diff --git a/.gitignore b/.gitignore
index 6aee0ed8e0..062358ef24 100644
--- a/.gitignore
+++ b/.gitignore
@@ -36,6 +36,7 @@ a.out
 config.json
 clcache
 compile_commands.json
+compare_runtime*.html
 
 persist
 board/obj/
diff --git a/.gitmodules b/.gitmodules
index bc439b451c..26f93ef164 100644
--- a/.gitmodules
+++ b/.gitmodules
@@ -16,3 +16,6 @@
 [submodule "body"]
   path = body
   url = ../../commaai/body.git
+[submodule "tinygrad"]
+  path = tinygrad_repo
+  url = https://github.com/geohot/tinygrad.git
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index 50554acceb..347216f2fb 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -28,7 +28,7 @@ repos:
     rev: v0.931
     hooks:
     -   id: mypy
-        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)/'
+        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)/|(tinygrad/)|(tinygrad_repo/)'
         additional_dependencies: ['types-PyYAML', 'lxml', 'numpy', 'types-atomicwrites', 'types-pycurl', 'types-requests', 'types-certifi']
         args:
         - --warn-redundant-casts
@@ -40,7 +40,7 @@ repos:
     rev: 4.0.1
     hooks:
     -   id: flake8
-        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)|(selfdrive/debug/)/'
+        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(opendbc/)|(laika_repo/)|(rednose_repo/)|(tinygrad/)|(tinygrad_repo/)|(selfdrive/debug/)/'
         additional_dependencies: ['flake8-no-implicit-concat']
         args:
         - --indent-size=2
@@ -55,7 +55,7 @@ repos:
         entry: pylint
         language: system
         types: [python]
-        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(laika_repo/)|(rednose_repo/)'
+        exclude: '^(pyextra/)|(cereal/)|(rednose/)|(panda/)|(laika/)|(laika_repo/)|(rednose_repo/)|(tinygrad/)|(tinygrad_repo/)'
         args:
         - -rn
         - -sn
diff --git a/SConstruct b/SConstruct
index d4155f4f83..66a94f9b1b 100644
--- a/SConstruct
+++ b/SConstruct
@@ -49,6 +49,11 @@ AddOption('--no-thneed',
           dest='no_thneed',
           help='avoid using thneed')
 
+AddOption('--pc-thneed',
+          action='store_true',
+          dest='pc_thneed',
+          help='use thneed on pc')
+
 AddOption('--no-test',
           action='store_false',
           dest='test',
diff --git a/common/clutil.cc b/common/clutil.cc
index b8f9fde4cf..9d3447d807 100644
--- a/common/clutil.cc
+++ b/common/clutil.cc
@@ -78,7 +78,8 @@ cl_program cl_program_from_file(cl_context ctx, cl_device_id device_id, const ch
 }
 
 cl_program cl_program_from_source(cl_context ctx, cl_device_id device_id, const std::string& src, const char* args) {
-  cl_program prg = CL_CHECK_ERR(clCreateProgramWithSource(ctx, 1, (const char*[]){src.c_str()}, NULL, &err));
+  const char *csrc = src.c_str();
+  cl_program prg = CL_CHECK_ERR(clCreateProgramWithSource(ctx, 1, &csrc, NULL, &err));
   if (int err = clBuildProgram(prg, 1, &device_id, args, NULL, NULL); err != 0) {
     cl_print_build_errors(prg, device_id);
     assert(0);
@@ -87,7 +88,7 @@ cl_program cl_program_from_source(cl_context ctx, cl_device_id device_id, const
 }
 
 cl_program cl_program_from_binary(cl_context ctx, cl_device_id device_id, const uint8_t* binary, size_t length, const char* args) {
-  cl_program prg = CL_CHECK_ERR(clCreateProgramWithBinary(ctx, 1, &device_id, &length, (const uint8_t*[]){binary}, NULL, &err));
+  cl_program prg = CL_CHECK_ERR(clCreateProgramWithBinary(ctx, 1, &device_id, &length, &binary, NULL, &err));
   if (int err = clBuildProgram(prg, 1, &device_id, args, NULL, NULL); err != 0) {
     cl_print_build_errors(prg, device_id);
     assert(0);
diff --git a/release/build_release.sh b/release/build_release.sh
index 1000e607fa..80106eefb2 100755
--- a/release/build_release.sh
+++ b/release/build_release.sh
@@ -78,7 +78,7 @@ find . -name 'moc_*' -delete
 find . -name '__pycache__' -delete
 rm -rf panda/board panda/certs panda/crypto
 rm -rf .sconsign.dblite Jenkinsfile release/
-rm selfdrive/modeld/models/supercombo.dlc
+rm selfdrive/modeld/models/supercombo.onnx
 
 # Move back signed panda fw
 mkdir -p panda/board/obj
diff --git a/release/files_common b/release/files_common
index ca6e91fb65..5be07b1c75 100644
--- a/release/files_common
+++ b/release/files_common
@@ -352,7 +352,7 @@ selfdrive/modeld/models/driving.cc
 selfdrive/modeld/models/driving.h
 selfdrive/modeld/models/dmonitoring.cc
 selfdrive/modeld/models/dmonitoring.h
-selfdrive/modeld/models/supercombo.dlc
+selfdrive/modeld/models/supercombo.onnx
 selfdrive/modeld/models/dmonitoring_model_q.dlc
 
 selfdrive/modeld/transforms/loadyuv.cc
@@ -561,3 +561,16 @@ opendbc/vw_mqb_2010.dbc
 opendbc/tesla_can.dbc
 opendbc/tesla_radar.dbc
 opendbc/tesla_powertrain.dbc
+
+tinygrad_repo/openpilot/compile.py
+tinygrad_repo/accel/opencl/*
+tinygrad_repo/extra/onnx.py
+tinygrad_repo/extra/utils.py
+tinygrad_repo/tinygrad/llops/ops_gpu.py
+tinygrad_repo/tinygrad/llops/ops_opencl.py
+tinygrad_repo/tinygrad/helpers.py
+tinygrad_repo/tinygrad/mlops.py
+tinygrad_repo/tinygrad/ops.py
+tinygrad_repo/tinygrad/shapetracker.py
+tinygrad_repo/tinygrad/tensor.py
+tinygrad_repo/tinygrad/nn/__init__.py
diff --git a/selfdrive/modeld/SConscript b/selfdrive/modeld/SConscript
index 4feb17f238..2544607aa4 100644
--- a/selfdrive/modeld/SConscript
+++ b/selfdrive/modeld/SConscript
@@ -62,25 +62,65 @@ else:
 
 common_model = lenv.Object(common_src)
 
-# build thneed model
-if use_thneed and arch == "larch64":
-  fn = File("models/supercombo").abspath
-  compiler = lenv.Program('thneed/compile', ["thneed/compile.cc"]+common_model, LIBS=libs)
-  cmd = f"cd {Dir('.').abspath} && {compiler[0].abspath} --in {fn}.dlc --out {fn}.thneed --binary --optimize"
-
-  lib_paths = ':'.join(Dir(p).abspath for p in lenv["LIBPATH"])
-  kernel_path = os.path.join(Dir('.').abspath, "thneed", "kernels")
-  cenv = Environment(ENV={'LD_LIBRARY_PATH': f"{lib_paths}:{lenv['ENV']['LD_LIBRARY_PATH']}", 'KERNEL_PATH': kernel_path})
-
-  kernels = [os.path.join(kernel_path, x) for x in os.listdir(kernel_path) if x.endswith(".cl")]
-  cenv.Command(fn + ".thneed", [fn + ".dlc", kernels, compiler], cmd)
-
 lenv.Program('_dmonitoringmodeld', [
     "dmonitoringmodeld.cc",
     "models/dmonitoring.cc",
   ]+common_model, LIBS=libs)
 
-lenv.Program('_modeld', [
+# build thneed model
+if use_thneed and arch == "larch64" or GetOption('pc_thneed'):
+  fn = File("models/supercombo").abspath
+
+  if GetOption('pc_thneed'):
+    cmd = f"cd {Dir('#').abspath}/tinygrad_repo && NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
+  else:
+    cmd = f"cd {Dir('#').abspath}/tinygrad_repo && FLOAT16=1 PYOPENCL_NO_CACHE=1 MATMUL=1 NATIVE_EXPLOG=1 OPTWG=1 UNSAFE_FLOAT4=1 DEBUGCL=1 python3 openpilot/compile.py {fn}.onnx {fn}.thneed"
+
+  # is there a better way then listing all of tinygrad?
+  lenv.Command(fn + ".thneed", [fn + ".onnx",
+    "#tinygrad_repo/openpilot/compile.py",
+    "#tinygrad_repo/accel/opencl/conv.cl",
+    "#tinygrad_repo/accel/opencl/matmul.cl",
+    "#tinygrad_repo/accel/opencl/ops_opencl.py",
+    "#tinygrad_repo/accel/opencl/preprocessing.py",
+    "#tinygrad_repo/extra/onnx.py",
+    "#tinygrad_repo/extra/utils.py",
+    "#tinygrad_repo/tinygrad/llops/ops_gpu.py",
+    "#tinygrad_repo/tinygrad/llops/ops_opencl.py",
+    "#tinygrad_repo/tinygrad/helpers.py",
+    "#tinygrad_repo/tinygrad/mlops.py",
+    "#tinygrad_repo/tinygrad/ops.py",
+    "#tinygrad_repo/tinygrad/shapetracker.py",
+    "#tinygrad_repo/tinygrad/tensor.py",
+    "#tinygrad_repo/tinygrad/nn/__init__.py"
+  ], cmd)
+
+  # old thneed compiler. TODO: remove this once tinygrad stuff is stable
+
+  #compiler = lenv.Program('thneed/compile', ["thneed/compile.cc"]+common_model, LIBS=libs)
+  #cmd = f"cd {Dir('.').abspath} && {compiler[0].abspath} --in {fn}.dlc --out {fn}.thneed --binary --optimize"
+
+  #lib_paths = ':'.join(Dir(p).abspath for p in lenv["LIBPATH"])
+  #kernel_path = os.path.join(Dir('.').abspath, "thneed", "kernels")
+  #cenv = Environment(ENV={'LD_LIBRARY_PATH': f"{lib_paths}:{lenv['ENV']['LD_LIBRARY_PATH']}", 'KERNEL_PATH': kernel_path})
+
+  #kernels = [os.path.join(kernel_path, x) for x in os.listdir(kernel_path) if x.endswith(".cl")]
+  #cenv.Command(fn + ".thneed", [fn + ".dlc", kernels, compiler], cmd)
+
+llenv = lenv.Clone()
+if GetOption('pc_thneed'):
+  pc_thneed_src = [
+    "thneed/thneed_common.cc",
+    "thneed/thneed_pc.cc",
+    "thneed/serialize.cc",
+    "runners/thneedmodel.cc",
+  ]
+  llenv['CFLAGS'].append("-DUSE_THNEED")
+  llenv['CXXFLAGS'].append("-DUSE_THNEED")
+  common_model += llenv.Object(pc_thneed_src)
+  libs += ['dl']
+
+llenv.Program('_modeld', [
     "modeld.cc",
     "models/driving.cc",
   ]+common_model, LIBS=libs + transformations)
diff --git a/selfdrive/modeld/models/supercombo.dlc b/selfdrive/modeld/models/supercombo.dlc
deleted file mode 100644
index fe133523fc..0000000000
--- a/selfdrive/modeld/models/supercombo.dlc
+++ /dev/null
@@ -1,3 +0,0 @@
-version https://git-lfs.github.com/spec/v1
-oid sha256:93d265fc88f05746ce47257e15fc2afe43b250b44715313049f829e8aa30a9d6
-size 94302331
diff --git a/selfdrive/modeld/thneed/serialize.cc b/selfdrive/modeld/thneed/serialize.cc
index 0b8b02c6ac..afc84ee769 100644
--- a/selfdrive/modeld/thneed/serialize.cc
+++ b/selfdrive/modeld/thneed/serialize.cc
@@ -33,11 +33,14 @@ void Thneed::load(const char *filename) {
       assert(mobj["needs_load"].bool_value() == false);
     } else {
       if (mobj["needs_load"].bool_value()) {
-        //printf("loading %p %d @ 0x%X\n", clbuf, sz, ptr);
         clbuf = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, sz, &buf[ptr], NULL);
+        if (debug >= 1) printf("loading %p %d @ 0x%X\n", clbuf, sz, ptr);
         ptr += sz;
       } else {
-        clbuf = clCreateBuffer(context, CL_MEM_READ_WRITE, sz, NULL, NULL);
+        // TODO: is there a faster way to init zeroed out buffers?
+        void *host_zeros = calloc(sz, 1);
+        clbuf = clCreateBuffer(context, CL_MEM_COPY_HOST_PTR | CL_MEM_READ_WRITE, sz, host_zeros, NULL);
+        free(host_zeros);
       }
     }
     assert(clbuf != NULL);
diff --git a/selfdrive/modeld/thneed/thneed.h b/selfdrive/modeld/thneed/thneed.h
index 19dc46e8f6..2a5800f302 100644
--- a/selfdrive/modeld/thneed/thneed.h
+++ b/selfdrive/modeld/thneed/thneed.h
@@ -122,7 +122,7 @@ class Thneed {
 
     // all CL kernels
     void find_inputs_outputs();
-    void copy_inputs(float **finputs);
+    void copy_inputs(float **finputs, bool internal=false);
     void copy_output(float *foutput);
     cl_int clexec();
     vector<shared_ptr<CLQueuedKernel> > kq;
diff --git a/selfdrive/modeld/thneed/thneed_common.cc b/selfdrive/modeld/thneed/thneed_common.cc
index b751cdf665..a3f5c908f9 100644
--- a/selfdrive/modeld/thneed/thneed_common.cc
+++ b/selfdrive/modeld/thneed/thneed_common.cc
@@ -30,17 +30,16 @@ cl_int Thneed::clexec() {
   return clFinish(command_queue);
 }
 
-void Thneed::copy_inputs(float **finputs) {
-  //cl_int ret;
+void Thneed::copy_inputs(float **finputs, bool internal) {
   for (int idx = 0; idx < inputs.size(); ++idx) {
     if (debug >= 1) printf("copying %lu -- %p -> %p (cl %p)\n", input_sizes[idx], finputs[idx], inputs[idx], input_clmem[idx]);
 
-    // TODO: fix thneed caching
-    if (finputs[idx] != NULL) memcpy(inputs[idx], finputs[idx], input_sizes[idx]);
-    //if (finputs[idx] != NULL) CL_CHECK(clEnqueueWriteBuffer(command_queue, input_clmem[idx], CL_TRUE, 0, input_sizes[idx], finputs[idx], 0, NULL, NULL));
-
-    // HACK
-    //if (input_sizes[idx] == 16) memset((char*)inputs[idx] + 8, 0, 8);
+    if (internal) {
+      // if it's internal, using memcpy is fine since the buffer sync is cached in the ioctl layer
+      if (finputs[idx] != NULL) memcpy(inputs[idx], finputs[idx], input_sizes[idx]);
+    } else {
+      if (finputs[idx] != NULL) CL_CHECK(clEnqueueWriteBuffer(command_queue, input_clmem[idx], CL_TRUE, 0, input_sizes[idx], finputs[idx], 0, NULL, NULL));
+    }
   }
 }
 
@@ -202,8 +201,8 @@ void CLQueuedKernel::debug_print(bool verbose) {
             assert(slice_pitch == 0);
 
             clGetImageInfo(val, CL_IMAGE_BUFFER, sizeof(buf), &buf, NULL);
-            size_t sz;
-            clGetMemObjectInfo(buf, CL_MEM_SIZE, sizeof(sz), &sz, NULL);
+            size_t sz = 0;
+            if (buf != NULL) clGetMemObjectInfo(buf, CL_MEM_SIZE, sizeof(sz), &sz, NULL);
             printf(" image %zu x %zu rp %zu @ %p buffer %zu", width, height, row_pitch, buf, sz);
           } else {
             size_t sz;
diff --git a/selfdrive/modeld/thneed/thneed_pc.cc b/selfdrive/modeld/thneed/thneed_pc.cc
new file mode 100644
index 0000000000..e32dd289ec
--- /dev/null
+++ b/selfdrive/modeld/thneed/thneed_pc.cc
@@ -0,0 +1,40 @@
+#include "selfdrive/modeld/thneed/thneed.h"
+
+#include <cassert>
+
+#include "common/clutil.h"
+#include "common/timing.h"
+
+Thneed::Thneed(bool do_clinit, cl_context _context) {
+  context = _context;
+  if (do_clinit) clinit();
+  char *thneed_debug_env = getenv("THNEED_DEBUG");
+  debug = (thneed_debug_env != NULL) ? atoi(thneed_debug_env) : 0;
+}
+
+void Thneed::execute(float **finputs, float *foutput, bool slow) {
+  uint64_t tb, te;
+  if (debug >= 1) tb = nanos_since_boot();
+
+  // ****** copy inputs
+  copy_inputs(finputs);
+
+  // ****** run commands
+  clexec();
+
+  // ****** copy outputs
+  copy_output(foutput);
+
+  if (debug >= 1) {
+    te = nanos_since_boot();
+    printf("model exec in %lu us\n", (te-tb)/1000);
+  }
+}
+
+void Thneed::stop() {
+}
+
+void Thneed::find_inputs_outputs() {
+  // thneed on PC doesn't work on old style inputs/outputs
+}
+
diff --git a/selfdrive/modeld/thneed/thneed_qcom2.cc b/selfdrive/modeld/thneed/thneed_qcom2.cc
index e79bb77edf..f35317d2a7 100644
--- a/selfdrive/modeld/thneed/thneed_qcom2.cc
+++ b/selfdrive/modeld/thneed/thneed_qcom2.cc
@@ -269,7 +269,7 @@ void Thneed::execute(float **finputs, float *foutput, bool slow) {
   if (debug >= 1) tb = nanos_since_boot();
 
   // ****** copy inputs
-  copy_inputs(finputs);
+  copy_inputs(finputs, true);
 
   // ****** set power constraint
   int ret;
diff --git a/selfdrive/test/process_replay/model_replay_ref_commit b/selfdrive/test/process_replay/model_replay_ref_commit
index 80be9b464f..e2ee2e7bb6 100644
--- a/selfdrive/test/process_replay/model_replay_ref_commit
+++ b/selfdrive/test/process_replay/model_replay_ref_commit
@@ -1 +1 @@
-ca90e11f8d59902af38d3785ddd91a27d0fbb411
+cffb4e720b0379bedd4ff802912d998ace775c37
diff --git a/tinygrad b/tinygrad
new file mode 120000
index 0000000000..cb003823c6
--- /dev/null
+++ b/tinygrad
@@ -0,0 +1 @@
+tinygrad_repo/tinygrad
\ No newline at end of file
diff --git a/tinygrad_repo b/tinygrad_repo
new file mode 160000
index 0000000000..2e9b7637b3
--- /dev/null
+++ b/tinygrad_repo
@@ -0,0 +1 @@
+Subproject commit 2e9b7637b3c3c8895fda9f964215db3a35fe3441