Allreduce latency sweep across topologies and data sizes

Adds test_allreduce_latency_sweep that runs the existing intercube
allreduce kernel under three SIP topologies (ring_1d, torus_2d,
mesh_2d_no_wrap, all at n_sips=4) across 11 data sizes from 256 B/SIP
up to 1 MB/SIP. For each point, captures max(pe_exec_ns) — the
critical-path kernel time — and emits CSV plus log-x and linear-x
plots, both per-topology and combined overview, with KB/MB-formatted
tick labels. Reuses run_allreduce + _write_temp_configs and adds a
slot_size auto-bump when n_elem*2 exceeds the default IPCQ slot.

Sweep skips n_elem=16 because the runtime's dim_map scalar-arg
remapping (context.py:761) collides any int-valued kernel scalar that
matches a global tensor dim with its local shard size.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

docs/adr/ADR-0009-kernel-execution-messaging.md

@@ -67,6 +67,51 @@ Completion semantics:
 
 ---
 
+### D5. Launch timing is endpoint-synchronized
+
+All PEs targeted by a single kernel launch MUST begin executing the kernel
+body at the same simulated time, regardless of their dispatch path length
+from the launch entry point.
+
+Rationale. The dispatch tree Host → IO_CPU → M_CPU → PE_CPU has variable
+latency at every level. PEs near their M_CPU receive the launch earlier
+than PEs farther away; cubes near an IO_CPU receive it earlier than cubes
+farther away. Without synchronization, each PE's kernel begins at a
+different `env.now`, making per-PE metrics such as `pe_exec_ns` a function
+of dispatch-path geometry rather than of the kernel's behavior —
+producing measurement artifacts in benchmarks that time kernel-internal
+waits (for example `tl.recv` on cross-cube or cross-SIP hops).
+
+Mechanism.
+
+- `KernelLaunchMsg` carries an optional `target_start_ns: float | None`.
+- **IO_CPU** is the canonical stamper. On fan-out to M_CPUs, it
+  computes `target_start_ns = env.now + max_latency` where `max_latency`
+  is the maximum `ComponentContext.compute_path_latency_ns(path)` across
+  every target (sip, cube, pe) tuple — `path = find_node_path(io_cpu,
+  pe_cpu_id)`. The stamped value is placed on the request carried by
+  every fanned-out sub-Transaction.
+- **M_CPU** passes an already-stamped `target_start_ns` through
+  unchanged. Only when the value is absent (e.g. a direct
+  launch-to-M_CPU unit test) does M_CPU compute a per-cube barrier
+  `env.now + max(local command-path latency)`.
+- **PE_CPU** yields `env.timeout(target_start_ns - env.now)` at the top
+  of `_execute_kernel`, before recording `pe_exec_start` and invoking
+  the kernel body.
+- When `target_start_ns is None`, PE_CPU falls through to the legacy
+  unsynchronized behavior — preserving backward compatibility.
+
+IO_CPU-level stamping guarantees every PE across every targeted cube
+uses the same barrier sim-time, eliminating both the within-cube
+dispatch-offset artifact *and* the cross-cube offset artifact in
+multi-cube launches. Models a real-hardware timed-broadcast launch
+(latency-equalized dispatch tree).
+
+The synchronization is internal to the engine / IO_CPU / M_CPU / PE_CPU
+control plane — runtime API and application kernels are unchanged.
+
+---
+
 ## Links
 
 - SPEC R1, R2, R7, R8

docs/adr/ADR-0023-ipcq-pe-collective.en.md

@@ -420,11 +420,21 @@ fan-out (see `IpcqInitMsg` in D12).
 #### PE_DMA's added responsibility
 
 When `vc_comm` receives a token, PE_DMA processes it as the following
-**atomic** sequence. **No SimPy yield is allowed between the two steps**
+sequence: pay the Transaction's terminal BW drain, then atomically
-(invariant I6):
+write data and forward metadata. **No SimPy yield is allowed between
+the data write and the metadata forward** (invariant I6). The drain
+yield must sit before the atomic block, not inside it:
 
 ```python
-def _on_vc_comm_recv(self, env, token):
+def _on_vc_comm_recv(self, env, txn):
+    # Pay the terminal BW drain (nbytes / bottleneck_bw stamped by the
+    # sender PE_DMA). MUST happen before the atomic block so recv only
+    # wakes after the bytes have "landed".
+    drain = getattr(txn, "drain_ns", 0.0)
+    if drain > 0:
+        yield env.timeout(drain)
+
+    token = txn.request
     # ── ATOMIC: no yield between these two operations ──
     data = self._memory_store.read(token.src_space, token.src_addr,
                                    shape=..., dtype=...)
@@ -439,6 +449,33 @@ The final `put` is yieldable but uses an unbounded internal store, so
 it completes in a single step. That `put` is the closing call of the
 atomic block; nothing may be inserted before it.
 
+#### Drain-at-inbound semantics (D9 timing model)
+
+The Transaction carries `drain_ns = nbytes / bottleneck_bw_on_path`
+stamped at send-side PE_DMA. In this simulator per-hop `overhead_ns`
+is paid at each forwarding component via `run()`, and the remaining
+BW drain is paid once at the Transaction's terminal. Every non-IPCQ
+Transaction (raw DMA, kernel-launch fanout, etc.) pays this drain via
+`ComponentBase._forward_txn` at the terminal node. For IPCQ the
+destination PE_DMA intercepts the Transaction with `_handle_ipcq_inbound`
+(so IPCQ-specific data write + metadata forward can happen), so **the
+drain MUST be paid explicitly at the top of that handler** to keep
+IPCQ's timing model on par with every other fabric Transaction.
+
+Side-effects of paying drain here:
+
+- **SRC `tl.send`** is unchanged — fire-and-forget semantics are
+  preserved because the sender PE_DMA does not `yield sub_done`. The
+  `sub_done.succeed()` call (made after metadata forward below) is an
+  event with no listener on the sender side.
+- **DST `tl.recv`** unblocks `drain_ns` later. Since recv wakes only
+  when `IpcqMetaArrival` reaches its local PE_IPCQ, and the metadata
+  forward now happens after the drain, recv observes the full fabric
+  transfer time including bandwidth cost.
+
+Matches the physical picture: send dispatches and leaves; recv waits
+until the bytes have actually been drained into its inbox.
+
 ### D9.5. ADR-0020 (2-pass) integration
 
 `tl.send` / `tl.recv` integrates with ADR-0020's two-pass model. Phase

docs/adr/ADR-0023-ipcq-pe-collective.md

@@ -426,11 +426,22 @@ backend init에서 IpcqInitMsg fan-out 시 양방향 fast path channel을 함께
 
 #### PE_DMA의 책임 추가
 
-PE_DMA(vc_comm)는 token 수신 시 다음 atomic 시퀀스로 처리한다.
+PE_DMA(vc_comm)는 token 수신 시 다음 시퀀스로 처리한다: Transaction
-**두 동작 사이에 SimPy yield를 두어서는 안 된다** (I6 MUST 규칙 참조):
+terminal의 BW drain을 먼저 지불하고, 이어서 atomic하게 data write +
+metadata forward 수행. **data write와 metadata forward 사이에는 SimPy
+yield를 두어서는 안 된다** (I6 MUST 규칙 참조). drain yield는 atomic
+구간 안이 아니라 그 앞에 위치해야 한다:
 
 ```python
-def _on_vc_comm_recv(self, env, token):
+def _on_vc_comm_recv(self, env, txn):
+    # Sender PE_DMA가 찍어 둔 drain_ns (= nbytes / bottleneck_bw) 를
+    # 여기서 지불. atomic 구간보다 앞이어야 한다 — recv는 bytes가
+    # "도착"한 이후에만 깨어나야 하므로.
+    drain = getattr(txn, "drain_ns", 0.0)
+    if drain > 0:
+        yield env.timeout(drain)
+
+    token = txn.request
     # ── ATOMIC: 두 동작 사이에 yield 금지 ──
     # 1. data를 dst_addr에 write (dst의 메모리 공간은 token.dst_endpoint.buffer_kind)
     data = self._memory_store.read(token.src_space, token.src_addr,
@@ -446,6 +457,32 @@ wire로 capacity가 unbounded인 store를 사용하므로 즉시 완료된다 (
 single-step). 이 최종 put이 atomic 구간의 끝이며, 그 이전에 다른 yield가
 삽입되면 안 된다.
 
+#### Drain-at-inbound semantics (D9 timing model)
+
+Transaction은 sender PE_DMA가 `drain_ns = nbytes / bottleneck_bw_on_path`
+를 찍어 둔 상태로 fabric에 들어간다. 이 simulator에서 per-hop `overhead_ns`
+는 각 forwarding component의 `run()` 에서 지불되고, 남은 BW drain은
+Transaction의 terminal node에서 한 번 지불된다. IPCQ가 아닌 모든
+Transaction (raw DMA, kernel-launch fanout 등) 은
+`ComponentBase._forward_txn` 이 terminal에서 이 drain을 지불한다. IPCQ의
+경우 목적지 PE_DMA가 `_handle_ipcq_inbound` 핸들러로 Transaction을
+가로채서 (IPCQ 전용 data write + metadata forward를 해야 하므로)
+**이 핸들러 최상단에서 drain을 명시적으로 지불해야 한다** — 그래야 IPCQ의
+timing model이 다른 모든 fabric Transaction과 동일선상에 놓인다.
+
+여기서 drain을 지불할 때의 side-effect:
+
+- **SRC `tl.send`**: 동작 불변. sender PE_DMA가 `sub_done` 을 `yield`
+  하지 않으므로 fire-and-forget 의미가 보존된다. metadata forward 이후
+  호출되는 `sub_done.succeed()` 는 sender 입장에서 listener가 없는 이벤트.
+- **DST `tl.recv`**: `drain_ns` 만큼 늦게 깨어난다. recv는 local PE_IPCQ
+  의 `IpcqMetaArrival` 수신 시에만 wake되며, metadata forward가 drain
+  이후로 이동했으므로 recv는 bandwidth까지 포함한 전체 fabric transfer
+  시간을 관측하게 된다.
+
+물리적 그림과 일치: send는 dispatch하고 바로 반환; recv는 bytes가 실제로
+자신의 inbox로 drain될 때까지 대기.
+
 #### Backpressure latency 정확도
 
 backpressure 해제까지 걸리는 시간:

src/kernbench/components/builtin/io_cpu.py

@@ -58,7 +58,18 @@ class IoCpuComponent(ComponentBase):
             self._pending[key] = (expected, received, parent_done)
 
     def _dispatch_to_m_cpus(self, env: simpy.Environment, txn: Any) -> Generator:
-        """Fan out sub-Transactions to target cube M_CPUs, wait for responses."""
+        """Fan out sub-Transactions to target cube M_CPUs, wait for responses.
+
+        ADR-0009 D5 (extended): for KernelLaunchMsg, stamp a single global
+        target_start_ns = env.now + max(IO_CPU → any target PE_CPU path
+        latency across all target cubes). M_CPU passes this value through
+        unchanged; every PE in every cube yields until the same sim-time
+        before beginning kernel execution. Without this, cross-cube
+        launches would have each cube's M_CPU compute its own per-cube
+        barrier relative to its local env.now, leaving PEs on different
+        cubes out of sync (the "h3/h4 dispatch-offset artifact").
+        """
+        import dataclasses
         from kernbench.runtime_api.kernel import KernelLaunchMsg, MemoryReadMsg, MemoryWriteMsg
 
         request = txn.request
@@ -72,6 +83,36 @@ class IoCpuComponent(ComponentBase):
             txn.done.succeed()
             return
 
+        # For KernelLaunchMsg, compute the global barrier once here so
+        # every downstream PE_CPU uses the same target_start_ns.
+        if isinstance(request, KernelLaunchMsg):
+            global_max_latency = 0.0
+            pe_ids = self._resolve_pe_ids(
+                getattr(request, "target_pe", "all")
+            )
+            for sip, cube in cube_targets:
+                for pe_id in pe_ids:
+                    pe_cpu_id = (
+                        f"sip{sip}.cube{cube}.pe{pe_id}.pe_cpu"
+                    )
+                    try:
+                        path = self.ctx.router.find_node_path(
+                            self.node.id, pe_cpu_id,
+                        )
+                    except Exception:
+                        continue
+                    if len(path) < 2:
+                        continue
+                    latency = self.ctx.compute_path_latency_ns(
+                        path, nbytes=0,
+                    )
+                    if latency > global_max_latency:
+                        global_max_latency = latency
+            request = dataclasses.replace(
+                request,
+                target_start_ns=float(env.now) + global_max_latency,
+            )
+
         # Setup aggregation
         self._pending[request.request_id] = (len(cube_targets), 0, txn.done)
 
@@ -91,6 +132,19 @@ class IoCpuComponent(ComponentBase):
             )
             yield self.out_ports[path[1]].put(sub_txn.advance())
 
+    def _resolve_pe_ids(self, target_pe: Any) -> list[int]:
+        """Resolve target_pe → list of PE indices (mirrors M_CPU logic)."""
+        if isinstance(target_pe, int):
+            return [target_pe]
+        if isinstance(target_pe, tuple):
+            return list(target_pe)
+        # "all": all PEs in a cube
+        n_slices = 8
+        if self.ctx and self.ctx.spec:
+            mm = self.ctx.spec.get("cube", {}).get("memory_map", {})
+            n_slices = mm.get("hbm_slices_per_cube", 8)
+        return list(range(n_slices))
+
     def _resolve_cube_targets(self, request: Any) -> list[tuple[int, int]]:
         """Return list of (sip, cube) pairs to fan out to."""
         from kernbench.runtime_api.kernel import (

src/kernbench/components/builtin/m_cpu.py

@@ -162,7 +162,11 @@ class MCpuComponent(ComponentBase):
         Routes through find_node_path (M_CPU → NOC → PE_CPU command edges).
         PE_CPU sends ResponseMsg back via NOC → M_CPU on completion.
         Then sends aggregate ResponseMsg back to IO_CPU on the reverse path.
+
+        ADR-0009 D5: stamps target_start_ns so every PE in this fanout
+        starts executing at the same env.now regardless of dispatch path.
         """
+        import dataclasses
         request = txn.request
         target_pe = getattr(request, "target_pe", "all")
         cube_prefix = self.node.id.rsplit(".", 1)[0]  # e.g. "sip0.cube0"
@@ -172,9 +176,13 @@ class MCpuComponent(ComponentBase):
             txn.done.succeed()
             return
 
-        # Fan out to each PE_CPU, using response-based aggregation
+        # Resolve per-PE paths. If IO_CPU already stamped a global
-        sub_txns: list[Transaction] = []
+        # target_start_ns (ADR-0009 D5 extended), pass it through
-        n_dispatched = 0
+        # unchanged so every PE across every cube uses the same barrier.
+        # Otherwise (e.g. direct-to-M_CPU launch in a unit test) compute
+        # a per-cube barrier from env.now.
+        per_pe: list[tuple[int, list[str], float]] = []
+        max_latency = 0.0
         for pe_id in pe_ids:
             pe_cpu_id = f"{cube_prefix}.pe{pe_id}.pe_cpu"
             try:
@@ -183,8 +191,24 @@ class MCpuComponent(ComponentBase):
                 continue
             if len(path) < 2:
                 continue
+            latency = self.ctx.compute_path_latency_ns(path, nbytes=0)
+            per_pe.append((pe_id, path, latency))
+            if latency > max_latency:
+                max_latency = latency
+
+        if getattr(request, "target_start_ns", None) is not None:
+            stamped_request = request
+        else:
+            stamped_request = dataclasses.replace(
+                request, target_start_ns=float(env.now) + max_latency,
+            )
+
+        # Fan out to each PE_CPU, using response-based aggregation
+        sub_txns: list[Transaction] = []
+        n_dispatched = 0
+        for pe_id, path, _lat in per_pe:
             sub_txn = Transaction(
-                request=request, path=path, step=0,
+                request=stamped_request, path=path, step=0,
                 nbytes=0, done=env.event(),
             )
             yield self.out_ports[path[1]].put(sub_txn.advance())
@@ -204,16 +228,21 @@ class MCpuComponent(ComponentBase):
         yield all_done
         del self._parent_txns[request.request_id]
 
-        # Aggregate PE-internal metrics (max across PEs)
+        # Aggregate PE-internal metrics (max across PEs and across cubes).
+        # Multiple M_CPUs share the same result_data dict via IO_CPU fanout;
+        # merge against the existing value so cubes don't clobber each other.
         pe_exec_values = [st.result_data.get("pe_exec_ns", 0.0) for st in sub_txns]
         if pe_exec_values:
-            txn.result_data["pe_exec_ns"] = max(pe_exec_values)
+            cur = txn.result_data.get("pe_exec_ns", 0.0) or 0.0
+            txn.result_data["pe_exec_ns"] = max(cur, max(pe_exec_values))
         dma_values = [st.result_data.get("dma_ns", 0.0) for st in sub_txns]
         if dma_values:
-            txn.result_data["dma_ns"] = max(dma_values)
+            cur = txn.result_data.get("dma_ns", 0.0) or 0.0
+            txn.result_data["dma_ns"] = max(cur, max(dma_values))
         compute_values = [st.result_data.get("compute_ns", 0.0) for st in sub_txns]
         if compute_values:
-            txn.result_data["compute_ns"] = max(compute_values)
+            cur = txn.result_data.get("compute_ns", 0.0) or 0.0
+            txn.result_data["compute_ns"] = max(cur, max(compute_values))
 
         # Send aggregate response on reverse command path back to IO_CPU
         reverse_path = list(reversed(txn.path))

src/kernbench/components/builtin/pe_cpu.py

@@ -95,6 +95,13 @@ class PeCpuComponent(ComponentBase):
         request = txn.request
         yield from self.run(env, 0)
 
+        # ADR-0009 D5: synchronized launch barrier. If M_CPU stamped a
+        # target_start_ns, wait until then so every PE in this launch
+        # begins pe_exec measurement at the same simulated time.
+        target_start = getattr(request, "target_start_ns", None)
+        if target_start is not None and target_start > env.now:
+            yield env.timeout(float(target_start) - env.now)
+
         kernel_fn = get_kernel(request.kernel_ref.name)
         num_programs = self._derive_num_programs(request)
         kernel_args = self._unpack_kernel_args(request)

src/kernbench/components/builtin/pe_dma.py

@@ -186,13 +186,37 @@ class PeDmaComponent(PeEngineBase):
     # ── IPCQ inbound (fabric → PE_DMA → MemoryStore + PE_IPCQ) ──────
 
     def _handle_ipcq_inbound(self, env: simpy.Environment, txn: Any) -> Generator:
-        """At destination PE_DMA: atomically write data and forward metadata.
+        """At destination PE_DMA: pay terminal drain, then atomically write
+        data and forward metadata.
+
+        ADR-0023 D9 (drain at inbound terminal): the Transaction carries
+        ``drain_ns = nbytes / bottleneck_bw_on_path`` stamped by the sender
+        PE_DMA. Like every other Transaction terminal in the simulator (see
+        ``ComponentBase._forward_txn``), this drain must be paid when the
+        Transaction reaches its destination. SRC-side ``tl.send`` is
+        fire-and-forget — it never yields on ``sub_done`` — so paying the
+        drain here does NOT delay the sender. What it DOES delay is the
+        IpcqMetaArrival forwarded below: that delay is the only signal
+        ``tl.recv`` on DST blocks on, which is exactly the desired
+        semantics — "send dispatches and returns; recv waits until the
+        bytes have actually landed in its inbox".
+
+        The drain MUST be paid before the atomic block — inserting a yield
+        inside would break invariant I6.
 
         I6 (MUST): no SimPy yield between MemoryStore.write and the
         IpcqMetaArrival put into PE_IPCQ.
         """
         from kernbench.common.ipcq_types import IpcqMetaArrival
 
+        # Pay terminal BW drain before the atomic write/metadata forward.
+        # Without this, IPCQ effectively got fabric bandwidth for free at
+        # the terminal (only intermediate-hop overhead_ns was charged),
+        # making IPCQ lower than raw DMA at large sizes in benchmarks.
+        drain = getattr(txn, "drain_ns", 0.0)
+        if drain > 0:
+            yield env.timeout(drain)
+
         token = txn.request
 
         # ── ATOMIC: do not introduce yield between these two operations ──

src/kernbench/components/context.py

@@ -26,6 +26,9 @@ class ComponentContext:
     spec: dict = field(default_factory=dict)  # topology spec (cube layout, PE count, etc.)
     memory_store: Any = None  # MemoryStore for Phase 1 data-aware execution (ADR-0020)
     op_logger: Any = None     # OpLogger for Phase 1 op recording (ADR-0020)
+    # node_id -> overhead_ns (ADR-0009 D5: used by M_CPU to compute per-PE
+    # dispatch latency when stamping target_start_ns on KernelLaunchMsg).
+    node_overhead_ns: dict[str, float] = field(default_factory=dict)
 
     def get_shared_resource(
         self, env: simpy.Environment, key: str, capacity: int = 1,
@@ -52,3 +55,19 @@ class ComponentContext:
         if min_bw == float("inf"):
             return 0.0
         return nbytes / min_bw
+
+    def compute_path_latency_ns(self, path: list[str], nbytes: int = 0) -> float:
+        """Formula latency along path: wire + per-node overhead + drain.
+
+        ADR-0009 D5: M_CPU uses this to compute per-PE dispatch latency
+        when stamping target_start_ns on KernelLaunchMsg fanout.
+        """
+        total = 0.0
+        for i in range(len(path) - 1):
+            edge = self.edge_map.get((path[i], path[i + 1]))
+            if edge:
+                total += edge.distance_mm * self.ns_per_mm
+        for node_id in path:
+            total += self.node_overhead_ns.get(node_id, 0.0)
+        total += self.compute_drain_ns(path, nbytes)
+        return total

src/kernbench/components/legacy/builtin/io_cpu.py

@@ -58,7 +58,13 @@ class IoCpuComponent(ComponentBase):
             self._pending[key] = (expected, received, parent_done)
 
     def _dispatch_to_m_cpus(self, env: simpy.Environment, txn: Any) -> Generator:
-        """Fan out sub-Transactions to target cube M_CPUs, wait for responses."""
+        """Fan out sub-Transactions to target cube M_CPUs, wait for responses.
+
+        ADR-0009 D5 (extended): stamp a global target_start_ns on
+        KernelLaunchMsg so every PE across every target cube starts at
+        the same env.now. See the non-legacy builtin for full rationale.
+        """
+        import dataclasses
         from kernbench.runtime_api.kernel import KernelLaunchMsg, MemoryReadMsg, MemoryWriteMsg
 
         request = txn.request
@@ -72,6 +78,34 @@ class IoCpuComponent(ComponentBase):
             txn.done.succeed()
             return
 
+        if isinstance(request, KernelLaunchMsg):
+            global_max_latency = 0.0
+            pe_ids = self._resolve_pe_ids(
+                getattr(request, "target_pe", "all")
+            )
+            for sip, cube in cube_targets:
+                for pe_id in pe_ids:
+                    pe_cpu_id = (
+                        f"sip{sip}.cube{cube}.pe{pe_id}.pe_cpu"
+                    )
+                    try:
+                        path = self.ctx.router.find_node_path(
+                            self.node.id, pe_cpu_id,
+                        )
+                    except Exception:
+                        continue
+                    if len(path) < 2:
+                        continue
+                    latency = self.ctx.compute_path_latency_ns(
+                        path, nbytes=0,
+                    )
+                    if latency > global_max_latency:
+                        global_max_latency = latency
+            request = dataclasses.replace(
+                request,
+                target_start_ns=float(env.now) + global_max_latency,
+            )
+
         # Setup aggregation
         self._pending[request.request_id] = (len(cube_targets), 0, txn.done)
 
@@ -91,6 +125,18 @@ class IoCpuComponent(ComponentBase):
             )
             yield self.out_ports[path[1]].put(sub_txn.advance())
 
+    def _resolve_pe_ids(self, target_pe: Any) -> list[int]:
+        """Resolve target_pe → list of PE indices (mirrors M_CPU logic)."""
+        if isinstance(target_pe, int):
+            return [target_pe]
+        if isinstance(target_pe, tuple):
+            return list(target_pe)
+        n_slices = 8
+        if self.ctx and self.ctx.spec:
+            mm = self.ctx.spec.get("cube", {}).get("memory_map", {})
+            n_slices = mm.get("hbm_slices_per_cube", 8)
+        return list(range(n_slices))
+
     def _resolve_cube_targets(self, request: Any) -> list[tuple[int, int]]:
         """Return list of (sip, cube) pairs to fan out to."""
         from kernbench.runtime_api.kernel import (

src/kernbench/components/legacy/builtin/m_cpu.py

@@ -162,7 +162,11 @@ class MCpuComponent(ComponentBase):
         Routes through find_node_path (M_CPU → NOC → PE_CPU command edges).
         PE_CPU sends ResponseMsg back via NOC → M_CPU on completion.
         Then sends aggregate ResponseMsg back to IO_CPU on the reverse path.
+
+        ADR-0009 D5: stamps target_start_ns so every PE in this fanout
+        starts executing at the same env.now regardless of dispatch path.
         """
+        import dataclasses
         request = txn.request
         target_pe = getattr(request, "target_pe", "all")
         cube_prefix = self.node.id.rsplit(".", 1)[0]  # e.g. "sip0.cube0"
@@ -172,9 +176,10 @@ class MCpuComponent(ComponentBase):
             txn.done.succeed()
             return
 
-        # Fan out to each PE_CPU, using response-based aggregation
+        # Resolve per-PE paths. If IO_CPU already stamped a global
-        sub_txns: list[Transaction] = []
+        # target_start_ns (ADR-0009 D5 extended), pass it through.
-        n_dispatched = 0
+        per_pe: list[tuple[int, list[str], float]] = []
+        max_latency = 0.0
         for pe_id in pe_ids:
             pe_cpu_id = f"{cube_prefix}.pe{pe_id}.pe_cpu"
             try:
@@ -183,8 +188,24 @@ class MCpuComponent(ComponentBase):
                 continue
             if len(path) < 2:
                 continue
+            latency = self.ctx.compute_path_latency_ns(path, nbytes=0)
+            per_pe.append((pe_id, path, latency))
+            if latency > max_latency:
+                max_latency = latency
+
+        if getattr(request, "target_start_ns", None) is not None:
+            stamped_request = request
+        else:
+            stamped_request = dataclasses.replace(
+                request, target_start_ns=float(env.now) + max_latency,
+            )
+
+        # Fan out to each PE_CPU, using response-based aggregation
+        sub_txns: list[Transaction] = []
+        n_dispatched = 0
+        for pe_id, path, _lat in per_pe:
             sub_txn = Transaction(
-                request=request, path=path, step=0,
+                request=stamped_request, path=path, step=0,
                 nbytes=0, done=env.event(),
             )
             yield self.out_ports[path[1]].put(sub_txn.advance())
@@ -204,16 +225,21 @@ class MCpuComponent(ComponentBase):
         yield all_done
         del self._parent_txns[request.request_id]
 
-        # Aggregate PE-internal metrics (max across PEs)
+        # Aggregate PE-internal metrics (max across PEs and across cubes).
+        # Multiple M_CPUs share the same result_data dict via IO_CPU fanout;
+        # merge against the existing value so cubes don't clobber each other.
         pe_exec_values = [st.result_data.get("pe_exec_ns", 0.0) for st in sub_txns]
         if pe_exec_values:
-            txn.result_data["pe_exec_ns"] = max(pe_exec_values)
+            cur = txn.result_data.get("pe_exec_ns", 0.0) or 0.0
+            txn.result_data["pe_exec_ns"] = max(cur, max(pe_exec_values))
         dma_values = [st.result_data.get("dma_ns", 0.0) for st in sub_txns]
         if dma_values:
-            txn.result_data["dma_ns"] = max(dma_values)
+            cur = txn.result_data.get("dma_ns", 0.0) or 0.0
+            txn.result_data["dma_ns"] = max(cur, max(dma_values))
         compute_values = [st.result_data.get("compute_ns", 0.0) for st in sub_txns]
         if compute_values:
-            txn.result_data["compute_ns"] = max(compute_values)
+            cur = txn.result_data.get("compute_ns", 0.0) or 0.0
+            txn.result_data["compute_ns"] = max(cur, max(compute_values))
 
         # Send aggregate response on reverse command path back to IO_CPU
         reverse_path = list(reversed(txn.path))

src/kernbench/components/legacy/builtin/pe_cpu.py

@@ -71,6 +71,13 @@ class PeCpuComponent(ComponentBase):
         request = txn.request
         yield from self.run(env, 0)
 
+        # ADR-0009 D5: synchronized launch barrier. If M_CPU stamped a
+        # target_start_ns, wait until then so every PE in this launch
+        # begins pe_exec measurement at the same simulated time.
+        target_start = getattr(request, "target_start_ns", None)
+        if target_start is not None and target_start > env.now:
+            yield env.timeout(float(target_start) - env.now)
+
         kernel_fn = get_kernel(request.kernel_ref.name)
         num_programs = self._derive_num_programs(request)
         kernel_args = self._unpack_kernel_args(request)

src/kernbench/policy/address/pe_mmu.py

@@ -19,7 +19,14 @@ class PageFault(Exception):
 
 
 class PeMMU:
-    """Per-PE MMU with page-aligned VA→PA translation table.
+    """Per-PE MMU with sub-page-capable VA→PA translation table.
+
+    Each page-table entry is a list of (start_in_page, end_in_page,
+    pa_at_offset_zero) regions. This is a SIMULATOR STOPGAP — real MMUs
+    store one PA per page-table entry. Sub-page regions exist here so
+    DPPolicy layouts that shard below page granularity (e.g. 128 B
+    payloads with 4 KB pages) don't silently mis-route through last-
+    write-wins overwrites. Memory note: project_mmu_subpage_stopgap.md.
 
     Args:
         page_size: Page size in bytes (default 2 MB).
@@ -34,7 +41,11 @@ class PeMMU:
         self._page_size = page_size
         self._page_shift = (page_size - 1).bit_length()
         self._page_mask = page_size - 1
-        self._table: dict[int, int] = {}  # va_page_number → pa_page_base
+        # vpn → list of (start_in_page, end_in_page, pa_at_offset_zero).
+        # pa_at_offset_zero is the PA that offset 0 of the page would map
+        # to under this region — i.e. translate(off) = pa_at_offset_zero
+        # + off when start <= off < end.
+        self._table: dict[int, list[tuple[int, int, int]]] = {}
         self._overhead_ns = overhead_ns
 
     @property
@@ -46,21 +57,67 @@ class PeMMU:
         return len(self._table)
 
     def map(self, va: int, pa: int, size: int) -> None:
-        """Register VA→PA mapping for a contiguous range."""
+        """Register VA→PA mapping for a contiguous range.
-        for off in range(0, size, self._page_size):
+
-            vpn = (va + off) >> self._page_shift
+        Sub-page-aware: a single page can hold multiple disjoint regions,
-            self._table[vpn] = pa + off
+        each pointing to a different PA. Later map() calls APPEND a new
+        region; on overlap with an existing region, the new region wins
+        for the overlapping offsets (translate iterates in reverse so the
+        last write takes precedence — matches legacy single-PA behavior
+        when a full page is re-mapped).
+        """
+        end_va = va + size
+        cur = va
+        while cur < end_va:
+            vpn = cur >> self._page_shift
+            page_base_va = vpn << self._page_shift
+            page_end_va = page_base_va + self._page_size
+            region_start = cur - page_base_va
+            region_end = min(end_va, page_end_va) - page_base_va
+            # PA seen at offset 0 of page if this region's mapping covered it
+            pa_at_offset_zero = pa + (cur - va) - region_start
+            self._table.setdefault(vpn, []).append(
+                (region_start, region_end, pa_at_offset_zero)
+            )
+            cur = page_base_va + region_end
 
     def unmap(self, va: int, size: int) -> None:
-        """Remove VA mapping for a contiguous range."""
+        """Remove VA mapping for a contiguous range.
-        for off in range(0, size, self._page_size):
+
-            vpn = (va + off) >> self._page_shift
+        Drops any region whose extent is contained within the unmapped
-            self._table.pop(vpn, None)
+        range. Partial overlaps (region straddles the range boundary)
+        are left in place — caller is expected to unmap on the same
+        boundaries it mapped on.
+        """
+        end_va = va + size
+        cur = va
+        while cur < end_va:
+            vpn = cur >> self._page_shift
+            page_base_va = vpn << self._page_shift
+            page_end_va = page_base_va + self._page_size
+            unmap_start = cur - page_base_va
+            unmap_end = min(end_va, page_end_va) - page_base_va
+            regions = self._table.get(vpn)
+            if regions is not None:
+                kept = [
+                    r for r in regions
+                    if not (r[0] >= unmap_start and r[1] <= unmap_end)
+                ]
+                if kept:
+                    self._table[vpn] = kept
+                else:
+                    del self._table[vpn]
+            cur = page_base_va + unmap_end
 
     def translate(self, va: int) -> int:
         """Translate VA to PA. Raises PageFault if unmapped."""
         vpn = va >> self._page_shift
-        pa_page_base = self._table.get(vpn)
+        regions = self._table.get(vpn)
-        if pa_page_base is None:
+        if regions is None:
+            raise PageFault(va)
+        offset = va & self._page_mask
+        # Iterate latest-first so newer map() calls win on overlap
+        for start, end, pa_at_offset_zero in reversed(regions):
+            if start <= offset < end:
+                return pa_at_offset_zero + offset
         raise PageFault(va)
-        return pa_page_base + (va & self._page_mask)

src/kernbench/runtime_api/kernel.py

@@ -90,6 +90,11 @@ class KernelLaunchMsg:
     args: tuple[KernelArg, ...]
     target_cubes: tuple[int, ...] | Literal["all"] = "all"
     target_pe: int | tuple[int, ...] | Literal["all"] = "all"
+    # ADR-0009 D5: synchronized kernel start. When set, each PE_CPU yields
+    # until env.now >= target_start_ns before beginning kernel execution,
+    # so every PE in a launch starts at the same simulated time regardless
+    # of its M_CPU dispatch path length. Stamped by M_CPU fan-out.
+    target_start_ns: float | None = None
     msg_type: Literal["kernel_launch"] = "kernel_launch"
 
 

src/kernbench/sim_engine/engine.py

@@ -67,6 +67,10 @@ class GraphEngine:
             spec=graph.spec,
             memory_store=self._memory_store,
             op_logger=self._op_logger,
+            node_overhead_ns={
+                nid: float(n.attrs.get("overhead_ns", 0.0))
+                for nid, n in graph.nodes.items()
+            },
         )
         self._components: dict[str, ComponentBase] = {
             node_id: ComponentRegistry.create(node, overrides, ctx)

tests/test_allreduce_multidevice.py

@@ -179,7 +179,9 @@ CONFIGS = [
 ]
 
 
-def _write_temp_configs(tmp_path, sip_topology, n_sips, algorithm):
+def _write_temp_configs(
+    tmp_path, sip_topology, n_sips, algorithm, n_elem_override=None,
+):
     """Write temp topology.yaml and ccl.yaml with the given overrides."""
     with open(TOPOLOGY_PATH) as f:
         topo_cfg = yaml.safe_load(f)
@@ -193,6 +195,15 @@ def _write_temp_configs(tmp_path, sip_topology, n_sips, algorithm):
     with open(ccl_path) as f:
         ccl_cfg = yaml.safe_load(f)
     ccl_cfg["defaults"]["algorithm"] = algorithm
+    if n_elem_override is not None:
+        ccl_cfg.setdefault("algorithms", {}).setdefault(
+            algorithm, {},
+        )["n_elem"] = int(n_elem_override)
+        # Ensure IPCQ slot is big enough for the per-message payload.
+        per_msg_bytes = int(n_elem_override) * 2  # f16
+        default_slot = int(ccl_cfg["defaults"].get("slot_size", 4096))
+        if per_msg_bytes > default_slot:
+            ccl_cfg["defaults"]["slot_size"] = per_msg_bytes
     tmp_ccl = tmp_path / "ccl.yaml"
     with open(tmp_ccl, "w") as f:
         yaml.dump(ccl_cfg, f, default_flow_style=False)
@@ -220,3 +231,191 @@ def test_allreduce(tmp_path, algorithm, sip_topology, n_sips):
             algorithm=algorithm, ccl_yaml=ccl_path,
         )
         assert result["ok_cubes"] > 0
+
+
+# ── Latency sweep ─────────────────────────────────────────────────────
+
+# avoid 16 (== n_cubes, dim_map collision). Goes up to 1 MB per SIP:
+# bytes_per_sip = n_cubes * n_elem * 2 = 32 * n_elem.
+_SWEEP_N_ELEM = [
+    8, 32, 64, 128, 512, 1024, 2048,
+    4096, 8192, 16384, 32768,
+]
+_ELEM_BYTES_F16 = 2
+
+
+def test_allreduce_latency_sweep(tmp_path):
+    """Sweep n_elem across each SIP topology; record max(pe_exec_ns)
+    as the critical-path kernel latency. Emits CSV + PNG plots to
+    tests/allreduce_latency_plots/.
+    """
+    import csv
+
+    import matplotlib.pyplot as plt
+    from matplotlib.ticker import FuncFormatter
+
+    def _fmt_bytes(x, _pos):
+        """Format tick as B / KB / MB."""
+        if x <= 0:
+            return "0"
+        if x >= 1024 * 1024:
+            return f"{x / (1024 * 1024):.0f} MB"
+        if x >= 1024:
+            return f"{x / 1024:.0f} KB"
+        return f"{x:.0f} B"
+
+    _bytes_fmt = FuncFormatter(_fmt_bytes)
+
+    out_dir = Path(__file__).parent / "allreduce_latency_plots"
+    out_dir.mkdir(parents=True, exist_ok=True)
+    records: list[dict] = []
+
+    # Apples-to-apples: same n_sips across all three topologies.
+    for algorithm, sip_topology, n_sips in [
+        ("intercube_allreduce", "ring_1d", 4),
+        ("intercube_allreduce", "torus_2d", 4),
+        ("intercube_allreduce", "mesh_2d_no_wrap", 4),
+    ]:
+        for n_elem in _SWEEP_N_ELEM:
+            sub = tmp_path / f"{sip_topology}_{n_elem}"
+            sub.mkdir()
+            topo_path, ccl_path = _write_temp_configs(
+                sub, sip_topology, n_sips, algorithm,
+                n_elem_override=n_elem,
+            )
+            topo = resolve_topology(topo_path)
+            engine = GraphEngine(topo.topology_obj, enable_data=True)
+            spec = topo.topology_obj.spec
+
+            with RuntimeContext(
+                engine=engine,
+                target_device=DeviceSelector("all"),
+                correlation_id=f"sweep_{algorithm}_{sip_topology}_{n_elem}",
+                spec=spec,
+            ) as ctx:
+                result = run_allreduce(
+                    ctx, engine, spec,
+                    algorithm=algorithm, ccl_yaml=ccl_path,
+                )
+                assert result["ok_cubes"] > 0
+
+            pe_exec_vals = [
+                float(tr.get("pe_exec_ns", 0.0) or 0.0)
+                for _, (_, tr) in engine._results.items()
+                if isinstance(tr, dict)
+            ]
+            crit_ns = max(pe_exec_vals) if pe_exec_vals else 0.0
+
+            cm = spec["sip"]["cube_mesh"]
+            n_cubes = int(cm["w"]) * int(cm["h"])
+            bytes_per_sip = n_cubes * n_elem * _ELEM_BYTES_F16
+            # pe="replicate" + num_pes=1 → one active PE per cube owns
+            # the whole cube row. Per-PE bytes == per-cube-tile bytes ==
+            # per-message bytes over the IPCQ fabric.
+            bytes_per_pe = n_elem * _ELEM_BYTES_F16
+
+            records.append({
+                "algorithm": algorithm,
+                "sip_topology": sip_topology,
+                "n_sips": n_sips,
+                "n_elem": n_elem,
+                "bytes_per_pe": bytes_per_pe,
+                "bytes_per_sip": bytes_per_sip,
+                "latency_ns": crit_ns,
+            })
+            print(
+                f"[{sip_topology:<16} n_sips={n_sips} n_elem={n_elem:>5}  "
+                f"bytes/pe={bytes_per_pe:>7}  bytes/sip={bytes_per_sip:>9}]  "
+                f"pe_exec_max = {crit_ns:8.1f} ns"
+            )
+
+    with open(out_dir / "summary.csv", "w", newline="", encoding="utf-8") as f:
+        w = csv.DictWriter(f, fieldnames=[
+            "algorithm", "sip_topology", "n_sips", "n_elem",
+            "bytes_per_pe", "bytes_per_sip", "latency_ns",
+        ])
+        w.writeheader()
+        for r in records:
+            w.writerow(r)
+
+    topologies = sorted({r["sip_topology"] for r in records})
+    # Per-topology plots: log-scale + linear-scale side-by-side.
+    # X-axis = bytes per PE (per-message payload size).
+    for topo_name in topologies:
+        rs = sorted(
+            [r for r in records if r["sip_topology"] == topo_name],
+            key=lambda r: r["bytes_per_pe"],
+        )
+        xs = [r["bytes_per_pe"] for r in rs]
+        ys = [r["latency_ns"] for r in rs]
+        title = (
+            f"Allreduce latency — {topo_name} "
+            f"(n_sips={rs[0]['n_sips']})"
+        )
+        # Log-scale
+        fig, ax = plt.subplots(figsize=(8, 5))
+        ax.plot(xs, ys, marker="o", color="tab:blue")
+        ax.set_xscale("log", base=2)
+        ax.set_xlabel("Bytes per PE (log scale)")
+        ax.set_ylabel("max pe_exec_ns (critical path)")
+        ax.set_title(title)
+        ax.grid(True, alpha=0.3)
+        ax.xaxis.set_major_formatter(_bytes_fmt)
+        fig.tight_layout()
+        fig.savefig(out_dir / f"{topo_name}.png", dpi=120)
+        plt.close(fig)
+        # Linear-scale companion
+        fig, ax = plt.subplots(figsize=(8, 5))
+        ax.plot(xs, ys, marker="o", color="tab:blue")
+        ax.set_xlabel("Bytes per PE")
+        ax.set_ylabel("max pe_exec_ns (critical path)")
+        ax.set_title(title + " [linear scale]")
+        ax.grid(True, alpha=0.3)
+        ax.xaxis.set_major_formatter(_bytes_fmt)
+        fig.tight_layout()
+        fig.savefig(out_dir / f"{topo_name}_linear.png", dpi=120)
+        plt.close(fig)
+
+    # Combined overview — two variants: log-scale (overview.png) and
+    # linear-scale (overview_linear.png).
+    colors = {"ring_1d": "tab:blue", "torus_2d": "tab:orange",
+              "mesh_2d_no_wrap": "tab:green"}
+
+    def _draw_overview(log_x: bool, filename: str, title_suffix: str) -> None:
+        fig, ax = plt.subplots(figsize=(9, 6))
+        for topo_name in topologies:
+            rs = sorted(
+                [r for r in records if r["sip_topology"] == topo_name],
+                key=lambda r: r["bytes_per_pe"],
+            )
+            ax.plot(
+                [r["bytes_per_pe"] for r in rs],
+                [r["latency_ns"] for r in rs],
+                marker="o",
+                label=f"{topo_name} (n_sips={rs[0]['n_sips']})",
+                color=colors.get(topo_name),
+            )
+        if log_x:
+            ax.set_xscale("log", base=2)
+            ax.set_xlabel("Bytes per PE (log scale)")
+        else:
+            ax.set_xlabel("Bytes per PE")
+        ax.set_ylabel("max pe_exec_ns (critical path)")
+        ax.set_title("Multi-device allreduce latency by topology" + title_suffix)
+        ax.grid(True, alpha=0.3)
+        ax.legend()
+        ax.xaxis.set_major_formatter(_bytes_fmt)
+        fig.tight_layout()
+        fig.savefig(out_dir / filename, dpi=120)
+        plt.close(fig)
+
+    _draw_overview(log_x=True, filename="overview.png", title_suffix="")
+    _draw_overview(
+        log_x=False, filename="overview_linear.png",
+        title_suffix=" [linear scale]",
+    )
+
+    print(
+        f"\nWrote {out_dir / 'overview.png'} + "
+        f"{out_dir / 'overview_linear.png'}"
+    )

tests/test_kernel_launch_sync.py

@@ -0,0 +1,62 @@
+"""ADR-0009 D5: synchronized launch barrier.
+
+M_CPU stamps KernelLaunchMsg with target_start_ns = env.now + max path
+latency; PE_CPU yields until that time before recording pe_exec_start.
+Every PE in a single launch MUST begin kernel execution at the same
+env.now regardless of its dispatch path length.
+
+We verify this indirectly: for a no-op kernel, pe_exec_ns = env.now -
+pe_exec_start. If every PE's pe_exec_start is identical and every PE
+runs the same no-op body, every pe_exec_ns value must be identical.
+Without D5, pe_exec_start varies by dispatch-path length and so does
+pe_exec_ns.
+"""
+from __future__ import annotations
+
+from pathlib import Path
+
+import numpy as np
+
+from kernbench.policy.placement.dp import DPPolicy
+from kernbench.runtime_api.context import RuntimeContext
+from kernbench.runtime_api.types import DeviceSelector
+from kernbench.sim_engine.engine import GraphEngine
+from kernbench.topology.builder import resolve_topology
+
+TOPOLOGY_PATH = Path(__file__).parent.parent / "topology.yaml"
+
+
+def test_kernel_launch_sync_all_pes_have_equal_exec_time():
+    """No-op kernel: every PE's pe_exec_ns must be identical under D5."""
+    topo = resolve_topology(str(TOPOLOGY_PATH))
+    engine = GraphEngine(topo.topology_obj, enable_data=True)
+    spec = topo.topology_obj.spec
+
+    with RuntimeContext(engine=engine, target_device=DeviceSelector("all"),
+                        correlation_id="sync_test", spec=spec) as ctx:
+        dp = DPPolicy(cube="row_wise", pe="column_wise",
+                      num_cubes=16, num_pes=8)
+
+        def kernel(t_ptr, n_elem, tl):
+            pass  # no-op
+
+        ctx.ahbm.set_device(0)
+        t = ctx.zeros((16, 8 * 64), dtype="f16", dp=dp, name="probe")
+        t.copy_(ctx.from_numpy(np.zeros((16, 8 * 64), dtype=np.float16)))
+
+        pending = ctx.launch("sync_probe", kernel, t, 64, _defer_wait=True)
+        for h, _sip, meta in pending:
+            ctx.wait(h, _meta=meta)
+
+        pe_exec_vals = []
+        for h, _sip, _meta in pending:
+            _, trace = engine.get_completion(h)
+            if trace and trace.get("pe_exec_ns") is not None:
+                pe_exec_vals.append(float(trace["pe_exec_ns"]))
+
+    assert pe_exec_vals, "expected completion traces with pe_exec_ns"
+    spread = max(pe_exec_vals) - min(pe_exec_vals)
+    assert spread < 1e-6, (
+        f"ADR-0009 D5 violated: pe_exec_ns spread across PEs = "
+        f"{spread:.6f} ns (expected 0). Values: {pe_exec_vals}"
+    )

tests/test_pe_to_pe_latency.py

@@ -0,0 +1,358 @@
+"""PE-to-PE latency sweep across hop types and data sizes.
+
+Compares IPCQ send/recv vs raw-DMA (tl.load + tl.store) latency for five
+hop types:
+
+  H1 Intra-cube horizontal   pe0 → pe1
+  H2 Intra-cube vertical     pe0 → pe4
+  H3 Inter-cube horizontal   sip0.cube0.pe0 → sip0.cube1.pe0
+  H4 Inter-cube vertical     sip0.cube0.pe0 → sip0.cube4.pe0
+  H5 Inter-SIP               sip0.cube0.pe0 → sip1.cube0.pe0   (IPCQ only —
+                                                                raw needs
+                                                                cross-SIP MMU)
+
+Sizes: 128..10240 bytes. Emits PNGs with both lines plus a CSV.
+"""
+from __future__ import annotations
+
+import csv
+from dataclasses import dataclass
+from pathlib import Path
+
+import numpy as np
+import pytest
+
+from kernbench.ccl.install import load_ccl_config, resolve_algorithm_config
+from kernbench.ccl.sfr_config import configure_sfr_intercube_multisip
+from kernbench.policy.placement.dp import DPPolicy
+from kernbench.runtime_api.context import RuntimeContext
+from kernbench.runtime_api.types import DeviceSelector
+from kernbench.sim_engine.engine import GraphEngine
+from kernbench.topology.builder import resolve_topology
+
+TOPOLOGY_PATH = Path(__file__).parent.parent / "topology.yaml"
+PLOT_DIR = Path(__file__).parent / "pe2pe_latency_plots"
+
+SIZES = [128, 256, 384, 512, 768, 1024, 2048, 4096, 8192, 10240]
+
+N_CUBES = 16
+N_PES = 8
+ELEM_BYTES = 2  # f16
+
+
+@dataclass(frozen=True)
+class Hop:
+    id: str
+    label: str
+    src: tuple[int, int, int]
+    dst: tuple[int, int, int]
+    send_dir: str
+    recv_dir: str
+    supports_raw: bool   # False for cross-SIP (DPPolicy intra-device only)
+
+
+HOPS = [
+    Hop("h1_intra_horizontal", "Intra-cube horizontal (pe0 to pe1)",
+        (0, 0, 0), (0, 0, 1), "intra_E", "intra_W", True),
+    Hop("h2_intra_vertical", "Intra-cube vertical (pe0 to pe4)",
+        (0, 0, 0), (0, 0, 4), "intra_S", "intra_N", True),
+    Hop("h3_inter_cube_horizontal", "Inter-cube horizontal (cube0 to cube1)",
+        (0, 0, 0), (0, 1, 0), "E", "W", True),
+    Hop("h4_inter_cube_vertical", "Inter-cube vertical (cube0 to cube4)",
+        (0, 0, 0), (0, 4, 0), "S", "N", True),
+    Hop("h5_inter_sip", "Inter-SIP (sip0 to sip1, same cube/pe)",
+        (0, 0, 0), (1, 0, 0), "global_E", "global_W", False),
+]
+
+
+def _make_engine():
+    topo = resolve_topology(str(TOPOLOGY_PATH))
+    engine = GraphEngine(topo.topology_obj, enable_data=True)
+    return engine, topo.topology_obj.spec
+
+
+# ── IPCQ path ────────────────────────────────────────────────────────
+
+
+def _measure_ipcq(hop: Hop, nbytes: int) -> float:
+    engine, spec = _make_engine()
+
+    cfg = load_ccl_config()
+    merged = resolve_algorithm_config(cfg, name="intercube_allreduce")
+    merged["slot_size"] = max(int(merged.get("slot_size", 4096)), nbytes)
+
+    n_elem = nbytes // ELEM_BYTES
+    src_sip, src_cube, src_pe = hop.src
+    dst_sip, dst_cube, dst_pe = hop.dst
+    send_dir, recv_dir = hop.send_dir, hop.recv_dir
+
+    with RuntimeContext(
+        engine=engine,
+        target_device=DeviceSelector("all"),
+        correlation_id=f"ipcq_{hop.id}_{nbytes}",
+        spec=spec,
+    ) as ctx:
+        configure_sfr_intercube_multisip(engine, spec, merged)
+
+        dp = DPPolicy(
+            cube="row_wise", pe="column_wise",
+            num_cubes=N_CUBES, num_pes=N_PES,
+        )
+
+        def kernel(t_ptr, n_elem, tl):
+            pe_id = tl.program_id(axis=0)
+            cube_id = tl.program_id(axis=1)
+            if cube_id == src_cube and pe_id == src_pe:
+                data = tl.load(t_ptr, shape=(n_elem,), dtype="f16")
+                tl.send(dir=send_dir, src=data)
+            elif cube_id == dst_cube and pe_id == dst_pe:
+                tl.recv(dir=recv_dir, shape=(n_elem,), dtype="f16")
+
+        tensors = []
+        for s in sorted({src_sip, dst_sip}):
+            ctx.ahbm.set_device(s)
+            t = ctx.zeros(
+                (N_CUBES, N_PES * n_elem), dtype="f16",
+                dp=dp, name=f"sip{s}",
+            )
+            t.copy_(ctx.from_numpy(
+                np.full((N_CUBES, N_PES * n_elem), 1.0, dtype=np.float16),
+            ))
+            tensors.append(t)
+
+        all_pending = []
+        for t in tensors:
+            pending = ctx.launch(
+                f"{hop.id}_ipcq", kernel, t, n_elem, _defer_wait=True,
+            )
+            all_pending.extend(pending)
+        for h, sip_id, meta in all_pending:
+            ctx.wait(h, _meta=meta)
+
+        # Per-PE kernel execution time (excludes launch dispatch and
+        # response aggregation). IPCQ: DST blocks on tl.recv until the
+        # send arrives, so max across SIPs = DST's transfer time.
+        pe_exec_vals = []
+        for h, _sip, _meta in all_pending:
+            _, trace = engine.get_completion(h)
+            if trace and trace.get("pe_exec_ns") is not None:
+                pe_exec_vals.append(float(trace["pe_exec_ns"]))
+
+    return max(pe_exec_vals) if pe_exec_vals else 0.0
+
+
+# ── Raw DMA path (intra-SIP only) ────────────────────────────────────
+
+
+def _measure_raw(hop: Hop, nbytes: int) -> float:
+    """tl.load from source slice + tl.store to destination slice. The VA
+    mapping spans the cube mesh within one SIP (MmuMapMsg broadcasts to all
+    cubes of the SIP), so the store goes through the fabric to the
+    destination PE's HBM.  No IPCQ protocol involved.
+    """
+    if not hop.supports_raw:
+        raise RuntimeError(f"hop {hop.id} does not support raw path")
+
+    engine, spec = _make_engine()
+
+    n_elem = nbytes // ELEM_BYTES
+    src_sip, src_cube, src_pe = hop.src
+    dst_sip, dst_cube, dst_pe = hop.dst
+    assert src_sip == dst_sip
+
+    # Slice offsets in the (N_CUBES, N_PES * n_elem) tensor:
+    #   row = cube, slice within row = pe * n_elem .. (pe+1)*n_elem
+    # Byte offsets from va_base:
+    src_off = (src_cube * N_PES + src_pe) * n_elem * ELEM_BYTES
+    dst_off = (dst_cube * N_PES + dst_pe) * n_elem * ELEM_BYTES
+
+    with RuntimeContext(
+        engine=engine,
+        target_device=DeviceSelector("all"),
+        correlation_id=f"raw_{hop.id}_{nbytes}",
+        spec=spec,
+    ) as ctx:
+        dp = DPPolicy(
+            cube="row_wise", pe="column_wise",
+            num_cubes=N_CUBES, num_pes=N_PES,
+        )
+        ctx.ahbm.set_device(src_sip)
+        t = ctx.zeros(
+            (N_CUBES, N_PES * n_elem), dtype="f16",
+            dp=dp, name="raw_tensor",
+        )
+        t.copy_(ctx.from_numpy(
+            np.full((N_CUBES, N_PES * n_elem), 1.0, dtype=np.float16),
+        ))
+
+        def kernel(t_ptr, n_elem, tl):
+            pe_id = tl.program_id(axis=0)
+            cube_id = tl.program_id(axis=1)
+            if cube_id == src_cube and pe_id == src_pe:
+                data = tl.load(
+                    t_ptr + src_off, shape=(n_elem,), dtype="f16",
+                )
+                tl.store(t_ptr + dst_off, data)
+
+        pending = ctx.launch(
+            f"{hop.id}_raw", kernel, t, n_elem, _defer_wait=True,
+        )
+        for h, sip_id, meta in pending:
+            ctx.wait(h, _meta=meta)
+
+        # Per-PE kernel execution time. Raw: only SRC does real work
+        # (tl.load + tl.store, store is blocking), so max across all PEs
+        # = SRC's transfer time. Idle PEs contribute only overhead_ns.
+        pe_exec_vals = []
+        for h, _sip, _meta in pending:
+            _, trace = engine.get_completion(h)
+            if trace and trace.get("pe_exec_ns") is not None:
+                pe_exec_vals.append(float(trace["pe_exec_ns"]))
+
+    return max(pe_exec_vals) if pe_exec_vals else 0.0
+
+
+# ── CSV + plotting ───────────────────────────────────────────────────
+
+
+def _write_csv(records, path: Path) -> None:
+    path.parent.mkdir(parents=True, exist_ok=True)
+    with open(path, "w", newline="", encoding="utf-8") as f:
+        w = csv.DictWriter(
+            f, fieldnames=["hop", "label", "size_bytes", "path", "total_ns"],
+        )
+        w.writeheader()
+        for r in records:
+            w.writerow(r)
+
+
+def _plot_per_hop(records, hop: Hop, path: Path) -> None:
+    import matplotlib.pyplot as plt
+
+    ipcq = sorted(
+        [r for r in records if r["hop"] == hop.id and r["path"] == "ipcq"],
+        key=lambda r: r["size_bytes"],
+    )
+    raw = sorted(
+        [r for r in records if r["hop"] == hop.id and r["path"] == "raw"],
+        key=lambda r: r["size_bytes"],
+    )
+
+    fig, ax = plt.subplots(figsize=(8, 5))
+    if ipcq:
+        ax.plot(
+            [r["size_bytes"] for r in ipcq],
+            [r["total_ns"] for r in ipcq],
+            marker="o", label="IPCQ (send/recv)", color="tab:blue",
+        )
+    if raw:
+        ax.plot(
+            [r["size_bytes"] for r in raw],
+            [r["total_ns"] for r in raw],
+            marker="s", label="Raw DMA (load+store)", color="tab:orange",
+        )
+    else:
+        ax.text(
+            0.98, 0.02, "(Raw DMA unavailable for cross-SIP)",
+            transform=ax.transAxes, ha="right", va="bottom",
+            fontsize=9, color="gray",
+        )
+    ax.set_xlabel("Data size (bytes)")
+    ax.set_ylabel("Latency (ns)")
+    ax.set_title(hop.label)
+    ax.grid(True, alpha=0.3)
+    ax.legend()
+    fig.tight_layout()
+    fig.savefig(path, dpi=120)
+    plt.close(fig)
+
+
+def _plot_overview(records, path: Path) -> None:
+    import matplotlib.pyplot as plt
+
+    fig, axes = plt.subplots(2, 3, figsize=(16, 9))
+    axes = axes.flatten()
+    for i, hop in enumerate(HOPS):
+        ax = axes[i]
+        ipcq = sorted(
+            [r for r in records if r["hop"] == hop.id and r["path"] == "ipcq"],
+            key=lambda r: r["size_bytes"],
+        )
+        raw = sorted(
+            [r for r in records if r["hop"] == hop.id and r["path"] == "raw"],
+            key=lambda r: r["size_bytes"],
+        )
+        if ipcq:
+            ax.plot(
+                [r["size_bytes"] for r in ipcq],
+                [r["total_ns"] for r in ipcq],
+                marker="o", label="IPCQ", color="tab:blue",
+            )
+        if raw:
+            ax.plot(
+                [r["size_bytes"] for r in raw],
+                [r["total_ns"] for r in raw],
+                marker="s", label="Raw", color="tab:orange",
+            )
+        ax.set_title(hop.label, fontsize=10)
+        ax.set_xlabel("bytes")
+        ax.set_ylabel("ns")
+        ax.grid(True, alpha=0.3)
+        ax.legend(fontsize=8)
+    for j in range(len(HOPS), len(axes)):
+        axes[j].axis("off")
+    fig.suptitle(
+        "PE-to-PE latency: IPCQ vs raw DMA",
+        fontsize=14,
+    )
+    fig.tight_layout()
+    fig.savefig(path, dpi=120)
+    plt.close(fig)
+
+
+# ── Test entry ───────────────────────────────────────────────────────
+
+
+def test_pe_to_pe_latency_sweep():
+    records: list[dict] = []
+
+    for hop in HOPS:
+        for size in SIZES:
+            # IPCQ path
+            ipcq_ns = _measure_ipcq(hop, size)
+            records.append({
+                "hop": hop.id, "label": hop.label,
+                "size_bytes": size, "path": "ipcq",
+                "total_ns": ipcq_ns,
+            })
+
+            raw_s = "n/a"
+            if hop.supports_raw:
+                raw_ns = _measure_raw(hop, size)
+                records.append({
+                    "hop": hop.id, "label": hop.label,
+                    "size_bytes": size, "path": "raw",
+                    "total_ns": raw_ns,
+                })
+                raw_s = f"{raw_ns:7.1f}ns"
+
+            print(
+                f"[{hop.id}] size={size:5d}  "
+                f"ipcq={ipcq_ns:7.1f}ns  raw={raw_s}"
+            )
+
+    PLOT_DIR.mkdir(parents=True, exist_ok=True)
+    _write_csv(records, PLOT_DIR / "summary.csv")
+    for hop in HOPS:
+        _plot_per_hop(records, hop, PLOT_DIR / f"{hop.id}.png")
+    _plot_overview(records, PLOT_DIR / "overview.png")
+
+    for hop in HOPS:
+        rs = sorted(
+            [r for r in records if r["hop"] == hop.id and r["path"] == "ipcq"],
+            key=lambda r: r["size_bytes"],
+        )
+        for r in rs:
+            assert r["total_ns"] > 0, f"{hop.id}: total_ns must be > 0"
+
+    print(f"\n  Plots + CSV written to {PLOT_DIR}")