From 923cf9559459dc212d943173c9ab53599cbaeaed Mon Sep 17 00:00:00 2001
From: =?UTF-8?q?=E6=9D=A8=E5=87=AF=E4=BC=9F?= <21376218@buaa.edu.cn>
Date: Thu, 10 Sep 2026 12:01:32 +0800
Subject: [PATCH] target: add NVIDIA H20 support

---
 schedule/ir.py           | 25 +++++++++++++++++++++++++
 tests/test_h20_target.py | 25 +++++++++++++++++++++++++
 2 files changed, 50 insertions(+)
 create mode 100755 tests/test_h20_target.py

diff --git a/schedule/ir.py b/schedule/ir.py
index b63b31e..ed9a522 100644
--- a/schedule/ir.py
+++ b/schedule/ir.py
@@ -489,6 +489,31 @@ def measured_bandwidth_bound_us(self, weight_bytes: int) -> float:
         max_regs_per_thread=255, l2_bytes=50 * 1024 * 1024, hbm_bytes=192 * 1024**3,
         hbm_bandwidth_gbs=8000.0, fp16_tflops=2250.0, clock_ghz=1.86,
         supports_cooperative=True, wddm_tdr=False, note="Datacenter Blackwell, HBM3e. Spec only."),
+    "h20": GpuTarget(
+        name="h20",
+        sm_arch=90,
+        num_sms=78,
+        smem_bytes_per_sm=233472,
+        smem_bytes_per_block_optin=232448,
+        regs_per_sm=65536,
+        max_threads_per_sm=2048,
+        max_regs_per_thread=255,
+        l2_bytes=60 * 1024 * 1024,
+        hbm_bytes=96 * 1024**3,
+        hbm_bandwidth_gbs=4000.0,
+        fp16_tflops=148.0,
+        clock_ghz=0.0,
+        supports_cooperative=True,
+        wddm_tdr=False,
+        measured_bw_gbs=3712.3,
+        note=(
+            "NVIDIA H20, Hopper sm_90, 96GB HBM3. "
+            "78 SMs, 60 MiB L2, and SMEM limits observed on real H20 silicon. "
+            "Spec HBM bandwidth/FP16 throughput are retained as target parameters; "
+            "measured_bw_gbs=3712.3 from eval/peak_bandwidth.py "
+            "(median of three independent peak_gbs runs)."
+        ),
+    ),
     "h100": GpuTarget(
         name="h100", sm_arch=90, num_sms=132, smem_bytes_per_sm=233472,
         smem_bytes_per_block_optin=232448, regs_per_sm=65536, max_threads_per_sm=2048,
diff --git a/tests/test_h20_target.py b/tests/test_h20_target.py
new file mode 100755
index 0000000..070e60e
--- /dev/null
+++ b/tests/test_h20_target.py
@@ -0,0 +1,25 @@
+from schedule.ir import TARGETS
+
+
+def test_h20_target_profile():
+    target = TARGETS["h20"]
+
+    assert target.name == "h20"
+    assert target.sm_arch == 90
+    assert target.num_sms == 78
+
+    assert target.smem_bytes_per_sm == 233472
+    assert target.smem_bytes_per_block_optin == 232448
+
+    assert target.regs_per_sm == 65536
+    assert target.max_threads_per_sm == 2048
+    assert target.max_regs_per_thread == 255
+
+    assert target.l2_bytes == 60 * 1024 * 1024
+    assert target.hbm_bytes == 96 * 1024**3
+
+    assert target.hbm_bandwidth_gbs == 4000.0
+    assert target.measured_bw_gbs == 3712.3
+
+    assert target.supports_cooperative is True
+    assert target.wddm_tdr is False