PSU redundancy in GPU servers: N+1, 3+1, 4+2 and N+N, and which mode survives a lost feed
Eurokommerz, Vienna, since 2006: Private AI/ML · IT Managed Services · Enterprise Training · AI Hardware & Software
- N is the number of power supplies that carries the server, N+1 adds one spare and N+N a second full set: Supermicro lists four 3,200 W supplies as 3+1 and six 2,700 W supplies as 4+2, and NVIDIA lists the twelve supplies of the DGX B300 as N+N with six as the minimum
- Only an N+N design can be split over feeds A and B and keep enough supplies when one feed fails; NVIDIA grades two feeds "Not acceptable" for the DGX H100, which needs four of its six supplies
- By our arithmetic three 3,200 W supplies carry 9,600 W and four 2,700 W supplies 10,800 W at full rating; the 3,200 W supplies of Supermicro and Dell give 2,900 W each at the bottom of the range, from 200 to 207 V and from 200 to 220 V
- Dell supports A/B Grid (4+4) for eight 600 W cards in the PowerEdge XE7740 with 3,200 W supplies and not with 2,400 W supplies, and after the loss of a grid the server runs "with reduced performance"
- 80 PLUS Titanium for 230 V redundant supplies means 90 % efficiency at 10 % load, 94 % at 20 %, 96 % at 50 % and 91 % at 100 %
Supplied by Eurokommerz: AI servers, built to order Request a configuration →
N+1 and N+N power supply redundancy in a GPU server
PSU redundancy in a GPU server is written as two numbers: N supplies carry the load, and the number after the plus sign is the reserve. With N+1 the server survives the failure of one supply. With N+N it has a second full set and survives the loss of half its supplies, which is what happens when one of two power feeds fails. For eight-GPU servers the makers' documents give 3+1 for four 3,200 W supplies (Supermicro SYS-422GA-NRT), 4+2 or 3+3 for six supplies (Supermicro's 5U servers), 7+1 and 4+4 for eight supplies (Dell PowerEdge XE7740) and N+N for the twelve supplies of the NVIDIA DGX B300.
Three things decide what a mode is worth in a given room: the rating of each supply, the voltage at the outlet, and whether the supplies hang on one feed or on two. The rack side, with feed capacity and outlets, is in GPU rack power and cooling and GPU server power on 16 A, 32 A and three-phase feeds; this page stays with the power supplies.
How the makers write 3+1, 4+2, 7+1, A/B Grid and N+N
| MODE | SURVIVES | FEEDS A AND B | DOCUMENTED EXAMPLE |
|---|---|---|---|
| Not redundant (N+0) | no failure | no | Dell XE7740: "Not Redundant (8+0)" |
| N+1 (3+1, 7+1) | one supply | no: a lost feed takes more than one supply | Supermicro SYS-422GA-NRT: "Redundant (3 + 1)"; Dell XE7740: "PSU Redundant (7+1)" |
| 4+2 | two supplies | no: three of six remain, four are needed | Supermicro SYS-522GA-NRT: "Redundant (4 + 2)" |
| N+N (3+3, 4+4, 6+6) | half of the supplies, or one feed | yes | Supermicro: "(3+3)"; Dell: "A/B Grid (4+4)"; NVIDIA DGX B300: "N+N" |
Supermicro's product pages for the SYS-422GA-NRT and SYS-522GA-NRT and the overview of manual MNL-2769 (published 15 May 2026); Dell PowerEdge XE7740 Technical Guide, Rev. A04, April 2026; NVIDIA DGX B300 User Guide (Introduction); read 12 November 2026. The column "Feeds A and B" is our count of the supplies left when one of two feeds fails.
Dell defines its three policies in the technical guides for the XE7745 and XE7740. "Not Redundant" has "No failover capability", and "If PSU power is insufficient, the system shuts down immediately." "PSU Redundant", in Dell's words, "Provides failover capability". "A/B Grid" is the "High-availability configuration", in which "PSUs are divided into two independent grids (A and B), each with its own power source." Both servers have eight bays, two in a CPU zone and six in a GPU zone, and Dell writes the rule in N+N terms: "To ensure full redundancy, install N+N number of PSUs in each zone, i.e., 1+1 in CPU zone and 3+3 in GPU zone."
NVIDIA's User Guide for the DGX B300 says: "The system includes 12 power supply units (PSU) configured for N+N redundancy", and "The minimum number of required PSUs is six."
At Supermicro the same chassis carries different modes. The six bays of the SYS-522GA-NRT hold 2,700 W supplies in 4+2 or 3,200 W supplies in 3+3. For the AS-5126GS-TNRT the datasheet prints "6x 2700W Redundant (3+3 or 4+2)" with the note "(Power supply redundancy based on configuration and application load)".
Some documents stop at the word "redundant". Gigabyte's page for the G494-SB4-AAP2, one of the Gigabyte servers for the RTX PRO 6000 Server Edition, lists "Quad 3000 W 80 PLUS Titanium redundant power supplies" and Supermicro's page for the SYS-521GE-TNRT "4x 2700W Redundant Titanium Level (96%) power supplies", both without numbers, so we assign no mode to them. For the ProLiant Compute DL380a Gen12, covered in HPE ProLiant servers for AI, the part of the QuickSpecs we could read on 12 November 2026 names 2,400 W and 3,200 W supplies and no mode, and the power data on ASUS's page for the ESC8000A-E13P could not be read that day. In these cases the mode is a line to ask for in the quote.
Capacity by mode for four, six, eight and twelve supplies
| SERVER AND SUPPLIES | MODE | CARRYING | CAPACITY | AT 200 TO 220 V |
|---|---|---|---|---|
| SYS-422GA-NRT, 4 × 3,200 W | 3+1 | 3 | 9,600 W at 219.1 to 240 V | 9,000 W at 207.1 to 219 V; 8,700 W at 200 to 207 V |
| SYS-522GA-NRT, 6 × 2,700 W | 4+2 | 4 | 10,800 W | the same: one band, 200 to 240 V |
| SYS-522GA-NRT, 6 × 3,200 W | 3+3 | 3 | 9,600 W | not stated for this option |
| AS-5126GS-TNRT, 6 × 2,700 W | 3+3 or 4+2 | 3 or 4 | 8,100 W or 10,800 W | the same: one band, 200 to 240 V |
| Dell GPU zone, 6 × 3,200 W | 3+3 | 3 | 9,600 W at 220.1 to 240 V | 8,700 W at 200 to 220 V |
| Dell GPU zone, 6 × 2,400 W | 3+3 | 3 | 7,200 W | the same: one band, 200 to 240 V |
| DGX B300, 12 supplies | N+N | 6 | 19.2 kW at 3.2 kW each; 19.8 kW at 3.3 kW each | the same: one band, 200 to 240 V |
our arithmetic: the supplies that carry the load times the output the maker prints for the input voltage (next table). Modes and ratings from Supermicro's product pages and datasheets, Dell's technical guides for the XE7745 (Rev. A09) and XE7740 (Rev. A04) and the NVIDIA DGX B300 User Guide, read 12 November 2026.
The table is our arithmetic on printed ratings; no maker states these sums. The Dell rows count the six bays of the GPU zone; the CPU zone has its own two supplies in 1+1. The DGX B300 row has two values because the User Guide prints two ratings on the same page: "12 x 3.2 kW" in the physical table and "3.3 kW @ 200-240 V, 16 A, 50-60 Hz" per supply in the power table. The same page gives "15 kW system max." and a power consumption of 14.5 kW.
A larger reserve lowers the capacity. Six 2,700 W supplies carry 10,800 W as 4+2 and 8,100 W as 3+3, and Supermicro's note on the AS-5126GS-TNRT says which of the two applies depends on configuration and load. Eight 600 W cards are 4,800 W before processors, fans and conversion losses; our wall-power estimates per configuration are in the GPU server power guide linked above. Dell prints the outcome for its own servers in a support matrix, shown in the section on feeds.
We supply Supermicro, Gigabyte, ASUS, Dell and HPE servers per project, with redundant power supplies; tell us the voltage at your rack and whether it has one feed or two, and we reply with a configuration and its redundancy mode.
Output of one supply between 200 and 240 V
| SUPPLY | INPUT VOLTAGE | OUTPUT |
|---|---|---|
| Supermicro 3,200 W | 219.1 to 240 V | 3,200 W |
| Supermicro 3,200 W | 207.1 to 219 V | 3,000 W |
| Supermicro 3,200 W | 200 to 207 V | 2,900 W |
| Supermicro 3,200 W | 100 to 127 V | 1,400 W |
| Supermicro 2,700 W | 200 to 240 V | 2,700 W |
| Dell 3,200 W Titanium | 220.1 to 240 V | 3,200 W |
| Dell 3,200 W Titanium | 200 to 220 V | 2,900 W |
| Dell 2,400 W Titanium | 200 to 240 V | 2,400 W |
| Dell 2,400 W Titanium | 100 to 120 V | 1,400 W |
| Gigabyte 3,000 W | 220 to 240 V | 3,000 W |
| Gigabyte 3,000 W | 200 to 220 V | 2,600 W |
| Gigabyte 3,000 W | 100 to 127 V | 1,000 W |
| NVIDIA DGX B300 | 200 to 240 V | 3.3 kW in the power table, 3.2 kW in the physical table |
Supermicro's product pages for the SYS-422GA-NRT (3,200 W module) and SYS-522GA-NRT (2,700 W module); Dell PowerEdge XE7745 Technical Guide Rev. A09 and XE7740 Technical Guide Rev. A04; Gigabyte's product page for the G494-SB4-AAP2; NVIDIA DGX B300 User Guide; read 12 November 2026.
The figure in a supply's name is its output at the top of the input range. Dell's 3,200 W supply is a 2,900 W supply between 200 and 220 V, Supermicro's module steps down at 219 V and again at 207 V, and Gigabyte's 3,000 W supply gives 2,600 W between 200 and 220 V. A circuit at the nominal 230 V lies in the full-rating band of all three. The bands begin at 219.1 V (Supermicro), 220 V (Gigabyte) and 220.1 V (Dell), so measure the voltage at the PDU outlet and state it in the enquiry. We read no standard on the tolerance of the mains for this article and give no figure for how far the voltage may fall.
The low-line rows (100 to 127 V at Supermicro and Gigabyte, 100 to 120 V at Dell) show what such an input costs: 1,000 to 1,400 W per supply remain. Supermicro's 2,700 W module and the DGX B300 list 200 to 240 V only.
Feeds A and B need an N+N design
Redundancy exists on two levels. Inside the server, the mode says how many supplies may fail. At the rack, the question is whether a whole feed with its PDU and UPS may fail. NVIDIA writes the feed level as 2N: its guide "Data Center Best Practices with DGX B300" (version 1.0) has "2N (each B300 System has Redundant PSUs)" for the rack and plans one three-phase circuit per utility source.
A server can be split over feeds A and B only if the supplies on one feed carry it alone, and that is N+N. NVIDIA documents the opposite case for the DGX H100, an older system that we do not supply. "Four of the six power supplies must be energized for the system to operate." With two feeds, "During a failure of one power source, the number of energized PSUs on a DGX H100 system would fall below four, resulting in a shutdown of that system." NVIDIA's data-centre design guide for the DGX H100 (DG-11301-001 v4, updated 19 November 2025) grades that layout "Not acceptable for DGX H100 systems" and asks for three power sources, each connected to two of the six supplies. By our count, a 3+1 server on two feeds likewise keeps two of the three supplies its mode counts on, and a 4+2 server three of four. Whether it stays up then depends on its load at that moment, and no maker document we read promises it.
The DGX B300 is the N+N case: twelve supplies, six needed, and the Service Manual states that "The system can operate at full capacity with six fully functional power supplies." NVIDIA's documents do not map the supplies to PDUs, so that is a question for the installation partner.
Dell's A/B Grid is N+N per zone, with two limits: it is not supported for every configuration, and it does not keep full performance. After the loss of a grid, the policy "Maintains system operation when multiple PSUs fail (up to four PSU drops), with reduced performance." Which bays belong to grid A and which to grid B is not stated in the guides. The support matrix of the XE7740 for eight 600 W cards, the NVIDIA RTX PRO 6000 Blackwell Server Edition or the NVIDIA H200 NVL, reads:
| POLICY | 3,200 W SUPPLIES | 2,400 W SUPPLIES |
|---|---|---|
| Not Redundant (8+0) | Supported | Supported |
| PSU Redundant (7+1) | Supported | Supported |
| A/B Grid (4+4) | Supported | Not Supported |
Dell PowerEdge XE7740 Technical Guide, Rev. A04, April 2026, Tables 26 and 27, rows for eight "RTX Pro 6000" or eight H200 NVL at 600 W; read 12 November 2026.
With A and B feeds, each feed, PDU and UPS has to carry the whole server; the sizes are in the GPU server power guide and in UPS for a GPU server. For three-phase PDUs the H100 guide adds that "The power draw across the phases of a three-phase circuit should be as balanced as possible."
What the server does when a supply or a grid fails
Dell's policy table has a column for the trigger point of a "GPU power brake". Under "Not Redundant" the brake is not supported. Under "PSU Redundant" it is supported, with the entry "Depends on iDRAC calculation when PSU drops"; under "A/B Grid" it is supported, with the entry "Four PSUs lost within a single grid". Neither guide explains what the brake does to the cards. Both guides also state that the server "does not support user power cap", so the load of these two servers cannot be limited with a user cap at the server. Limits at the card are a separate tool, described in RTX PRO 6000 Server Edition power at 600, 450 and 300 W.
For the DGX B300, NVIDIA's instructions are graded by the number of working supplies. With one failed: "If one PSU fails, troubleshoot the cause and replace the failed PSU immediately." Below six, the Service Manual says: "If insufficient power supplies are present and working, power off the system." The system boots only "if at least three PSUs are operational", and with fewer "only the BMC is available". A supply is swapped "in less than 30 seconds to avoid airflow disruptions", and "All PSUs in the system must be from the same manufacturer."
80 PLUS Titanium efficiency by load point
| LEVEL | 10 % LOAD | 20 % LOAD | 50 % LOAD | 100 % LOAD |
|---|---|---|---|---|
| 80 PLUS Platinum | 88 % | 90 % | 94 % | 91 % |
| 80 PLUS Titanium | 90 % | 94 % | 96 % | 91 % |
| 80 PLUS Ruby | 91 % | 95 % | 96.5 % | 92 % |
80 PLUS programme, page "80 PLUS PSU Ratings explained" (undated), table "230V, 277V Internal Redundant", read 12 November 2026. The page notes that the 10 % criteria below Titanium are suggested; Ruby also requires 90 % at 5 % load.
Supermicro writes "Titanium Level (96%)" and HPE "Up to 96% efficiency". The programme's table shows where that figure applies: at 50 % load. At 100 % load Titanium requires 91 %, at 10 % load 90 %. Redundancy moves the working point: where the supplies of a redundant set share the load, each runs at a lower load point than it would alone, and after a failure the remaining ones run at a higher one.
How to state the redundancy requirement in an enquiry
A quote can name the right supplies if the enquiry contains six facts:
- the number of feeds at the rack position and whether they come from separate sources
- the voltage measured at the PDU outlet and the breaker rating of each feed
- the number and type of free outlets on each feed
- the failure the server has to survive: one supply, or a whole feed
- whether reduced performance after that failure is acceptable
- the cards, their number and their power mode
Redundant power supplies are among the components we size and source for AI servers built to order, and that page states a rack, power and airflow check before we quote. Send us these six points, or write to request@
FAQ
What is the difference between N+1 and N+N PSU redundancy?
What does 3+1 PSU redundancy mean?
Is there 2+2 redundancy for an 8-GPU server?
Can a server with N+1 power supplies run on A and B feeds?
How many power supplies does a DGX B300 need to keep running?
Does a 3,200 W power supply deliver 3,200 W at 230 V?
Are Titanium power supplies 96 per cent efficient?
We supply GPU servers built to order across the EU, with redundant power supplies sized to the feeds of your site. Describe your task in a few sentences and our team will help. We reply within one business day.
Request a configurationWe reply within one business day