Home → Card classes
Consumer vs datacenter GPU
Spec sheets make consumer and datacenter cards look like neighbors; the feature list says otherwise. This page compares what actually differs, and which class fits which workload.
The reference consumer card: RTX 4090
| Spec | Value |
|---|---|
| Architecture | Ada (AD102) |
| Memory | 24GB GDDR6X |
| Memory bandwidth | 1008 GB/s |
| Board power | 450W |
| Interface | PCIe 4.0 |
| ECC memory | No |
| NVLink | No |
Source: gpufinder.dev (retrieved Oct 2026).
What datacenter cards add
Ampere/Hopper-class datacenter cards (A100, H100) differ from consumer cards in four ways that matter for ML workloads. These are public hardware facts:
- More memory of a different kind. 80GB of HBM-class memory versus 24GB of GDDR6X — the capacity and bandwidth classes are simply different.
- ECC memory. Error correction matters for long unattended training runs.
- NVLink. High-bandwidth inter-GPU links on SXM boards, which multi-GPU training uses to shard models; consumer multi-GPU is limited to PCIe.
- Datacenter form factor and cooling. SXM modules and server chassis versus consumer PCIe boards in workstations.
Which class fits which workload
| Workload | Usual fit | Why |
|---|---|---|
| Inference up to ~32B Q4 (see LLM table) | 24GB consumer | Fits in 24GB; no multi-GPU needed |
| Image generation (SD/SDXL/Flux) | 12–24GB consumer | Workflows are single-GPU and bursty |
| Video generation (14B+ or HunyuanVideo class) | 40GB+ datacenter | Exceeds consumer memory class |
| 70B Q4 and above | 2×24GB or 48GB+ class | 43GB file needs more than one card holds |
| Multi-node training | Datacenter + NVLink fabric | Interconnect bandwidth dominates |
Decision rule: pick by memory requirement first (the tables on this site), then by whether the job needs ECC and NVLink. If it fits a 24GB consumer card and runs in bursts, the consumer class is the economic choice; if it needs 80GB or multi-GPU bandwidth, that is datacenter territory.