Twelve CAD Workstations, One Circuit
Your firm is standardizing 12 engineering workstations for 2D and 3D CAD. The approved build is a 16-core CPU (125 W TDP, roughly 200 W package peak), a 320 W professional GPU, 2 x 32 GB DDR5-5600, an 850 W 80+ Gold PSU, and a motherboard with two M.2 slots and two x16-length PCIe slots. The manual states the second x16 slot is wired x4 and shares bandwidth with M.2_2. Pilot users report three things: benchmark scores are 15% lower after 30 minutes of load than on a cold run; a 10 GbE NIC in the second x16 slot is not detected; and two units have shut down during combined CPU-and-GPU renders. The chassis is small-form-factor with two 120 mm intake fans and a 240 mm AIO.
1. Benchmark scores drop 15% after 30 minutes of load. Walk through how you determine whether this is thermal, power, or software, and what you would change in the build.
Read the worked answer
Log before you touch anything. Record core clocks, package power, CPU temperature, GPU clock, GPU hotspot temperature, and fan RPM at one-second intervals across a 60-minute run. Then read the shape of the drop:
- Clocks fall and temperatures sit at a ceiling (CPU near 95-100 C, GPU hotspot near 100-105 C) -> thermal.
- Clocks fall while temperatures are flat and package power pins at its configured limit -> power/current limit, not cooling.
- Neither, and the drop mirrors a background task (indexing, antivirus scan, cloud sync) -> software.
Heat budget check: 200 W CPU + 320 W GPU + roughly 60 W of board, drives, and fans is about 580 W inside a small chassis. An open-air 320 W GPU dumps most of that into the case, and a 240 mm AIO mounted as front intake pre-heats the air the GPU breathes. That combination explains a slow-onset drop far more often than a bad paste job.
Change in this order, cheapest and most reversible first:
- Fix airflow - AIO as top/rear exhaust, dedicated cool intake to the GPU, add exhaust if pressure is positive and stagnant.
- Raise fan curves; confirm the case is not recirculating against a wall or under a desk.
- Undervolt or set a modest power limit - often recovers most of the lost clock for a few percent of peak performance.
- Repaste only if you measured a temperature delta against the other eleven units.
- Change chassis only if steps 1-4 fail.
Verification: a 60-minute sustained run whose clocks stay within 5% of the cold-run values.
2. The 10 GbE NIC in the second x16 slot is not detected, and one user has ordered a third NVMe drive plus a capture card. What do you do?
Read the worked answer
Prove the conflict before redesigning around it. Read the manual's lane-sharing clause literally: "shares bandwidth with M.2_2" usually means populating M.2_2 either disables the slot outright or drops it to x1. A 10 GbE NIC needs x4 (PCIe 3.0 x4 is about 3.9 GB/s, ample), so x1 will not do - the card may still not enumerate, or it may enumerate and run badly.
Test: power down, remove the M.2_2 drive, reseat the NIC, clear CMOS, and check for enumeration in BIOS and the OS. If it appears, the conflict is confirmed and the manual is telling the truth.
Then allocate the scarce lanes by business value, because you cannot have all four devices:
- GPU stays in the primary x16 slot at x16 - CAD performance depends on it.
- Replace the two 2 TB NVMe drives with a single 4 TB drive, freeing M.2_2. Cost is roughly the same; capacity is the same; the slot conflict disappears.
- Put the 10 GbE NIC in a chipset-connected slot if the board has one. It shares the chipset uplink, but for workstation traffic that is bursty rather than sustained, this is an acceptable trade.
- If the board has no chipset slot, the honest answer is no: 10 GbE plus a capture card plus three NVMe plus a x16 GPU is not possible on a mainstream consumer board. The alternative is a workstation/HEDT platform with 64+ CPU lanes, which is a different budget conversation.
Decline the third NVMe. Saying yes to an order that cannot be installed costs more goodwill than saying no now.
3. Twelve of these workstations will share a single 20 A / 120 V circuit. Verify the electrical plan and the PSU choice, and explain how you would distinguish a PSU shutdown from a breaker trip.
Read the worked answer
The circuit fails by a wide margin. A 20 A / 120 V circuit is 2.4 kVA, derated to 80% for continuous load = about 1.92 kW. Twelve workstations drawing roughly 600 W under render load is about 7.2 kW - nearly four times the circuit. Even at an idle-ish 350 W each, it is 4.2 kW. This is not a marginal case; it will trip in normal use and it is a fire and downtime risk.
Specification: two workstations per 20 A circuit gives comfortable headroom, or four per 30 A / 240 V circuit if the building allows it. Confirm with an electrician, and put renders on a schedule so all twelve do not peak simultaneously.
PSU check: 850 W against roughly 600 W continuous is about 71% load, which is a healthy operating point - provided it is a single 12 V rail delivering roughly 70 A and can absorb the GPU's microsecond transient spikes, which can approach twice the board's rated power. Check the spec sheet for a single rail and adequate OCP threshold, not just the wattage number.
Distinguishing the two failures:
- Breaker trip: other equipment on that circuit also dies, the breaker handle is mid-travel, and it resets immediately by hand.
- PSU OCP/OPP trip: only that workstation dies, the breaker is untouched, and it powers back on after a few seconds. It typically happens at the instant a render starts, not during steady load.
Do not fix a breaker problem with a bigger PSU. UPS or at least surge protection on every unit, and never put the AIO pump on a switched outlet.