Jetson Orin high RAM usage at idle — tegrastats and memory reduction

Jetson Orin memory accounting is confusing because it uses a unified memory architecture where GPU, video codec, and AI accelerators share the same physical LPDDR5. What looks like “high RAM usage” at idle is mostly hardware carveouts and IOVA reservation — not Linux process memory. Understanding the breakdown lets you recover memory where your use case doesn’t need certain hardware blocks.

Key Insights

• Unified memory means hardware blocks pre-reserve their working space — GPU, NVDEC, VIC, ISP, and NvMedia each carve out physically contiguous regions before Linux userspace starts
• tegrastats shows actual breakdown — look at RAM / SWAP / IOVA / GR3D fields separately
• CMA is the most recoverable reservation — default CMA is 512MB on Orin; if you don’t use hardware video decode/encode, reduce it in the kernel cmdline
• jtop gives a more readable view — the open-source jtop tool wraps tegrastats in a human-friendly ncurses UI
• GPU carveout scales with CUDA context count — each new CUDA context reserves GPU channel memory; limiting concurrent CUDA users reduces this

Reading tegrastats output

tegrastats --interval 1000

Example output:

RAM 2847/7772MB (lfb 512x4MB) SWAP 0/3886MB
CPU [12%@1574,5%@1574,8%@1574,4%@1574,3%@1574,3%@1574,2%@1574,2%@1574,0%@729,0%@729,0%@729,0%@729]
GR3D_FREQ 0%@0 NVDEC_FREQ 0%@0 NVENC_FREQ 0%@0 VIC_FREQ 0%@0
EMC_FREQ 0%@204 APE 25 MTS fg 0% bg 0%
IOVA 104 MB NVMAP 1124 MB AO 0 MB

Field	What it means
RAM 2847/7772MB	2847MB used of 7772MB visible (AGX Orin 64GB minus carveouts)
lfb 512x4MB	Largest Free Block count × size — key indicator of fragmentation
IOVA 104 MB	DMA address space currently mapped by SMMU
NVMAP 1124 MB	Memory mapped via NvMap (NVIDIA DMA-buf handle allocator) — includes video, GPU, ISP

Where idle memory goes

On a full JetPack install at idle (AGX Orin 32GB):

Consumer	Approximate size	Recoverable?
Linux kernel + userspace	~800MB	No
NVIDIA display driver (nvdisplay)	~200MB	Only if no display needed
CMA (kernel contiguous allocator)	512MB	Yes — reduce to 128MB if no HW video decode
GPU channel carveout	~300MB	Partially — reduce with fewer CUDA contexts
NVDEC/NVENC power domain init	~150MB	Yes — disable if not using HW video codec
VIC (Video Image Compositor)	~100MB	Yes — disable if not using Argus/DeepStream
NVMAP shared buffer pool	~300MB	Partially

For the Orin Nano 4GB, all of these compete in a smaller pool — the defaults leave less than 2GB visible to free -h.

Reducing CMA

CMA reserves physically contiguous memory for video codec buffers. If you don’t use hardware video decode/encode, reduce it:

# In /boot/extlinux/extlinux.conf, add to APPEND:
cma=128M
# Default is 512M; 128M is sufficient for display-only use

# Or disable CMA entirely (no hardware video decode/encode at all):
cma=0

Rebuild the kernel or edit extlinux.conf — no reflash needed.

Disabling unused video engine power domains

If your application does not use NVDEC/NVENC/VIC:

# Check which video engines are powered on
cat /sys/kernel/debug/bpmp/debug/clk/*/state 2>/dev/null | grep -E "nvdec|nvenc|vic"

# Disable NVDEC power domain at runtime
echo 0 > /sys/devices/platform/tegra-dce/power/control
# (device path varies by L4T version — check /sys/bus/platform/devices/)

Via Yocto, remove the NvMedia and multimedia packages if not needed:

# In your image recipe, replace demo-image-full with a custom base:
IMAGE_INSTALL:remove = " \
    tegra-multimedia-api \
    libargus \
    nvargus-daemon \
"

Monitoring memory under load

# Real-time combined view
tegrastats --interval 500

# Check memory fragmentation (lfb = largest free block)
# lfb dropping to single digits = severe fragmentation
cat /proc/buddyinfo

# Per-process memory (includes GPU mappings)
cat /proc/$(pidof your_app)/status | grep -E "VmRSS|VmPeak|VmSwap"

# IOVA usage per device
ls /sys/kernel/debug/iommu/

jtop: the readable alternative

# Install jtop
pip3 install jetson-stats

# Run
sudo jtop

# Programmatic access
from jtop import jtop
with jtop() as jetson:
    print(jetson.memory)

jtop shows GPU memory, CPU memory, swap, and hardware engine utilization in one view with historical graphs.

FAQ

Why does Jetson Orin show 2-4 GB RAM used at idle?

Most idle memory is hardware carveouts for GPU, NVDEC, VIC, CMA, and IOVA reservation — not Linux process memory. On a full JetPack install, the NVIDIA multimedia stack pre-allocates these regions before your application starts.

What does IOVA mean in Jetson tegrastats output?

IOVA is the DMA address space currently mapped by the SMMU for GPU, video codec, and ISP DMA engines. It does not directly equal physical memory consumption.

How do I reduce idle RAM usage on Jetson Orin?

Reduce CMA (cma=128M in extlinux.conf), remove unused NvMedia packages, and disable NVDEC/NVENC power domains. On a minimal production image this can recover 500MB–1GB of usable RAM.

What is the minimum usable RAM on Jetson Orin Nano 4GB?

On a full JetPack install, approximately 2.5–3GB is available to applications. A minimal image without CUDA/multimedia raises this to ~3.2–3.5GB.

---