The V100's memory hits 85 °C whatever I cool it with — I sold the card and I'm staying on the RTX 5070 Ti
The card is idle: model unloaded, nothing running, 48 W, and the die reads 82 °C, then 87 two minutes later — the point where it starts cutting clocks. Three weeks ago it idled at 50.
It's a Tesla V100 SXM2 32GB, ₽65K secondhand, on a PCIe adapter with a three-fan shroud. My last post about it was «The host froze on the first night…». This one is about cooling the HBM, and where that post was wrong.
The memory stops at 85 °C whatever I change
Rated limits from nvidia-smi -q: memory 85 °C, die slowdown 87 °C, shutdown 90 °C. The memory clock never dropped below 877 MHz in any sample, so throttling shows up only in the SM clock and the SW Thermal Slowdown flag. gpu_burn, 300 W cap, 120 seconds:
| What I changed | Memory, °C | Die, °C | Power, W | SM clock, MHz |
|---|---|---|---|---|
| 3 fans (2 Arctic P8 + stock Foxconn), 23.09 | 84-85 | 71-75 | 136-190 | 1000-1185 |
| 4 fans | 84-85 | 72 | 147-189 | 1035-1147 |
| 3 stock Foxconn, reseated pads (later found cracked) | 84-85 | 74-75 | 90-115 | 562-637 |
| 180 W cap, 03.10 | 86 in 75 s | 76 | 118 avg | 715 |
| new pads 3 and 4 mm, 06.10 | 87, 85 °C at 12 s | 83 | 90 avg | 475 |
Same memory ceiling in every row. Fan count and type never moved it (the stock Foxconns are 72 dB on paper, by the way). With the 180 W cap the card never drew 180: it cut clocks at 118 W because the memory hit 85 °C.
The new pads — 3 mm over the main HBM stacks, 4 mm over the outer ones — cut the gap between memory and die from about 10 K to 1-3 K. The memory still reached 85 °C in 12 seconds and the slowdown kicked in at four.
Real load looked the same. Two concurrent 28K-token requests for nine minutes: memory averaged 84 °C, peaked at 93, and was at or above 85 in 43 of 111 samples. Prompt processing dropped to about 450 tokens/s combined. On day one, unloaded, it did 141.9 tokens/s on one request and 57.8 tokens/s generating.
My last post was wrong
In that post I blamed power: two connectors engaging at once and freezing the host. I later ran the full cycle — reboot, cold start, both cards under load — and the freeze never came back. The problem was cooling.
The «6-7 month payback» in the same post assumed the card as a daily driver. It never got that role: on my private test set it scored 45% acceptance against an 80% gate. That is a model result, not a cooling one. For this card, that payback doesn't hold.
The server heatsink cooled the die, not the memory
Next I ordered a server SXM2 heatsink. It meant pulling the RTX 5070 Ti and putting the V100 in the x16 slot. It doesn't cover the memory (pads are needed there too), but the die cooled down:
120 s gpu_burn | Bent plate | Server heatsink |
|---|---|---|
| Memory ≥ 85 °C | at 12 s | at 37 s |
| First slowdown | at 4 s | at 53 s |
| Die, max | 83 °C | 74 °C |
| Mean SM clock | 475 MHz | 1144 MHz |
| Gflop/s (double precision) | 1543-1821 | 5637-5749 |
The die now has headroom and the memory doesn't: memory sits 12 K above the die, where the bent plate had 1-3 K with a hot die. That's where most of the loss is.
The memory clock never moved, so I expected generation speed to hold. With two streams it was 16-26 tokens/s per stream against 57.8 on one. Two concurrent requests and throttling are mixed together here; I didn't test a single stream under load.
Why I'm staying on the RTX 5070 Ti
My local model is for data that must never go to the cloud. Everything else goes to cloud models through a PII filter. Since 23.09 the local lane runs on the RTX 5070 Ti and the V100 is out of it. On the 22.09 test:
→ RTX 5070 Ti on a single request: time to first token 24% lower, generation 44% faster → V100 won on two parallel requests, and has 32 GB against 16
That wasn't worth the cooling work for a card that failed the quality gate.
A fair caveat. One card, one cooling kit. I didn't measure room temperature; the thermal-resistance estimate of about 0.7 K/W assumes 27 °C. The plate had been dropped and bent, one screw wouldn't tighten, the pads were 3-4 mm: I never separated which of those matters most, and I didn't check the plate with a straightedge or take a paste imprint. The measurements show that my cooling path doesn't work, not that the card is bad: ECC on, 0 errors, 0 retired pages, gpu_burn clean on every run.
I sold the card.