GPU architectures

Six generations on one page. Start with the architecture.

NVIDIA's data-center line spans rack-scale superchips, eight-GPU HGX boards and PCIe cards. Memory, bandwidth and precision support change most between architectures, so pick the generation first, then the form factor. Ranges and counts below are computed from the table.

Compare table

All 23 NVIDIA GPUs. Sort, filter, compare.

Tensor columns follow NVIDIA's convention: switch between sparse (with 2:4 structured sparsity, NVIDIA's headline figure) and dense. A dash (—) means NVIDIA does not publish that value for the GPU, so we leave it blank rather than estimate.

Tensor
Showing 23 of 23
Filters
Architecture
Form factor
Memory type
GPU memory (GB) 16–288
Max power (W) 60–1,400
Ssparse figure as stated by NVIDIA ÷Nper-GPU value derived from NVIDIA's system figure ‡NVIDIA does not state sparse or dense ½dense = ½ sparse, per NVIDIA's footnote

Tensor throughput in TFLOPS (1 PFLOPS = 1,000 TFLOPS). FP8 on Blackwell parts is NVIDIA's combined FP8/FP6 figure. FP64 shows the FP64 Tensor Core figure where NVIDIA lists one. Per-GPU figures for HGX boards and NVL72 racks are NVIDIA's system figures divided by the GPU count; NVIDIA's Blackwell Ultra architecture blog quotes 15 PFLOPS dense and 20 PFLOPS sparse FP4 per GPU, while HGX B300 system figures divided by eight give 13.5 and 18. Rubin CPX is announced, expected at the end of 2026 per NVIDIA. No GPU prices are shown: NVIDIA does not publish list prices and every XcellHost GPU configuration is quoted. Specifications per NVIDIA, retrieved 30 Oct 2026.

Use cases

Match the workload to the GPU.

Ten common AI and visual-computing jobs, the GPUs that suit them and why. GPU chips are pulled live from the table — tap one to add it to your comparison.

Tool · Which GPU?

Five questions. Three GPUs that fit.

The finder scores every current GPU in the table against your answers — memory fit, precision support, throughput and how you want to deploy — and explains each pick. It is a starting point for sizing, not a benchmark.

1Main workload
2Model size 70B parameters
1B8B70B405B1T
3Serving precision
4What matters most
5Deployment
Tool · model memory fit

Will your model fit? Count the GPUs before you buy.

Weights take parameters × bytes per parameter. Serving also needs room for the KV cache, activations and runtime, entered here as an overhead percentage. The overhead default is illustrative — long contexts and large batches need more.

B
Precision · bytes per parameter
Illustrative · edit to match your context length and batch size
Memory needed
168GB
GPUs needed, per GPU in the table

GPUs needed = ⌈memory needed ÷ GPU memory⌉. It assumes the model is split across GPUs (tensor or pipeline parallel) and ignores per-GPU framework overhead; NVL72 rows count GPUs inside one rack. Capacity only — not throughput.

XcellHost GPUs

Picked a GPU? We build, host and run it.

Three NVIDIA GPUs run in the XcellHost fleet today and B300 is open for pre-booking. Every other GPU in the table is sourced and quoted — availability is confirmed at quote.* GPU servers are always priced on quote.

AI FactoryFull-stack GPU platform Liquid-cooled AI data centreFor NVL72-class racks Rent AMD Instinct MI300XPrice on quote Host CPUs: AMD EPYCCompare every EPYC
Generation explainer

Ampere → Hopper → Blackwell → Blackwell Ultra → Rubin. What actually changed.

Platform figures are computed from the table. Performance statements are NVIDIA's own claims, quoted with attribution.

Glossary

Spec-sheet words, in plain English.

SXMSocketed module

A GPU module that mounts flat on a baseboard such as HGX instead of a PCIe slot. It allows higher power and connects every GPU over NVLink — the form factor for B300, B200, H200 and H100 training systems.

NVLinkGPU-to-GPU link

NVIDIA's high-speed link between GPUs. Per NVIDIA, fifth-generation NVLink gives each Blackwell GPU 1.8 TB/s; Hopper SXM gets 900 GB/s. PCIe cards use a bridge for two or four GPUs, or none.

NVL72Rack-scale domain

A liquid-cooled rack with 72 GPUs and 36 Grace CPUs in one NVLink domain, so the rack behaves like one giant GPU. GB200 NVL72 and GB300 NVL72 are built this way.

HBMHigh-bandwidth memory

DRAM stacked beside the GPU die. HBM3e and HBM4 deliver terabytes per second, which sets how fast large models generate tokens. PCIe cards such as L4 and RTX PRO use GDDR6 or GDDR7 instead.

MIGMulti-Instance GPU

Splits one GPU into isolated instances, each with its own memory and compute — for example up to seven on H200 or four 24 GB slices on RTX PRO 6000. Useful for many small models or tenants.

Transformer EngineMixed-precision engine

Hardware and software that pick FP8 (Hopper) or FP4 (Blackwell) per layer while keeping accuracy. Per NVIDIA, it is why newer GPUs serve transformers far faster at the same memory.

FP4 · NVFP44-bit floating point

Four bits per weight — a quarter of FP16. Blackwell and Rubin run it natively; NVFP4 is NVIDIA's scaled FP4 format. A 70B model's weights drop from about 140 GB at FP16 to about 35 GB.

Sparsity2:4 structured sparsity

If two of every four weights are zero, Tensor Cores skip them and double throughput. NVIDIA's headline figures are usually "with sparsity"; dense figures are half unless NVIDIA states otherwise. This page labels both.