RTX 4090, RTX 6000 and Data Center GPU Compared
Key Takeaway
These are not different price points on the same product line but three different engineering approaches: the RTX 4090/5090 is the consumer flagship (no ECC, no NVLink, restricted drivers), the RTX 6000 series is the workstation professional card (ECC, large memory, enterprise drivers), and only the H100/H200/B200 are data centre training cards (HBM, extremely high bandwidth, NVLink interconnect). The cost of choosing wrong is not slowness but a training run crashing halfway through, data corruption or licence violations.
Key Comparison
| Model | Architecture | Memory | ECC | Bandwidth | NVLink | Hong Kong Reference Price |
|---|---|---|---|---|---|---|
| RTX 4090 | Ada | 24GB GDDR6X | No | 1,008 GB/s | No | HK$13,000 to 18,000 |
| RTX 5090 | Blackwell | 32GB GDDR7 | No | 1,792 GB/s | No | HK$18,000 to 28,000 |
| RTX 6000 Ada | Ada | 48GB GDDR6 | Yes | 960 GB/s | Yes | HK$55,000 to 75,000 |
| RTX PRO 6000 Blackwell | Blackwell | 96GB GDDR7 | Yes | 1,792 GB/s | No | HK$130,000 to 190,000 |
| H200 SXM | Hopper | 141GB HBM3e | Yes | 4.8 TB/s | 900 GB/s | Requires a full-system solution |
| B200 | Blackwell | 180 to 192GB HBM3e | Yes | 8 TB/s | 1.8 TB/s | Requires a full-system solution |
Practical Takeaways
- Naming trap: NVIDIA dropped NVLink from the entire Ada-generation professional line. The RTX 6000 Ada (48GB) does not support NVLink, and neither does the RTX PRO 6000 Blackwell (96GB) — both connect over PCIe only. The card that does support NVLink is the older RTX A6000 (Ampere, 48GB), where two cards can be bridged into a single 96GB pool at 112.5 GB/s bidirectional. The naming similarity between "RTX A6000" and "RTX 6000 Ada" is exactly where buyers go wrong. Also note that the "RTX 6000D" is a China-market-only, de-rated SKU (84GB, 448-bit, about 1.57 TB/s) without the PRO designation, so always verify at procurement.
- ECC is the biggest weakness of consumer cards: graphics memory under prolonged high load can suffer bit flips, which ECC detects and corrects automatically. Without ECC, training errors can accumulate and may even reboot the whole machine. Consumer cards are fine for prototype validation; training that runs continuously for several days or more should use professional or data centre cards with ECC.
- NVLink decides multi-GPU efficiency: the RTX 4090/5090 and RTX PRO 5000/6000 Blackwell all lack NVLink support, so all inter-GPU communication goes over PCIe 5.0. If you plan to run two cards in tensor parallelism, the actual speed-up will be far below 2x.
- Licensing compliance: the EULA for GeForce-series drivers places restrictions on "deployment in a data centre". An enterprise that wants to offer GPU compute services in its server room in a compliant way should choose the RTX PRO series (enterprise drivers, ISV certification, longer support cycles) or the data centre series; this matters especially for finance, healthcare and government projects.
- Deployment details: for a 600W-class card, a power supply of 1,300W or more is recommended, with 30% headroom for peaks; 12VHPWR / 12V-2x6 connectors must be pushed fully home until they click, since a partial insertion can overheat and melt; heavy cards need a support bracket, as sagging over time damages the PCIe slot or the PCB.
Tip: Common misconceptions: "two 4090s = one A100" does not hold - the headline compute may be close, but they lack ECC, they lack NVLink, and bandwidth is only about half of an A100's. "The more VRAM the better" is also incomplete - for the same 48GB, the gap between GDDR7 (1,344 GB/s) and HBM (3.35 TB/s) can exceed two times.
