranking of 55 slightly above the M2090's 53.

Specification Differences

The two GPUs differ across nearly every specification category. The AMD Radeon 660M uses a 6 nm process node, while the NVIDIA Tesla M2090 uses 40 nm. Transistor counts are 13,100 million versus 3,000 million, and die sizes are 208 mm² versus 520 mm². The 660M has a base clock of 1500 MHz and boost clock of 1900 MHz, while the M2090 lists no base or boost clock values, its memory clock is specified at 924 MHz (3.7 Gbps effective), but core clocks are absent from the data.

Memory configurations are fundamentally different: the 660M uses system-shared memory with system-dependent bandwidth, while the M2090 has 6 GB of GDDR5 with a 384-bit bus and 177.4 GB/s bandwidth. The 660M has 384 shading units, 24 TMUs, and 16 ROPs, versus the M2090's 512 shading units, 64 TMUs, and 48 ROPs. The 660M includes 6 RT cores, which the M2090 does not have at all.

Pixel rate favors the 660M at 30.40 GPixel/s versus 20.83 GPixel/s for the M2090. Texture rate also favors the 660M at 45.60 GTexel/s versus 41.66 GTexel/s. FP32 performance is higher on the 660M at 1,459.2 GFLOPS compared to 1,332.2 GFLOPS on the M2090. The 660M also supports FP16 at 2.918 TFLOPS (2:1), while the M2090 has no FP16 data. Power consumption is dramatically different: 40 W TDP for the 660M versus 250 W for the M2090.

The 660M is an IGP with no power connectors and a PCIe 4.0 x8 interface, while the M2090 is a dual-slot card requiring 1x 6-pin and 1x 8-pin power connectors, a 600 W suggested PSU, and uses PCIe 2.0 x16. The M2090 measures 248 mm (9.8 inches) in length and has no display outputs; the 660M's display outputs are portable device dependent. API support also differs: the 660M supports DirectX 12 Ultimate (12_2) and Vulkan 1.4, while the M2090 only supports DirectX 12 (11_0) and has no Vulkan support. Release dates are 2022-01-03 for the 660M and 2011-07-24 for the M2090.

Where Each One Wins

The AMD Radeon 660M wins decisively in architectural modernity and efficiency. Its 6 nm process versus 40 nm means it delivers higher performance per watt by a substantial margin, the data shows 1,459.2 GFLOPS of FP32 performance at 40 W, versus 1,332.2 GFLOPS at 250 W for the M2090. The 660M also offers ray tracing support through its 6 RT cores, a feature completely absent from the M2090. For integrated graphics users in portable devices, the 660M's system-shared memory and lack of power connectors make it the only viable option in mobile form factors.

The 660M wins on raw compute metrics as well: higher pixel rate (30.40 vs 20.83 GPixel/s), higher texture rate (45.60 vs 41.66 GTexel/s), higher FP32 throughput (1,459.2 vs 1,332.2 GFLOPS), and the addition of FP16 support. It also supports modern APIs including Vulkan 1.4 and DirectX 12 Ultimate, which the M2090 cannot match. The 660M's 55th percentile ranking and higher average benchmark score (13812 vs 13075) further reinforce its overall performance advantage.

The NVIDIA Tesla M2090 wins in the specific area of dedicated memory. Its 6 GB of GDDR5 on a 384-bit bus with 177.4 GB/s bandwidth is a fixed, dedicated resource that does not compete with the system for memory access. For workloads that require consistent, high-bandwidth memory access without system interference, this is a clear advantage. The M2090 also holds the direct head-to-head OpenCL benchmark win at 13075 versus 12876, albeit by a narrow 1.5% margin. Its higher shading unit count (512 vs 384) and greater TMU/ROP counts (64/48 vs 24/16) suggest it may have advantages in certain compute-heavy workloads that do not rely on the architectural efficiencies of the newer GPU.

The Verdict

The data presents a clear split: the AMD Radeon 660M is the superior all-around GPU for modern workloads, while the NVIDIA Tesla M2090 retains niche advantages for legacy compute tasks. The 660M's higher average benchmark score (13812 vs 13075), higher percentile ranking (55 vs 53), and superior architectural features, including ray tracing, Vulkan support, and dramatically lower power consumption, make it the better choice for anyone building or using a system where efficiency and modern API support matter. Its FP32 performance advantage (1,459.2 vs 1,332.2 GFLOPS) and higher pixel/texture rates demonstrate that the newer architecture achieves more with fewer shading units.

However, the NVIDIA Tesla M2090 is not without merit. It wins the only direct head-to-head benchmark available, and its dedicated 6 GB GDDR5 memory with 177.4 GB/s bandwidth is a significant asset for applications that require predictable, high-bandwidth memory access. For users running legacy compute workloads that do not benefit from newer APIs or ray tracing, and where power consumption is not a primary concern, the M2090 remains a functional option, its dual-slot form factor, 250 W TDP, and 600 W PSU requirement are acceptable for server environments. The M2090's 512 shading units and 64 TMUs provide raw parallel resources that some compute tasks may still utilize effectively.

Ultimately, the choice depends on the use case. The 660M is the modern winner for mobile devices, general computing, and any workload that can leverage its newer feature set. The M2090 is a legacy compute card that still holds a narrow benchmark edge in OpenCL and offers dedicated memory bandwidth, but its architectural age shows in process node, power efficiency, and API support. For most users, the 660M is the clear recommendation based on the benchmark data and specification advantages.