vraminfo

Read NVIDIA GPU memory (VRAM) vendor, type and temperature on Linux — including VRAM temperatures that nvidia-smi does not report.

English | 中文文档

$ sudo ./vraminfo
PCI            DEVICE   MEMORY         MAKER     MEM TEMP      NOTE
02:00.0        0x2204   GDDR6X         Micron    58 C

$ sudo ./vraminfo --per-module
PCI            DEVICE   MEMORY         MAKER     MEM TEMP      NOTE
20:00.0        0x2b85   GDDR7          Samsung   39 C [38 38 36 36 38 38 36 39]

Installation

Requirements

• Linux, an NVIDIA GPU, a recent driver — the NVAPI bridge libnvidia-api.so.1 ships with the driver (present from ~R515 onwards).
• gcc (or any C compiler) and GNU make.
• root for the temperature part (/dev/mem is opened read-only).
• On kernels with a strict /dev/mem policy, add iomem=relaxed to the kernel command line (/etc/default/grub → GRUB_CMDLINE_LINUX_DEFAULT="... iomem=relaxed", then sudo update-grub && sudo reboot). Many distributions work without it.
• Secure Boot must be disabled, otherwise /dev/mem cannot be mapped.

Build and install

git clone https://github.com/xzwgit/vraminfo.git
cd vraminfo
make                 # or: gcc -O2 -Wall -o vraminfo vraminfo.c -ldl
sudo make install    # installs /usr/local/bin/vraminfo

Usage

sudo vraminfo                 # table, hotspot temperature per card
sudo vraminfo --per-module    # list every DRAM module's temperature
sudo vraminfo --json          # machine-readable output
sudo vraminfo --watch         # refresh every 2 seconds
vraminfo                      # as a normal user: vendor/type only, no temperature

JSON example:

{
  "gpus": [
    {
      "bdf": "0000:02:00.0",
      "device_id": "0x2204",
      "subsystem": "0x1043:0x87af",
      "board_vendor": "ASUSTeK Computer Inc.",
      "board_vendor_id": "0x1043",
      "memory_maker": "Micron",
      "memory_type": "GDDR6X",
      "memory_maker_id": 10,
      "memory_type_id": 15,
      "memory_temp_c": 52,
      "note": null
    }
  ]
}

What it reports

Field	Source	Root needed	Notes
Board vendor (ASUSTeK / GALAX / MSI / …)	PCI subsystem ID, resolved with the system pci.ids	no	every card, driver not involved; shown as N/A when the subsystem ID does not name a board partner (e.g. NVIDIA's generic reference ID 0x10de)
Memory maker (Samsung / Hynix / Micron / …)	NVAPI NvAPI_GPU_GetRamMaker	no	read from the driver, works on every card
Memory type (GDDR5 / GDDR6 / GDDR6X / GDDR7)	NVAPI NvAPI_GPU_GetRamType	no
Memory temperature	GPU registers over MMIO	yes	only on DRAM types that have a sensor (GDDR6X, GDDR7)
Per-module temperatures	per-module DRAM sensors (GDDR7)	yes	--per-module

The register path is selected from the memory type reported by NVAPI, not from a model whitelist, so new cards generally work without code changes:

• GDDR6X (Ampere / Ada — RTX 3080/3090/3090 Ti, RTX 4070 Ti/4080/4090, …) BAR0 + 0xE2A8, temperature in bits [11:0], Celsius = field / 32.
• GDDR7 (Blackwell — RTX 5090, …) per-module DRAM sensors (DQR): module p at BAR0 + 0x9024C0 + p*0x4000, validity nibble in bits [27:24] of +0x10 (must be 0xF), temperature as a GDDR MR-code in bits [23:16] (code 20 = 0 °C, +2 °C per unit above).
• GDDR6 / GDDR5 / … — no readable sensor; reported as unsupported (this is a hardware property of the DRAM, not a software limitation: GDDR6X and GDDR7 chips carry an on-die thermal sensor, plain GDDR6 does not).

How the temperature is read

nvidia-smi/NVML do not expose memory temperature on Linux for most setups (the open kernel modules report N/A, and the field is absent for many consumer SKUs). The memory temperature lives in the GPU's register aperture, which can be read from userspace:

1. find the GPU's BAR0 base in /sys/bus/pci/devices/<bdf>/resource;
2. open("/dev/mem", O_RDONLY) and mmap() the page containing the register (read-only);
3. read one 32-bit register and decode it as described above.

Nothing is written to the GPU. The tool only ever maps pages PROT_READ.

Verified on

GPU	Memory	Reported	Temperature source
RTX 3090	Micron GDDR6X	✔	0xE2A8 (cross-checked against another independent reading of the same register — values agree)
RTX 4090 (48 GB mod)	Micron GDDR6X	✔	0xE2A8 — readable even though nvidia-smi shows N/A
RTX 5090	Samsung GDDR7	✔	per-module DQR sensors
RTX 3060	Samsung GDDR6	✔	no sensor — correctly reported unsupported
CMP 50HX	Micron GDDR6	✔	no sensor (0xE2A8 reads 0 on this card) — reported unsupported

Credits

The register offsets and decoders were established by the public reverse-engineering work of the gddr6 project and the wider modding community (Paulo Gomes, Unwinder, asder00, Martin Malík / HWiNFO). This is an independent implementation that adds NVAPI-based maker/type reporting and type-driven register selection.

License

MIT — see LICENSE.