diff --git a/runtime/kernel.c b/runtime/kernel.c
index 194e3e1b..128f1b46 100644
--- a/runtime/kernel.c
+++ b/runtime/kernel.c
@@ -7,6 +7,7 @@
 
 #include "hle.h"
 #include "kernel.h"
+#include "platform.h"
 #include "prof.h"
 
 // ------------------------------------------------------------------ results
@@ -394,6 +395,7 @@ void rt_unwind_point(void) {
 static void *thread_main(void *arg) {
     KThread *t = arg;
     t_cur = t;
+    if (t->id == 1) plat_thread_elevate();
     {
         char pn[32];
         snprintf(pn, sizeof pn, "guest-%x", t->handle);
diff --git a/runtime/nvn/objects.c b/runtime/nvn/objects.c
index 8fe09649..0578a4c3 100644
--- a/runtime/nvn/objects.c
+++ b/runtime/nvn/objects.c
@@ -5,6 +5,7 @@
 #include "../framepace.h"
 #include "../game_sites.h"
 #include "../kernel.h"
+#include "../platform.h"
 #include "../prof.h"
 #include "../perfmon.h"
 #include "nvn.h"
@@ -224,6 +225,7 @@ static void qop_run(const QOp *o) {
 static void *queue_thread(void *arg) {
     (void)arg;
     prof_thread_name("nvn-queue");
+    plat_thread_elevate();
     kernel_sample_host_thread();
     extern void control_prof_thread(const char *name);  // runtime/control_prof.c (`threads`)
     control_prof_thread("render");
diff --git a/runtime/nvn/vulkan/record.c b/runtime/nvn/vulkan/record.c
index 2d13fae9..6f6dfa31 100644
--- a/runtime/nvn/vulkan/record.c
+++ b/runtime/nvn/vulkan/record.c
@@ -181,6 +181,7 @@ static uint64_t spin_for_work(uint64_t t) {  // the published end once it moves
 static void *run(void *arg) {
     (void)arg;
     prof_thread_name("vk-record");
+    plat_thread_elevate();
     extern void control_prof_thread(const char *name);  // runtime/control_prof.c (`threads`)
     control_prof_thread("recorder");
     uint64_t t = 0;
diff --git a/runtime/platform.h b/runtime/platform.h
index a359f0d7..108f366b 100644
--- a/runtime/platform.h
+++ b/runtime/platform.h
@@ -54,6 +54,7 @@ uint64_t plat_now_ns(void);                // monotonic
 void plat_sleep_until_ns(uint64_t t);      // precise wait for plat_now_ns() >= t
 void plat_wait_until_ns(uint64_t t);       // wait for plat_now_ns() >= t without spinning (the OS timer's precision)
 void plat_thread_realtime(uint64_t period_ns);  // periodic timing thread (vsync): don't coalesce its timers
+void plat_thread_elevate(void);            // latency-critical thread (game main, render, recorder): favor P-cores on hybrid architectures
 int plat_cpu_count(void);       // logical processors this process may run on (its affinity, after plat_cpu_placement)
 void plat_cpu_placement(void);  // (game start) one logical processor per core where plat_one_per_core_default says
 // Whether the game runs on one logical processor per core when SAVAGEWIND_ONE_PER_CORE doesn't say (cores: physical
diff --git a/runtime/platform/linux.c b/runtime/platform/linux.c
index 9cf53677..158c5a8e 100644
--- a/runtime/platform/linux.c
+++ b/runtime/platform/linux.c
@@ -134,6 +134,8 @@ void plat_thread_realtime(uint64_t period_ns) {
     setpriority(PRIO_PROCESS, (id_t)gettid(), -10);  // (per thread on Linux; fails quietly without CAP_SYS_NICE)
 }
 
+void plat_thread_elevate(void) {}
+
 // ---- process
 // stdout / stderr go to savagewind.log in the data folder (runtime/paths.h; rt_paths_init runs later, so the same rule
 // is applied here). Processes started by this one (the start screen re-executing itself for the game) append to the
diff --git a/runtime/platform/macos.c b/runtime/platform/macos.c
index e251e879..1d3c0d09 100644
--- a/runtime/platform/macos.c
+++ b/runtime/platform/macos.c
@@ -83,6 +83,8 @@ void plat_thread_realtime(uint64_t period_ns) {
                       THREAD_TIME_CONSTRAINT_POLICY_COUNT);
 }
 
+void plat_thread_elevate(void) {}
+
 const char *plat_log_path(void) { return ""; }  // (the log goes to stderr)
 
 bool plat_exe_path(char *buf, size_t size) {
diff --git a/runtime/platform/windows.c b/runtime/platform/windows.c
index 446efd9f..75c2bc7b 100644
--- a/runtime/platform/windows.c
+++ b/runtime/platform/windows.c
@@ -325,6 +325,10 @@ void plat_thread_realtime(uint64_t period_ns) {
     SetThreadPriority(GetCurrentThread(), THREAD_PRIORITY_TIME_CRITICAL);
 }
 
+void plat_thread_elevate(void) {
+    SetThreadPriority(GetCurrentThread(), THREAD_PRIORITY_ABOVE_NORMAL);
+}
+
 // ---- process
 // Release builds are GUI-subsystem executables (no console window): stdout / stderr go to savagewind.log next to the
 // exe, and a fatal error shows a message box (plat_fatal_notice) instead of vanishing. Processes started by this one
@@ -433,9 +437,11 @@ void plat_init(void) {
 // never share a core with another busy thread of the game. With fewer cores than the game's busy threads (main, two job
 // workers, render, recorder, audio, file loading) they queue for one instead: by default with SMT and 4 or more
 // physical cores (plat_one_per_core_default). SAVAGEWIND_ONE_PER_CORE=1: always, 0: never. Called when the game
-// starts, so what the start screen runs (the game code build) keeps every logical processor. Under Wine / Proton the
-// Linux kernel schedules the threads, and there the Steam Deck measured one processor per core as a loss
-// (platform.h): the default keeps every logical processor, as the Linux build does.
+// starts, so what the start screen runs (the game code build) keeps every logical processor. On hybrid architectures
+// (Intel 12th+ Gen, Core Ultra, heterogeneous P/E cores) keeping all logical processors avoids starving P-core SMT
+// execution slots and forcing critical threads onto low-clock E-cores. Under Wine / Proton the Linux kernel schedules
+// the threads, and there the Steam Deck measured one processor per core as a loss (platform.h): the default keeps
+// every logical processor, as the Linux build does.
 void plat_cpu_placement(void) {
     DWORD_PTR all, sys, one = 0;
     if (!GetProcessAffinityMask(GetCurrentProcess(), &all, &sys)) return;
@@ -445,18 +451,28 @@ void plat_cpu_placement(void) {
     DWORD len = 0;
     GetLogicalProcessorInformationEx(RelationProcessorCore, NULL, &len);
     char *buf = len ? malloc(len) : NULL;
+    BYTE min_eff = 255, max_eff = 0;
     if (buf && GetLogicalProcessorInformationEx(RelationProcessorCore, (void *)buf, &len))
         for (DWORD o = 0; o < len;) {
             SYSTEM_LOGICAL_PROCESSOR_INFORMATION_EX *e = (void *)(buf + o);
             DWORD_PTR m = e->Processor.GroupMask[0].Group == 0 ? e->Processor.GroupMask[0].Mask & all : 0;
-            if (m) cores++, one |= m & -m;  // (the core's first logical processor)
+            if (m) {
+                cores++;
+                one |= m & -m;  // (the core's first logical processor)
+                BYTE eff = e->Processor.EfficiencyClass;
+                if (eff < min_eff) min_eff = eff;
+                if (eff > max_eff) max_eff = eff;
+            }
             o += e->Size;
         }
     free(buf);
+    bool hybrid = (cores > 0 && min_eff != max_eff);
     const char *x = getenv("SAVAGEWIND_ONE_PER_CORE");
-    bool on = one && (x ? atoi(x) != 0 : !wine && plat_one_per_core_default(cores, logical));
-    fprintf(stderr, "[cpu] %d cores, %d logical processors: %s%s\n", cores, logical,
-            on ? "one logical processor per core" : "all logical processors", wine && !x ? " (Wine)" : "");
+    bool on = one && (x ? atoi(x) != 0 : !wine && !hybrid && plat_one_per_core_default(cores, logical));
+    fprintf(stderr, "[cpu] %d cores%s, %d logical processors: %s%s\n", cores,
+            hybrid ? " (hybrid P/E)" : "", logical,
+            on ? "one logical processor per core" : "all logical processors",
+            wine && !x ? " (Wine)" : hybrid && !x ? " (hybrid architecture)" : "");
     if (on) SetProcessAffinityMask(GetCurrentProcess(), one);
 }