GeForce, Quadro, RTX, and the software behind the hardware
Abstract
GeForce, Quadro, GTX, RTX, and CUDA describe different parts of the NVIDIA ecosystem. This brief explains the product families, architecture generations, practical workstation examples, and software-support checks that matter when evaluating a refurbished PC or GPU upgrade.
Read the whole name before comparing the numbers
A Quadro P620, a GeForce GTX 1080, and an RTX A2000 are all NVIDIA GPUs, but their names describe different product families, generations, and intended uses. Understanding those layers is more useful than assuming that a larger number or a professional label always means a better fit.
This brief follows the same approach as our processor guides: identify the family, understand the generation, then match the exact model to the workload. For the fundamentals, start with TB-011: What Is a GPU?.
CleanSpec examples: K620, P620, and P1000
About these examples: These are configurations and GPU models used in CleanSpec’s refurbishment work, not a live inventory list. The suggested uses below are selection guidance, not new benchmark results or a claim that every application has been tested.
The Quadro names become more useful when connected to actual computers. Our examples include a Dell Precision 3420 with a Quadro K620 and an HP Z2 G5 SFF with a Xeon W-1250 and Quadro P1000. We also use the P620 in our workstation work. These are compact professional graphics options with different limits—not interchangeable versions of a modern RTX card.
| GPU we use | Practical role to evaluate | Important limit |
|---|---|---|
| Quadro K620 | Display output, office work, and modest or older professional graphics workloads in a compatible software environment. | Its 2 GB of memory and older architecture call for checking the application version and driver before assuming CAD or rendering suitability. |
| Quadro P620 | A compact workstation or multiple-display setup where the exact applications fit its capability. | Its 2 GB of memory can constrain larger models or scenes. Four physical mini DisplayPort outputs do not establish performance in a demanding 3D application. |
| Quadro P1000 | An SFF workstation needing a compact card with 4 GB of dedicated graphics memory. | More memory than the P620 gives a larger capacity budget, not a promise of twice the speed or suitability for every 4 GB workload. |
NVIDIA’s desktop Quadro specifications describe the P620 and P1000; PNY’s K620 specification documents that older card. Check the actual board, bracket, ports, and adapters in the machine.
What this means in practice: opening drawings, interacting with a small model, rendering a large scene, and running a local AI model are separate requirements. The Precision’s Xeon processor or the Z2’s workstation name does not erase the graphics card’s memory, software, or support limits. CPU-and-memory testing also does not establish GPU application performance.
See how these configurations compare with our Radeon and Intel examples in Choosing the Right GPU.
1. GeForce, Quadro, and professional RTX
GeForce is NVIDIA's consumer graphics family, commonly associated with gaming and content creation. It spans a wide range of capabilities. A family name is a starting point, not a performance score.
Quadro is the familiar professional graphics name found in many refurbished workstations. Later professional products use names such as NVIDIA RTX, RTX A-series, RTX Ada Generation, and RTX PRO. The market did not disappear when the branding changed. NVIDIA's professional graphics portfolio and legacy Quadro catalog help identify the exact product.
Professional positioning can bring application-specific drivers, certifications, and features. It does not mean every professional card outruns every GeForce card. For CAD, check the software vendor's certified hardware configurations, including the workstation, operating system, GPU, and driver. A badge alone does not establish certification for your application version.
2. Architecture and product family are different things
A GPU architecture describes the underlying design. A product family describes how particular implementations are sold. NVIDIA's architecture history and GeForce comparison provide these useful reference points:
| Architecture | Consumer examples | Practical distinction |
|---|---|---|
| Maxwell | GTX 900 Series | Older architecture; verify current driver and application support |
| Pascal | GTX 10 Series | Also used by Quadro P-series workstation cards |
| Turing | GTX 16 and RTX 20 Series | GTX 16 lacks the dedicated RT and Tensor cores found in RTX 20 |
| Ampere | RTX 30 Series | A later RTX generation; individual models still vary substantially |
| Ada Lovelace | RTX 40 Series | Further changes to graphics, AI, and media capabilities |
| Blackwell | RTX 50 Series | Another architecture generation, not a universal performance guarantee |
This is a desktop-oriented reference, not a complete history of every NVIDIA chip. Data-center architectures and products do not always map directly to consumer cards. Even within one generation, memory capacity, processing resources, power limits, and physical dimensions can differ dramatically.
Naming exception: the Quadro K620 is a Maxwell GPU despite its K prefix. PNY's hardware configuration reference identifies the architecture. Do not turn a useful naming pattern into an absolute rule.
3. What changed between generations?
NVIDIA's GTX 1080 launch documentation provides a historical comparison with the GTX 980:
| Specification | GTX 980 · Maxwell | GTX 1080 · Pascal |
|---|---|---|
| CUDA cores | 2,048 | 2,560 |
| Graphics memory | 4 GB | 8 GB |
| Memory bandwidth | 224 GB/s | 320 GB/s |
| Listed TDP | 165 W | 180 W |
Several parts of the design changed together. These specifications illustrate why generation matters; they are not a CleanSpec benchmark and do not predict an exact speed increase in your application. Core count, clock speed, and memory capacity each describe only part of the GPU.
4. CUDA cores, RT cores, and Tensor cores
CUDA cores are programmable processing resources used for graphics and supported compute workloads. RT cores accelerate particular ray-tracing operations. Tensor cores accelerate supported matrix operations used in AI and other workloads. NVIDIA's RTX product specifications show these as separate resources.
Adding their counts together would not produce a meaningful total. They do different jobs, and software must use the relevant hardware. A GPU with more resources can still be limited by memory, data movement, the CPU, or application design.
Similarly, a teraflop describes a trillion floating-point operations per second under a particular calculation convention. Theoretical throughput is useful context, but it is not an application benchmark. Compare the same precision, operation type, and workload before drawing conclusions.
5. Familiar Quadro cards show the range within a generation
The following published specifications illustrate three Pascal workstation models. See NVIDIA's legacy product specifications for the complete details.
| Model | Memory / power | Physical format / outputs |
|---|---|---|
| Quadro P620 | 2 GB GDDR5 40 W maximum | Low-profile, single-slot 4 × Mini DisplayPort |
| Quadro P1000 | 4 GB GDDR5 47 W maximum | Low-profile, single-slot 4 × Mini DisplayPort |
| Quadro P2200 | 5 GB GDDR5X 75 W maximum | Full-height, single-slot 4 × DisplayPort |
The shared P-series name does not make these interchangeable. A P2200's full-height board may rule it out of a small-form-factor chassis even if its performance is attractive. A P620 or P1000 still needs the correct bracket, sufficient clearance, an appropriate slot, and compatible display adapters.
Memory capacity matters when a model or scene must fit on the GPU. More unused memory does not automatically make a task faster. Check the actual application's memory behavior rather than treating gigabytes as a universal ranking.
6. Video acceleration is a separate capability
NVIDIA uses NVENC for hardware video encoding and NVDEC for hardware decoding. Those are distinct from general 3D performance. A card can support decoding a format without supporting encoding of that same format.
Use NVIDIA's encode/decode support matrix for the exact model and codec. Then check resolution, bit depth, chroma format, and whether your application actually uses the available hardware path. Filters, effects, audio processing, or unsupported formats may still involve other parts of the system.
This matters for streaming, video editing, and media servers. A gaming result does not establish transcoding capability, and a media feature does not establish good gaming performance.
7. CUDA compatibility is more than the NVIDIA name
CUDA is NVIDIA's computing platform. Many rendering and AI applications use it, but each application or framework has supported versions, GPU generations, operating systems, and driver requirements. Check the intended software before buying a card for a CUDA workflow.
OptiX illustrates the importance of version-specific checks. NVIDIA's OptiX 9.1 download page lists Turing-or-newer GPUs and an R590-or-newer driver requirement. An older Quadro can function as a display adapter while failing that particular software requirement. An application using an earlier backend can have different requirements.
For local AI, the model, precision, memory requirement, framework, and software backend all matter. Owning an NVIDIA GPU does not guarantee that a particular model will fit or that a current package supports it. For rendering, test a representative scene, including memory use and completion time, rather than relying only on a synthetic score.
8. Older cards have a support horizon
NVIDIA's Quadro support plan identifies RTX Enterprise Release 580 as the final branch for Maxwell, Pascal, and Volta Quadro GPUs. It specifies one year of full support and critical security updates through October 2028. This is the enterprise/Quadro policy; consumer GeForce products have a separate support plan.
A physically healthy card can therefore remain useful for a defined supported workload without being a sensible choice for every new software release. Driver support, operating-system support, and application support are separate checks. Refurbishment does not reset the hardware's software lifecycle.
9. Apply the information to the complete computer
Start with the job: office displays, an engineering viewport, rendering, gaming, media processing, or local AI. Identify the application and version, then establish the required features and memory. Only then compare exact cards.
For a compact workstation, verify board height, length, thickness, slot power limits, available auxiliary power, airflow, and display connections. A card fitting into the connector does not prove the complete installation is suitable. Use the OEM documentation and evaluate the finished machine under a relevant sustained workload.
A general GPU stress test provides evidence of behavior within that test's scope. It does not establish CAD certification, a game's frame rate, every codec feature, or compatibility with an AI package. CleanSpec's goal is to describe the configuration and the testing clearly, including limitations.
Continue with TB-015: Choosing the Right GPU to apply the NVIDIA, AMD, and Intel guides to practical selection scenarios.