How to read GPU names
RTX 4090, GTX 1080 Ti, H100, RX 7900 XTX. A name can identify a family, generation or variant. The rules change across brands and sometimes within the same brand.
Use the full name to find the exact specification, then compare matched models and workloads. Names alone do not establish performance, memory capacity or software compatibility.
NVIDIA GeForce
Read the family, series, model and full variant together. A familiar number can appear on substantially different desktop and laptop products.
- RTX / GTX
RTX is NVIDIA’s ray-tracing and AI graphics-platform brand. GTX is an older GeForce family label; the official comparison lists dedicated RT and Tensor cores for RTX generations rather than GTX 10/16.
Treat these as product-family labels. The sources do not provide a universal letter-by-letter expansion; Some GTX models gained basic DXR support through programmable shaders; check the supported card, driver and application.
- 50 and 70 in RTX 5070
Within recent GeForce desktop names, 50 places the product in the 50 Series; 70 identifies its model position within that series.
The digits are not VRAM capacity or a cross-brand performance score. Series and architecture naming have changed over time.
- Ti
Ti has a historical Titanium connection: NVIDIA called the GTX 650 Ti a “titanium” edition in its 2012 launch release. In current names, it distinguishes a model variant.
It does not promise titanium construction, a fixed speed increase or more memory. The RTX 4070 and 4070 Ti both have 12 GB in the official family table.
- SUPER / Ti SUPER
These mark separate GeForce variants. The RTX 4070 family includes the original, SUPER, Ti and Ti SUPER, with different specifications.
There is no fixed SUPER multiplier. Preserve the entire suffix when matching specifications or benchmark results.
- Laptop GPU / Max-Q
“RTX 4090 Laptop GPU” is a separately specified product from desktop RTX 4090. In the cited GeForce laptop range, Max-Q describes a suite of technologies for power, thermals and system performance.
Keep the laptop model and its system configuration. Max-Q is not a universal laptop-only suffix: NVIDIA also sells a Max-Q Workstation Edition in its professional desktop range.
NVIDIA workstation names
Professional products follow their own sequence. GeForce and workstation names with similar digits are not interchangeable.
- Quadro RTX / RTX A / RTX PRO
Quadro RTX 6000, RTX A6000, RTX 6000 Ada and RTX PRO 6000 Blackwell are distinct products. The full name separates the family and architecture.
Quadro RTX 6000 is Turing; RTX A6000 is Ampere. The A in A6000 must not be expanded to Ada.
- Ada / Blackwell / edition qualifiers
Architecture names and qualifiers such as Workstation Edition, Max-Q Workstation Edition and SFF distinguish products and integration requirements.
Keep every qualifier. RTX PRO 6000 Blackwell’s two workstation editions have different power/form-factor specifications; SFF indicates a compact product form factor, with exact dimensions still worth checking.
NVIDIA data-center names
A GPU, a module, a bridged pair and a complete server can share part of a name. Identify the physical scope before comparing capacity or performance.
- H in H100 / H200
H100 and H200 are Hopper-based products. Hopper is the architecture named for Grace Hopper; the H is associated with that family in these examples.
This is a documented architecture association, not a universal character-by-character naming rule. H200 remains Hopper-based, so a larger number does not necessarily mean a new architecture.
- A100 / B200
A100 belongs to Ampere; B200 belongs to Blackwell. These are product-model names within their documented platforms.
The numbers are not memory amounts. A100 has separate 40 GB and 80 GB variants; retain memory and form factor rather than comparing the short name alone.
- SXM / PCIe
These distinguish form factors and system integration. H100 SXM and PCIe-based H100 products have different specifications and installation requirements.
Check exact memory, power, cooling and interconnect. Neither label turns an accelerator module into an ordinary interchangeable desktop card.
- NVL
H100 NVL uses PCIe cards with NVLink bridges. Its product brief specifies 94 GB on one card; the 188 GB description refers to paired capacity.
Keep one card, the bridged pair and the server distinct. NVL does not imply one universal GPU count across NVIDIA products.
- GB200 / NVL72 / DGX
GB200 combines a Grace CPU with two B200 GPUs. GB200 NVL72 is a rack-scale system; DGX B200 is an eight-GPU system.
A system score describes that system. Dividing it by GPU count does not establish a measured single-GPU result.
AMD Radeon and Instinct
Radeon RX consumer graphics, Radeon PRO workstations and Instinct data-center accelerators are separate families, even when model numbers look similar.
- Radeon RX / series digits
RX 7900 GRE belongs to Radeon RX 7000 and RDNA 3; RX 9070 and 9070 XT belong to RX 9000 and RDNA 4.
Product-series digits are not the RDNA architecture version. Use the stated generation’s pattern rather than a universal Radeon number parser.
- XT / XTX / GRE
These identify distinct variants. The RX 7900 XT, XTX and GRE have their own compute-unit and memory specifications.
They are useful variant labels, not fixed speed increments. The cited official pages do not establish universal acronym expansions for XT, XTX or GRE.
- M / S / M XT
In the RX 7000 laptop range, M names the main mobile series and S is positioned for thin-and-light, lower-power laptops. M XT is a distinct mobile variant.
Keep mobile status and the laptop’s configuration. Neither suffix supplies a universal wattage or makes a same-number desktop result comparable.
- Radeon PRO W
Radeon PRO W7900 belongs to the workstation PRO W7000 family, separate from consumer RX 7900 cards.
Similar digits do not establish the same board, memory or driver/application support.
- Instinct MI
Instinct is AMD’s data-center accelerator family. Its launch was framed around machine intelligence; later products address AI and high-performance computing.
That historical association does not establish a literal official expansion of the letters MI. Use the exact model and system specifications.
- MI300A / MI300X
MI300A is an accelerated processing unit combining CPU and GPU cores with unified memory. MI300X is a GPU accelerator.
The A/X distinction changes the system and memory architecture. It is not a universal decoding formula for every Instinct generation.
Intel Arc and Xe
Arc identifies products; Xe names architectures. Keep discrete desktop, mobile, professional and integrated products separate.
- Arc A-series / B580 / B570
The discrete A-series was code-named Alchemist and uses Xe-HPG. The B580/B570 desktop B-series was code-named Battlemage and uses Xe2.
These are scoped product-family mappings. The model digits are not memory capacity or a cross-generation performance score.
- Integrated Arc B390
Arc branding also appears on integrated GPUs. Intel explicitly identifies Arc B390 graphics as Xe3, unlike the Xe2 B580/B570 desktop examples.
The first letter B therefore cannot universally decode to Battlemage or Xe2 across all Arc products.
- Arc 3 / 5 / 7; M; memory
Intel introduced Arc 3, 5 and 7 mobile performance groups. Mobile models such as A770M are separately listed; desktop A770 itself has 8 GB and 16 GB versions.
Preserve the full model, mobile suffix and memory amount. A770 alone is not a complete tested configuration.
- Arc Pro / Xe-HPG / Xe2
Arc Pro is positioned for workstation applications and professional drivers. Xe-HPG and Xe2 describe architecture; Intel expands HPG as high-performance graphics.
Architecture and professional branding do not guarantee every OS, software API or application version is supported.
Apple’s integrated GPU names
Apple describes the GPU as part of an M-series system-on-a-chip rather than a separate PCIe graphics card.
- M generation / Pro / Max
Names such as M4 Pro and M4 Max identify chip families and tiers that include CPU, GPU and other engines.
Pro or Max alone does not identify the GPU configuration. Record the exact chip and GPU-core count.
- Ultra
For M3 Ultra, Apple documents UltraFusion joining two M3 Max dies and presenting them as one chip, with up to 80 GPU cores.
Scope that construction to the documented product. The word Ultra does not promise identical packaging across all generations or twice the measured speed.
- GPU cores / unified memory
The same chip tier can have multiple GPU-core options. Unified memory is a configurable shared chip/system resource, so include its capacity with the chip and core count.
Shared unified memory is not dedicated VRAM. GPU-core counts across Apple, NVIDIA, AMD and Intel are not interchangeable performance units.