GPU Core Cool but Memory Junction Scorching? GDDR6X Memory Temperature Mechanism & Cooling Guide
Key Takeaways
GDDR6X memory can run at 104–110°C while the GPU core stays at 60–70°C. HWiNFO's Memory Junction reading is the hottest VRAM chip's internal temperature, and throttling starts near 110°C regardless of core temp. Causes, consequences and safe handling — including why repadding thickness matters — explained.
Why GDDR6X Runs So Hot — The Mechanism
GDDR6X was developed through a close collaboration between Micron and Nvidia, and the first cards to carry it were the GeForce RTX 3080 and RTX 3090, with Micron as the only supplier at launch (Tom's Hardware, September 2020). The key innovation is PAM4 signalling: while GDDR6 uses NRZ encoding (one bit per symbol), GDDR6X uses four-level pulse amplitude modulation — two bits per symbol. That pushes per-pin speeds to 19–21 Gbps and gives the RTX 3090 roughly 936 GB/s of total memory bandwidth.
Here is the power story, and it needs careful wording: Micron's public data shows GDDR6X consumes about 15% less energy per transmitted bit than GDDR6 (approximately 7.25 pJ/bit versus about 7.5 pJ/bit, cited via wevolver). It would be wrong to conclude from this that GDDR6X is "more efficient and therefore cooler." Because the interface runs so much faster, the memory subsystem as a whole draws more total power and produces more heat. Less energy per bit, far more bits per second — net result, a hotter card.
PAM4 also brings engineering cost: it is more sensitive to the signal-to-noise ratio than NRZ, demanding more complex equalisation and signal-integrity design, which makes the chip implementation more complex and more expensive (Wikipedia). The speed is real, and so is the heat.
What "GPU Memory Junction Temperature" Actually Is
In HWiNFO, the reading labelled "GPU Memory Junction Temperature" is the internal junction temperature measured inside the silicon of the VRAM chips — not the external case temperature. The HWiNFO author's own words: "This is the internal junction temperature measured inside the silicon, NOT the external (case) temperature!!!" (quoted in an NVIDIA developer forum thread). That distinction matters, because it trips up a lot of comparisons.
Two more facts about this reading:
- According to the HWiNFO author, the figure shown is the hottest of all the VRAM chips, with an upper operating limit around 105–110°C (TechPowerUp forum, quoting the author).
- A technical community analysis notes that Micron's ~95°C specification refers to case temperature (Tcase), while the junction temperature HWiNFO shows runs roughly 10–15°C higher than the case. On that basis, Nvidia's firmware throttling at a 110°C junction reading is not in conflict with Micron's 95°C rating — they are measuring two different things (community analysis on vibingtalk, not an official document).
One practical upgrade: starting with HWiNFO v8.51 in 2025, per-chip VRAM temperature readings became available for AMD Navi and Nvidia GDDR6X/GDDR7 cards (Club386). Before that there was only a single aggregate reading, with no way to know whether it was an average or the hottest chip. Per-chip readings let you locate a problem: if one chip runs more than 10–15°C hotter than the others, that usually points to a thermal-pad contact problem at that spot, not a card-wide issue.
The RTX 3090 FE Phenomenon: Memory at ~110°C, Core at 60–70°C
This is the documented scenario that confused a generation of owners: public measurements showed the RTX 3090's memory junction hitting 110°C while the GPU core sat at a comfortable 60–70°C. Tom's Hardware recorded exactly that pairing; another test showed the RTX 3080 at 94°C and the 3090 at 104°C; Notebookcheck reported on YouTuber CryptoAtHome's unmodified 3090 FE reaching 110°C — and dropping to about 85°C after a thermal-pad replacement, a reduction of roughly 25°C.
Three structural causes combine here (researchers caution against blaming any single one):
- The 3090 has VRAM chips on the back of the PCB too. The 3090's 24 GDDR6X chips are split across the front and back of the board (the 3080 has only 10 chips), and the back-side chips rely on passive backplate cooling — there is no active airflow over them (Tom's Hardware forums).
- The Founders Edition's stock thermal pads are mediocre. Multiple owners reported the FE pads were on the thin side with poor contact; replacing them with quality pads improved temperatures immediately (vibingtalk; MSI forums).
- First-batch cooler designs did not keep up with GDDR6X. A widely echoed point in the Tom's Hardware forum discussion: "the first batch of cards, including the FE, clearly didn't put enough work into VRAM cooling" — backed up by the fact that well-cooled third-party designs (a Colorful 3080 Vulcan, for example) ran at just 96°C under mining load. The heat is a design difference, not a GDDR6X death sentence.
The 3090 Ti quietly confirmed this diagnosis: it switched to higher-density 16Gb chips so all 24GB could live on the front of the PCB, eliminating the back-side cooling problem (Tom's Hardware forums). And there is even a public record of a one-off factory defect — a Reddit user found a factory finger cot (finger cover) left between the 3090 FE's heatsink and the VRAM; removing it dropped temperatures from above 110°C to 86°C (progressiverobot blog).
GPU Hotspot Is Not Memory Junction — Do Not Confuse Them
Two different sensors, two different stories. GPU Hotspot is the hottest point on the GPU die itself, and it reflects die cooling (thermal paste, mounting pressure). Memory Junction is the internal temperature of the VRAM chips, and it reflects VRAM cooling (thermal pads, backplate). They have no necessary connection: a high hotspot does not mean hot VRAM, and hot VRAM does not mean a high hotspot — the classic 3090 situation is a normal core and hotspot with memory junction off the chart (TechPowerUp forums: "GPU Hotspot is GPU Hotspot and not VRAM Temp").
For reference (not hard standards): the core-vs-hotspot gap of around 20°C is typical; and a memory-junction-vs-core gap of more than 30°C is a documented common occurrence on 3080/3090 FE cards — it does not mean the card is dying this instant, but it is worth addressing (Tom's Hardware forums; a Steam user measured an RX 7900 XTX at 68°C core / 88°C hotspot / 90°C memory junction, showing the three can diverge independently).
The most common misconception is simply watching the core temperature and concluding "the card is cool" — while the VRAM is already past 105°C (dev.to).
What Overheating Does to the Card
Thermal throttling starts at around 110°C. Nvidia's firmware begins reducing memory and overall clocks when the memory junction reaches approximately 110°C (HWiNFO author: "Thermal throttling starts around 110C"); one MSI forum user reported normal operation below 107°C with slight throttling beginning at 108°C. The decisive point: throttling looks only at the memory junction. A 60°C core does not save you — VRAM at 110°C throttles regardless (dev.to).
Beyond throttling, the symptoms are familiar to any repair bench:
- Artifacts and driver crashes. Techbloat's symptom table lists overheating or unstable VRAM as one of the most likely causes of artifacts appearing minutes into gaming; aeanet notes that artifacts (coloured dots, lines, texture errors) are almost always a GPU or VRAM failure signal. A vibingtalk user reported a VRAM temperature of 110°C causing a straight system shutdown.
- Long-term aging. Tom's Hardware discussion quotes the warning that Micron rates GDDR6X at a maximum of about 95°C (operating temperature) and that "running over spec long-term is a recipe for disaster." Some users adopt the chip-maker's 95°C Tcase as their 24/7 operating guideline and treat Nvidia's 110°C as merely a functional limit within the warranty period (MSI forums). Put qualitatively, without inventing lifespan figures: running close to the throttle line for months accelerates aging.
What Your Readings Mean — A Practical Guide
The figures below are approximate, compiled from public records (reviews and forum reports). They are a reading guide, not manufacturer guarantees.
| Memory junction reading | Status | What to do |
|---|---|---|
| ~75–85°C | Normal. Around 80°C is considered safe; no pad replacement needed | Nothing — stop worrying about the number |
| ~85–95°C | Warm but typical for GDDR6X under sustained load | Improve case airflow, clear dust |
| ~95–105°C | High. Worth addressing if you run 24/7 sustained loads | Check airflow and pad contact; consider backplate cooling |
| ~105–110°C | Throttle zone. Slight throttling reported from ~108°C; ~110°C throttling starts | Repad, add backplate cooling, or lower memory clocks |
| Sustained 110°C+ | Throttling active; shutdown reports exist at this level | Stop heavy loads and investigate urgently |
Note that a steady 80°C reading is safe — there is no reason to dismantle the card over a number you don't like (TechPowerUp forums). Also remember the workload matters: memory temperatures that are fine in gaming can explode when you switch to mining or LLM inference, because those are continuous memory-intensive loads with no idle gaps. Never infer 24/7 compute temperatures from gaming temperatures (dev.to).
Handling Principles, From Lightest to Heaviest
1. Case airflow and dust cleaning — zero risk. Ambient temperature and intake airflow directly affect the backplate area. Reports describe temperature drops of a few degrees from improved case ventilation alone. Do this first.
2. Backplate heatsinks and fans. The 3090's back-side chips have only the passive backplate, so this is where many owners start. Publicly documented mods include sticking heatsinks to the backplate and pointing a fan at it; Igor's Lab used an infrared camera to locate hotspots and then added 3mm thermal pads at the exact positions (igorslab; myminingrig).
3. Lower the memory clock or power limit. Reducing the memory clock or the card's power limit directly cuts VRAM heat, and some 3090 owners reported stability after lowering power. The MSI forum user who pushed +975 MHz on the memory and watched temperatures climb to 98°C demonstrates the reverse (MSI forums).
4. Replace the thermal pads — with the critical thickness warning. Documented results are strong: the 3090 FE repad in the Notebookcheck report went from about 110°C to about 85°C, roughly 25°C. But the pads must be the right thickness — wrong thickness measurably worsens temperatures. A documented 3080 FE case used 2mm pads on the GPU side where 1.5mm was correct; the GPU temperature went up by nearly 20°C, because the too-thick pads lifted the heatsink and opened a gap against the die — a tiny distance change with huge consequences (myminingrig; Igor's Lab likewise notes that an excessively thick interface layer increases the distance between heatsink and memory modules). The correct procedure: after opening the card, measure the original pad thickness with a caliper (or obtain the manufacturer's official figure), use the same thickness at each location, and test-fit to check die contact and pad compression marks before final assembly (Tom's Hardware forums).
5. Understand the disassembly risk. Removing the heatsink may affect the manufacturer's warranty depending on region and vendor policy — one MSI forum user specifically wrote in to ask whether replacing pads voids warranty — and a botched job can damage the card. If you are not confident, do not open it yourself.
When It's a Hardware-Fault Signal — Send It In
Cleaning, airflow and stock clocks are the dividing line: if intermittent artifacts or driver crashes persist after you have cleaned the card, improved airflow and restored default clocks, that is the signal to send it in. Persistent artifacts with high VRAM temperatures point to a VRAM hardware problem — BGA solder joints or the chips themselves — which pad replacement cannot fix (techbloat; aeanet).
For diagnosis, OCCT includes a dedicated VRAM test that can help distinguish a GPU-core problem from a VRAM problem — a publicly documented diagnostic practice — but confirmation and repair need professional equipment (Tom's Hardware forums). One plain-language rule of thumb: after dust is cleared, airflow is fixed and clocks are back at stock, still artifacting or crashing means it needs a professional diagnosis.
Frequently Asked Questions
Q1: My memory junction reads 100°C. Is the card about to die?
Not necessarily. Throttling starts near 110°C, and 100°C under sustained load is a documented common occurrence on GDDR6X cards — but watch the trend and the symptoms. Climbing temperatures over weeks, new throttling, or artifacts mean it needs attention. Also note the reading is the internal junction temperature, roughly 10–15°C above the chip's case temperature, so don't compare it directly against Micron's ~95°C Tcase rating.
Q2: Why is the memory at 105°C when the core is only 65°C?
Because the two are cooled by different paths and measured by different sensors. The GPU core is directly under the heatsink with active fans; the VRAM — especially back-side chips on cards like the 3090 — may only get passive backplate cooling, and the stock pads on some designs make poor contact. Throttling watches the memory junction, not the core.
Q3: Does a high GPU Hotspot reading tell me anything about VRAM?
No. Hotspot reflects die cooling (paste, mounting pressure); memory junction reflects VRAM cooling (pads, backplate). Either can be fine while the other is not. Fix the one that is actually hot — repasting the die will not cool the VRAM.
Q4: Can I just buy any thermal pads? Thicker must be better, right?
No. Pad thickness must match the original specification per location. Too-thick pads lift the heatsink off the die and make the whole card run hotter — a documented case showed GPU temperatures rising nearly 20°C from a 2mm-vs-1.5mm mistake. Measure the original pads with a caliper or confirm the official thickness, and test-fit before closing up.
Q5: My card runs at 80°C memory junction. Should I replace the pads for safety?
No. Around 80°C is considered safe, and opening the card carries its own risks (including warranty questions). Handle the number with airflow and cleaning; save repadding for genuinely high readings.
Q6: What does OCCT's VRAM test tell me?
It is a preliminary diagnostic tool that can help distinguish whether instability comes from the GPU core or the VRAM. It does not repair anything, but it tells you — and a repair shop — where to look.
Need a Professional Diagnosis?
If your card is throttling at the memory junction limit, artifacting, or crashing after you've already cleaned it and restored stock settings, it may need chip-level VRAM diagnosis. WhatsApp 6558 6806, call 2186 8010, or visit the shop at Room 1209, Dali Plaza, Shek Mun. Repairs follow three-tier fees of HK$400–2,800, with the final quote confirmed after inspection.