diff --git a/README.md b/README.md
index c923b01..8dc215c 100644
--- a/README.md
+++ b/README.md
@@ -8,6 +8,7 @@
 - Sortable process columns (PID/CPU/MEM) with direction indicators
 - Toggleable process details pane (right on wide terminals, bottom on narrow)
 - CPU total bar + per-core mini-bars
+- GPU widget (NVIDIA via `nvidia-smi`, AMD via Linux sysfs when available)
 - Memory/swap bars and disk usage bars (top-N by usage)
 - Disk pseudo-filesystem filtering toggle (`tmpfs`/`devtmpfs`/`overlay`)
 - Compact network rates + totals + interface label
@@ -58,6 +59,7 @@ Accepted interval range: `200..=5000` ms.
 - Target platform is Linux.
 - Network throughput is computed as a delta between consecutive samples.
 - Disk and network metrics are system-wide aggregates.
+- GPU metrics are best-effort: NVIDIA requires `nvidia-smi`; AMD uses available `/sys/class/drm` fields.
 
 ## Next steps
 
diff --git a/src/collectors.rs b/src/collectors.rs
index b5cf260..1d1381b 100644
--- a/src/collectors.rs
+++ b/src/collectors.rs
@@ -1,9 +1,14 @@
-use std::time::{Duration, Instant};
+use std::{
+    fs,
+    path::{Path, PathBuf},
+    process::Command,
+    time::{Duration, Instant},
+};
 
 use anyhow::Result;
 use sysinfo::{CpuRefreshKind, Disks, MemoryRefreshKind, Networks, RefreshKind, System, Users};
 
-use crate::state::{DiskSample, ProcessSample, Snapshot, Totals};
+use crate::state::{DiskSample, GpuSample, ProcessSample, Snapshot, Totals};
 
 pub struct Collector {
     system: System,
@@ -136,6 +141,7 @@ impl Collector {
             .collect::<Vec<_>>();
 
         let load = System::load_average();
+        let gpus = collect_gpu_samples();
         let snapshot = Snapshot {
             host: self.host.clone(),
             uptime_seconds: System::uptime(),
@@ -153,8 +159,155 @@ impl Collector {
             processes,
             network_totals: Totals { rx_bytes, tx_bytes },
             network_interfaces,
+            gpus,
         };
 
         Ok((snapshot, elapsed))
     }
 }
+
+fn collect_gpu_samples() -> Vec<GpuSample> {
+    let mut gpus = collect_nvidia_gpus();
+    gpus.extend(collect_amd_gpus());
+    gpus
+}
+
+fn collect_nvidia_gpus() -> Vec<GpuSample> {
+    let output = Command::new("nvidia-smi")
+        .args([
+            "--query-gpu=name,utilization.gpu,memory.used,memory.total,temperature.gpu",
+            "--format=csv,noheader,nounits",
+        ])
+        .output();
+
+    let Ok(output) = output else {
+        return Vec::new();
+    };
+    if !output.status.success() {
+        return Vec::new();
+    }
+
+    let text = String::from_utf8_lossy(&output.stdout);
+    let mut out = Vec::new();
+    for line in text.lines() {
+        if let Some(gpu) = parse_nvidia_smi_line(line) {
+            out.push(gpu);
+        }
+    }
+    out
+}
+
+fn parse_nvidia_smi_line(line: &str) -> Option<GpuSample> {
+    let parts = line.split(',').map(|s| s.trim()).collect::<Vec<_>>();
+    if parts.len() < 5 {
+        return None;
+    }
+
+    Some(GpuSample {
+        vendor: "NVIDIA".to_string(),
+        name: parts[0].to_string(),
+        utilization_percent: parse_f32(parts[1]),
+        mem_used_bytes: parse_u64(parts[2]).map(|v| v * 1024 * 1024),
+        mem_total_bytes: parse_u64(parts[3]).map(|v| v * 1024 * 1024),
+        temperature_c: parse_f32(parts[4]),
+    })
+}
+
+fn collect_amd_gpus() -> Vec<GpuSample> {
+    let mut out = Vec::new();
+    let drm_path = Path::new("/sys/class/drm");
+    let Ok(entries) = fs::read_dir(drm_path) else {
+        return out;
+    };
+
+    for entry in entries.flatten() {
+        let name = entry.file_name();
+        let name = name.to_string_lossy();
+        if !name.starts_with("card") || name.contains('-') {
+            continue;
+        }
+
+        let device_path = entry.path().join("device");
+        if !device_path.exists() || !is_amd_device(&device_path) {
+            continue;
+        }
+
+        let gpu_name = read_trimmed(device_path.join("product_name"))
+            .or_else(|| read_trimmed(device_path.join("vendor")))
+            .unwrap_or_else(|| "AMD GPU".to_string());
+
+        let utilization_percent =
+            read_trimmed(device_path.join("gpu_busy_percent")).and_then(|s| parse_f32(&s));
+        let mem_used_bytes =
+            read_trimmed(device_path.join("mem_info_vram_used")).and_then(|s| parse_u64(&s));
+        let mem_total_bytes =
+            read_trimmed(device_path.join("mem_info_vram_total")).and_then(|s| parse_u64(&s));
+        let temperature_c = read_amd_temp_c(&device_path);
+
+        out.push(GpuSample {
+            vendor: "AMD".to_string(),
+            name: gpu_name,
+            utilization_percent,
+            mem_used_bytes,
+            mem_total_bytes,
+            temperature_c,
+        });
+    }
+
+    out
+}
+
+fn is_amd_device(device_path: &Path) -> bool {
+    read_trimmed(device_path.join("vendor"))
+        .map(|vendor| vendor.eq_ignore_ascii_case("0x1002"))
+        .unwrap_or(false)
+}
+
+fn read_amd_temp_c(device_path: &Path) -> Option<f32> {
+    let hwmon_dir = device_path.join("hwmon");
+    let Ok(entries) = fs::read_dir(hwmon_dir) else {
+        return None;
+    };
+
+    for entry in entries.flatten() {
+        let p = entry.path().join("temp1_input");
+        if let Some(raw) = read_trimmed(p).and_then(|v| parse_f32(&v)) {
+            return Some(raw / 1000.0);
+        }
+    }
+    None
+}
+
+fn read_trimmed(path: PathBuf) -> Option<String> {
+    fs::read_to_string(path).ok().map(|s| s.trim().to_string())
+}
+
+fn parse_f32(s: &str) -> Option<f32> {
+    if s.eq_ignore_ascii_case("N/A") || s.is_empty() {
+        return None;
+    }
+    s.parse::<f32>().ok()
+}
+
+fn parse_u64(s: &str) -> Option<u64> {
+    if s.eq_ignore_ascii_case("N/A") || s.is_empty() {
+        return None;
+    }
+    s.parse::<u64>().ok()
+}
+
+#[cfg(test)]
+mod tests {
+    use super::*;
+
+    #[test]
+    fn parse_nvidia_line_works() {
+        let line = "NVIDIA GeForce RTX 3080, 47, 1234, 10240, 67";
+        let gpu = parse_nvidia_smi_line(line).expect("expected a parsed gpu");
+        assert_eq!(gpu.vendor, "NVIDIA");
+        assert_eq!(gpu.name, "NVIDIA GeForce RTX 3080");
+        assert_eq!(gpu.utilization_percent, Some(47.0));
+        assert_eq!(gpu.mem_total_bytes, Some(10_240 * 1024 * 1024));
+        assert_eq!(gpu.temperature_c, Some(67.0));
+    }
+}
diff --git a/src/state.rs b/src/state.rs
index 8e6d37d..0225105 100644
--- a/src/state.rs
+++ b/src/state.rs
@@ -36,6 +36,16 @@ pub struct ProcessSample {
     pub cpu_time_seconds: Option<u64>,
 }
 
+#[derive(Debug, Clone)]
+pub struct GpuSample {
+    pub vendor: String,
+    pub name: String,
+    pub utilization_percent: Option<f32>,
+    pub mem_used_bytes: Option<u64>,
+    pub mem_total_bytes: Option<u64>,
+    pub temperature_c: Option<f32>,
+}
+
 #[derive(Debug, Clone)]
 pub struct Snapshot {
     pub host: String,
@@ -54,6 +64,7 @@ pub struct Snapshot {
     pub processes: Vec<ProcessSample>,
     pub network_totals: Totals,
     pub network_interfaces: Vec<String>,
+    pub gpus: Vec<GpuSample>,
 }
 
 #[derive(Debug, Clone, Copy, PartialEq, Eq)]
diff --git a/src/ui.rs b/src/ui.rs
index 5e5d7f4..bdef191 100644
--- a/src/ui.rs
+++ b/src/ui.rs
@@ -100,15 +100,25 @@ fn draw_body(frame: &mut Frame, area: Rect, state: &AppState) {
         if state.show_process_details {
             let split = Layout::default()
                 .direction(Direction::Vertical)
-                .constraints([Constraint::Min(8), Constraint::Length(6)])
+                .constraints([
+                    Constraint::Min(8),
+                    Constraint::Length(5),
+                    Constraint::Length(6),
+                ])
                 .split(sidebar);
             draw_process_details(frame, split[0], state);
-            draw_network(frame, split[1], state);
+            draw_gpu(frame, split[1], state);
+            draw_network(frame, split[2], state);
         } else {
-            draw_network(frame, sidebar, state);
+            let split = Layout::default()
+                .direction(Direction::Vertical)
+                .constraints([Constraint::Length(5), Constraint::Min(4)])
+                .split(sidebar);
+            draw_gpu(frame, split[0], state);
+            draw_network(frame, split[1], state);
         }
     } else {
-        let net_line = compact_network_line(state);
+        let net_line = compact_status_line(state);
         let net_area = Rect {
             x: main.x,
             y: main.y,
@@ -325,6 +335,50 @@ fn draw_disks(frame: &mut Frame, area: Rect, state: &AppState) {
     }
 }
 
+fn draw_gpu(frame: &mut Frame, area: Rect, state: &AppState) {
+    let block = Block::default().title("GPU").borders(Borders::ALL);
+    let lines = if state.snapshot.gpus.is_empty() {
+        vec![Line::from("No GPU detected")]
+    } else {
+        state
+            .snapshot
+            .gpus
+            .iter()
+            .take(2)
+            .map(|gpu| {
+                let util = gpu
+                    .utilization_percent
+                    .map(|u| format!("{u:.0}%"))
+                    .unwrap_or_else(|| "-".to_string());
+                let mem = match (gpu.mem_used_bytes, gpu.mem_total_bytes) {
+                    (Some(used), Some(total)) => {
+                        format!("{}/{}", fmt_bytes_compact(used), fmt_bytes_compact(total))
+                    }
+                    _ => "-".to_string(),
+                };
+                Line::from(format!(
+                    "{} {} u:{} m:{} t:{}",
+                    truncate(&gpu.vendor, 4),
+                    truncate(&gpu.name, 10),
+                    util,
+                    mem,
+                    gpu.temperature_c
+                        .map(|t| format!("{t:.0}C"))
+                        .unwrap_or_else(|| "-".to_string())
+                ))
+            })
+            .collect::<Vec<_>>()
+    };
+
+    frame.render_widget(
+        Paragraph::new(lines)
+            .block(block)
+            .alignment(Alignment::Left)
+            .wrap(Wrap { trim: true }),
+        area,
+    );
+}
+
 fn draw_network(frame: &mut Frame, area: Rect, state: &AppState) {
     let block = Block::default().title("Net").borders(Borders::ALL);
 
@@ -594,13 +648,25 @@ fn sort_header(name: &str, col: ProcessSort, state: &AppState) -> String {
     }
 }
 
-fn compact_network_line(state: &AppState) -> String {
+fn compact_status_line(state: &AppState) -> String {
+    let gpu = state.snapshot.gpus.first().map_or_else(
+        || "GPU -".to_string(),
+        |g| {
+            let util = g
+                .utilization_percent
+                .map(|u| format!("{u:.0}%"))
+                .unwrap_or_else(|| "-".to_string());
+            format!("GPU {util}")
+        },
+    );
+
     format!(
-        "RX {}/s | TX {}/s | tot {}/{}",
+        "RX {}/s | TX {}/s | tot {}/{} | {}",
         fmt_bytes_compact(state.network_rate.rx_bps as u64),
         fmt_bytes_compact(state.network_rate.tx_bps as u64),
         fmt_bytes_compact(state.snapshot.network_totals.rx_bytes),
-        fmt_bytes_compact(state.snapshot.network_totals.tx_bytes)
+        fmt_bytes_compact(state.snapshot.network_totals.tx_bytes),
+        gpu
     )
 }