新川サーバー | エンタープライズ サーバー ハードウェア サプライヤー
The Nameplate Says 2400W. The Server Draws 400W. You Are Paying for 2000W of Inefficiency.
Every data center operator has seen it: a server spec’d with dual 2400W power supplies that idles at 350W. The 2400W rating is the worst-case scenario – both CPUs at 100% load, all DIMM slots populated, every NVMe drive writing simultaneously, and every PCIe slot populated. In reality, that server spends 95% of its life at 15-25% of its nameplate power rating. At 15% load, a Titanium PSU is 90% 効率的. At 50% load, it is 96% 効率的. At 15% load, an 800W PSU is 93% 効率的 – because the load is a larger percentage of its rating. Matching the PSU to the actual power draw, not the worst-case peak, saves you 3-6% of your annual electricity bill – and on a fleet of 100 サーバー, that is tens of thousands of dollars.
80 PLUS Ratings – The 3% Difference That Compounds for 5 年
| 80 PLUS Rating | Efficiency at 20% Load | Efficiency at 50% Load | Efficiency at 100% Load | 5-Year Cost Difference (vs Gold, 100 サーバー, $0.12/kWh, 500W avg) |
|---|---|---|---|---|
| Gold | 87% | 90% | 87% | Baseline ($0) |
| 白金 | 90% | 92% | 89% | -$26,280 (saved vs Gold) |
| チタン | 92% | 94% | 90% | -$42,048 (saved vs Gold) |
The jump from Gold to Platinum saves roughly $263 per server per year. The jump from Platinum to Titanium saves another $158. On a single server, neither number justifies a procurement fight. On a 100-server fleet over 5 年, Titanium saves $42,000 – enough to buy three additional R760 servers. The efficiency argument is a fleet argument.
Redundancy Models – N+1, 1+1, 2+2, and When Each One Applies
| モデル | How It Works | Fault Tolerance | Best For |
|---|---|---|---|
| N+1 (Grid Redundancy) | All PSUs share load. One extra PSU in reserve. E.g., 3+1 = 4 psus, 3 needed, 1 spare. | Survives loss of 1 PSU or 1 power grid. | 4-socket servers (R860, R960, 5885H V7). |
| 1+1 (Active-Active) | 2 psus, both active, each supplying ~50% load. | Survives loss of 1 電源ユニット. Most common configuration. | All dual-socket servers (R660, R760, R770, 2288H V7). |
| 2+2 (Dual Grid) | 4 psus: 2 on Grid A, 2 on Grid B. Each grid independently redundant. | Survives loss of 1 PSU per grid or loss of 1 entire power grid. | GPU servers drawing 3kW+ (XE9680, G5500 V7). |
| 1+0 (Non-Redundant) | 1 電源ユニット. No redundancy. | なし. Single point of failure. | Lab, dev, test only. Never in production. |
Server-by-Server PSU Configuration Guide
| サーバ | Typical Draw (アイドル状態 / Peak) | Recommended PSU | Redundancy | Efficiency at Typical Load |
|---|---|---|---|---|
| Dell R470 (1-socket, no GPU) | 150W / 500W | 2x 600W or 2x 800W | 1+1 | Platinum at 30-60% load |
| Dell R660 (2-socket, no GPU) | 200W / 700W | 2x 800W or 2x 1100W | 1+1 | Platinum at 25-65% load |
| Dell R760 (2-socket, 2x GPU) | 250W / 1,200W | 2x 1400W | 1+1 | Titanium at 20-85% load |
| Dell R770 (2-socket, 2x 450W GPU) | 300W / 1,800W | 2x 2400W | 1+1 | チタン (only option at this wattage) |
| Dell R860 (4-socket, no GPU) | 450W / 2,000W | 4x 1100W or 2+2 1400W | 2+2 (dual grid) また 3+1 | チタン |
| Dell XE9680 (2-socket, 8x H100) | 800W / 8,500W | 4x 2800W Titanium | 2+2 (dual grid, 2 per grid) | チタン (only option) |
| xFusion G5500 V7 (2-socket, 10x GPU) | 700W / 10,000W | 4x 3000W Titanium | 2+2 また 3+1 | チタン |
| xFusion 1288H/2288H V7 (2-socket) | 200W / 800W | 2x 900W or 2x 1500W | 1+1 | 白金 |
The 80% PSU Utilization Myth
There is a persistent belief in data centers that a PSU should not be loaded beyond 80% of its rated capacity. This myth originated from early 2000s PSU designs that experienced accelerated capacitor aging above 80% load. Modern 80 PLUS Titanium and Platinum PSUs from Dell and xFusion are rated for 100% continuous load at the specified temperature and altitude. The warranty covers 100% load, 24/7, for 5 年. Undersizing a PSU because of the 80% myth forces the PSU to operate at a lower efficiency point (10-15% load instead of 40-50%) and wastes energy.
The counterpoint: a PSU running at 100% load has no headroom for a transient spike – a CPU turbo boost, a GPU compute burst, or simultaneous drive spin-up. A 100ms spike to 110% load will cause a properly designed PSU to engage over-current protection and shut down. The real rule is not “80% 最大” – it is “the PSU’s rated capacity must exceed the server’s worst-case transient load.” For most general-purpose servers, nameplate PSU rating at 120-130% of measured peak load is comfortable. For GPU servers, measure worst-case transient power with both GPUs at 100% Tensor Core utilization before sizing the PSU.
Three Questions to Ask Before Ordering Power Supplies
- What is the electricity rate at my colocation facility? At $0.08/kWh, Titanium’s premium over Platinum may not justify itself within a 3-year refresh cycle. At $0.25/kWh (common in European colocation), Titanium pays for itself in 18 months. Run the math against your actual rate.
- Do I have dual power grids (A+B feed)? そうであれば, 2+2 redundant PSUs configured as grid-redundant protect against both PSU failure and grid failure. If no (single grid), 1+1 redundant provides only PSU-level redundancy – a grid failure takes the server down regardless of PSU configuration.
- Am I paying for power by the amp or by the kilowatt-hour? If your colocation bills per amp of provisioned circuit capacity, an oversized PSU costs you money even if the server never draws near its rated power – because the circuit must be provisioned for the PSU’s maximum input current, not the server’s typical draw. Right-sizing the PSU is both an energy efficiency decision and a circuit provisioning decision.
Configure Power-Optimized Servers With Xincuan
We ship Dell PowerEdge and xFusion FusionServer servers with PSUs matched to your actual workload power profile – not the worst-case spec sheet maximum. 工場直送価格, 3-年保証, そして世界的な発送.
Request a power-optimized server configuration
Explore: Dell PowerEdge Servers | xFusion Servers | すべての製品