NVIDIA GeForce GTX 670M vs NVIDIA RTX PRO 6000 Blackwell Server Comparison

The NVIDIA GeForce GTX 670M and the NVIDIA RTX PRO 6000 Blackwell Server represent two extremes of NVIDIA’s mobile and server GPU lineage, separated by over a decade of architectural evolution. The GTX 670M, a Fermi 2.0 part from 2012, was a high-end mobile graphics solution for gaming laptops, while the RTX PRO 6000 Blackwell Server is a 2025 data-center and professional visualization powerhouse built on the Blackwell 2.0 architecture. The data shows a stark contrast in design goals: the GTX 670M was engineered for power-efficient 3D rendering in portable devices, whereas the RTX PRO 6000 is designed for massive parallel compute, ray tracing, and AI workloads. Benchmark results, however, present an unusual twist, as the two GPUs occupy similar percentile positions in their respective test suites, despite their generational and technical gulfs.

Where Each One Wins

The use-case split between these two GPUs is defined by their respective benchmark contexts and target platforms. The GTX 670M’s single benchmark result comes from Geekbench OpenCL, where it scores 6,513 points. This test measures general-purpose compute performance across a variety of workloads, including image processing and computational photography. The GTX 670M’s score places it at the 38th percentile of all GPUs, indicating that it performs competitively within its older mobile GPU cohort, edging out the NVIDIA GeForce GT 555M by 0.3% and the NVIDIA Quadro M5000M by 0.5%. Its strengths lie in legacy DirectX 11-era workloads and OpenCL compute tasks that do not require modern features like hardware ray tracing or tensor cores. The GTX 670M wins in scenarios where legacy software compatibility and low power draw (75 W TDP) are paramount, such as in older mobile workstations or embedded MXM-based systems.

Conversely, the RTX PRO 6000 Blackwell Server wins decisively in any modern, high-intensity compute or rendering scenario. Its benchmark result comes from 3DMark Steel Nomad DX12, a demanding DirectX 12 Ultimate test that stresses raw rasterization, mesh shaders, and asynchronous compute. Scoring 5,996 points, it sits at the 34th percentile of all GPUs, which seems low until one considers the extreme variance in the benchmark database, where its nearest rivals include the NVIDIA GeForce GTX 770M (0.1% behind) and the AMD Radeon RX 6400 (0.1% ahead). The RTX PRO 6000’s real wins are not reflected in its percentile but in its architectural capabilities; it is the only one of the two with 188 RT cores and 752 tensor cores, making it the sole choice for ray-traced rendering, AI inference, and neural network training. The data shows that the RTX PRO 6000 wins in any workload that leverages DirectX 12 Ultimate features, Vulkan 1.4, or CUDA-accelerated compute, while the GTX 670M wins in purely legacy OpenCL and DirectX 11 environments.

Architecture Differences

The architectural divide between these two GPUs is vast, rooted in a 13-year gap in design philosophy. The GTX 670M is built on the GF114 chip using the Fermi 2.0 architecture, fabricated on a 40 nm process at TSMC. It packs 1,950 million transistors into a 332 mm² die, resulting in a transistor density of 5.9M per mm². Fermi 2.0 introduced improvements over the original Fermi architecture, including better tessellation performance and enhanced CUDA compute capabilities, but it lacks dedicated hardware for ray tracing or tensor operations. The GTX 670M features 336 shading units, 56 texture mapping units (TMUs), and 24 raster output units (ROPs). Its memory subsystem consists of 1,536 MB of GDDR5 on a 192-bit bus, delivering 72.00 GB/s of bandwidth. The absence of RT and tensor cores means all compute is handled by the uniform shading units, which peak at 803.7 GFLOPS FP32.

The RTX PRO 6000 Blackwell Server, in contrast, utilizes the GB202 chip based on the Blackwell 2.0 architecture, manufactured on a 5 nm process, also at TSMC. This chip is a monolithic behemoth with 92,200 million transistors on a 750 mm² die, achieving a transistor density of 122.9M per mm²—roughly 20 times denser than the GTX 670M. Blackwell 2.0 is a complete departure, integrating 24,064 shading units, 752 TMUs, and 192 ROPs. Critically, it adds 188 RT cores and 752 tensor cores, enabling hardware-accelerated ray tracing and matrix math for AI workloads. The memory system is equally advanced: 96 GB of GDDR7 on a 512-bit bus provides 1.79 TB/s of bandwidth. The FP32 compute throughput is 126.0 TFLOPS, and FP16 is supported at a 1:1 ratio of 126.0 TFLOPS, a feature absent on the GTX 670M which has no FP16 capability listed. The process node shrink from 40 nm to 5 nm is the fundamental enabler of these massive resource increases.

Head-to-Head Benchmarks

Direct head-to-head benchmark comparisons are not available in the data, as the two GPUs were tested with entirely different suites—Geekbench OpenCL for the GTX 670M and 3DMark Steel Nomad DX12 for the RTX PRO 6000. However, the relative scores and nearest rival deltas provide a basis for interpretation. The GTX 670M scores 6,513 in Geekbench OpenCL, which is 0.3% higher than the GT 555M and 0.5% higher than the Quadro M5000M, but 0.6% lower than the Intel UHD Graphics P750. This indicates that within its peer group of older mobile GPUs, the GTX 670M is a mid-to-upper tier performer, slightly outpacing dedicated NVIDIA mobile parts but lagging a modern integrated GPU in that specific OpenCL test.

The RTX PRO 6000 scores 5,996 in 3DMark Steel Nomad DX12, a test that is significantly more demanding than Geekbench OpenCL. Its nearest rivals are the GTX 770M (0.1% behind), RX 6400 (0.1% behind), FirePro W4100 (0.2% behind), and Quadro K4000M (0.2% behind). The tiny delta percentages (all within 0.2%) suggest that in this specific DX12 benchmark, the RTX PRO 6000’s massive compute resources are not fully utilized due to the test’s nature or driver overhead, placing it in a statistical tie with much older and less powerful GPUs. This is a critical insight: the RTX PRO 6000’s 126.0 TFLOPS FP32 and 1.79 TB/s bandwidth do not translate to a dominant score in Steel Nomad, likely because the benchmark is more rasterization-bound and does not stress RT or tensor cores. The GTX 670M’s 803.7 GFLOPS FP32 is minuscule by comparison, yet its Geekbench score is higher in absolute terms, underscoring that cross-suite comparisons are meaningless. The data shows that the RTX PRO 6000 wins on raw compute headroom by a factor of roughly 157x in FP32 (126.0 TFLOPS vs 803.7 GFLOPS), but the benchmark scores do not reflect this due to differing test methodologies.

Specification Differences

The specification table reveals stark contrasts across every measurable metric. The process node differs dramatically: 40 nm for the GTX 670M versus 5 nm for the RTX PRO 6000, with transistor counts of 1,950 million versus 92,200 million, respectively. Die size grows from 332 mm² to 750 mm², while transistor density jumps from 5.9M / mm² to 122.9M / mm². Clock speeds are also divergent: the GTX 670M has no base or boost clock listed, while the RTX PRO 6000 runs at 1,590 MHz base and 2,617 MHz boost. Memory clocks are 750 MHz (3 Gbps effective) for the GTX 670M versus 1,750 MHz (28 Gbps effective) for the RTX PRO 6000. Memory capacity is 1,536 MB GDDR5 versus 96 GB GDDR7, with bus widths of 192-bit versus 512-bit, yielding bandwidths of 72.00 GB/s versus 1.79 TB/s—a 24.9x difference in favor of the RTX PRO 6000.

Compute resources differ by orders of magnitude: 336 shading units versus 24,064, 56 TMUs versus 752, and 24 ROPs versus 192. The RTX PRO 6000 adds 188 RT cores and 752 tensor cores, features entirely absent from the GTX 670M. Pixel rate improves from 8.372 GPixel/s to 502.5 GPixel/s, and texture rate from 33.49 GTexel/s to 1,968.0 GTexel/s. FP32 compute rises from 803.7 GFLOPS to 126.0 TFLOPS, with FP16 only available on the RTX PRO 6000 at 126.0 TFLOPS. Power and physical specs also diverge: the GTX 670M consumes 75 W and uses an MXM Module slot with no power connectors, while the RTX PRO 6000 consumes 600 W, is a dual-slot card with a single 16-pin connector, and requires a 1,000 W suggested PSU. The bus interface changes from MXM-B (3.0) to PCIe 5.0 x16. Display outputs are "Portable Device Dependent" on the GTX 670M versus 4x DisplayPort 2.1b on the RTX PRO 6000. API support differs as well: DirectX 12 (11_0) versus 12 Ultimate (12_2), and Vulkan is null on the older card but 1.4 on the newer one. Production status and release dates reinforce the generational gap: the GTX 670M is end-of-life from March 2012, while the RTX PRO 6000 is active from March 2025.

FAQ

Q: Which GPU has higher FP32 compute performance?

A: The RTX PRO 6000 Blackwell Server has a FP32 throughput of 126.0 TFLOPS, compared to the GTX 670M’s 803.7 GFLOPS, making the newer card roughly 157 times faster in raw single-precision compute.

Q: Does either GPU support hardware ray tracing?

A: Only the RTX PRO 6000 Blackwell Server includes 188 dedicated RT cores. The GTX 670M has no RT cores listed, meaning it cannot perform hardware-accelerated ray tracing.

Q: How do their memory bandwidths compare?

A: The RTX PRO 6000 offers 1.79 TB/s of bandwidth using 96 GB of GDDR7 on a 512-bit bus, while the GTX 670M provides 72.00 GB/s with 1,536 MB of GDDR5 on a 192-bit bus. The RTX PRO 6000 has approximately 24.9 times the bandwidth.

Q: What are the power consumption requirements for each?

A: The GTX 670M has a 75 W TDP and uses an MXM module with no power connectors. The RTX PRO 6000 has a 600 W TDP, requires a single 16-pin power connector, and lists a 1,000 W suggested PSU.

Q: Which API versions do they support?

A: The GTX 670M supports DirectX 12 (11_0) and OpenGL 4.6, with no Vulkan support listed. The RTX PRO 6000 supports DirectX 12 Ultimate (12_2), OpenGL 4.6, and Vulkan 1.4.

Q: How do their benchmark scores compare to their nearest rivals?

A: The GTX 670M scores 6,513 in Geekbench OpenCL, which is 0.3% above the GT 555M and 0.5% above the Quadro M5000M. The RTX PRO 6000 scores 5,996 in 3DMark Steel Nomad DX12, which is 0.1% below the GTX 770M and 0.2% above the FirePro W4100.