diff --git a/README.md b/README.md
index 2c856b4..5b8f85c 100644
--- a/README.md
+++ b/README.md
@@ -105,8 +105,8 @@ A `ValueError` will be raised if `parse_float` produces illegal types.
 - it's lil'
 - pure Python with zero dependencies
 - the fastest pure Python parser [\*](#performance):
-  15x as fast as [tomlkit](https://pypi.org/project/tomlkit/),
-  2.4x as fast as [toml](https://pypi.org/project/toml/)
+  16x as fast as [tomlkit](https://pypi.org/project/tomlkit/),
+  2.3x as fast as [toml](https://pypi.org/project/toml/)
 - outputs [basic data types](#how-do-toml-types-map-into-python-types) only
 - 100% spec compliant: passes all tests in
   [a test set](https://github.com/toml-lang/compliance/pull/8)
@@ -157,10 +157,10 @@ Running the benchmark on my personal computer output the following:
 
 ```console
 foo@bar:~/dev/tomli$ tox -e benchmark-pypi
-benchmark-pypi installed: attrs==19.3.0,click==7.1.2,pytomlpp==1.0.2,qtoml==0.3.0,rtoml==0.7.0,toml==0.10.2,tomli==1.1.0,tomlkit==0.7.2
-benchmark-pypi run-test-pre: PYTHONHASHSEED='2658546909'
+benchmark-pypi installed: attrs==21.4.0,click==8.0.3,pytomlpp==1.0.10,qtoml==0.3.1,rtoml==0.7.1,toml==0.10.2,tomli==2.0.1,tomlkit==0.9.2
+benchmark-pypi run-test-pre: PYTHONHASHSEED='3088452573'
 benchmark-pypi run-test: commands[0] | python -c 'import datetime; print(datetime.date.today())'
-2021-07-23
+2022-02-09
 benchmark-pypi run-test: commands[1] | python --version
 Python 3.8.10
 benchmark-pypi run-test: commands[2] | python benchmark/run.py
@@ -168,12 +168,12 @@ Parsing data.toml 5000 times:
 ------------------------------------------------------
     parser |  exec time | performance (more is better)
 -----------+------------+-----------------------------
-     rtoml |    0.901 s | baseline (100%)
-  pytomlpp |     1.08 s | 83.15%
-     tomli |     3.89 s | 23.15%
-      toml |     9.36 s | 9.63%
-     qtoml |     11.5 s | 7.82%
-   tomlkit |     56.8 s | 1.59%
+     rtoml |    0.891 s | baseline (100%)
+  pytomlpp |    0.969 s | 91.90%
+     tomli |        4 s | 22.25%
+      toml |     9.01 s | 9.88%
+     qtoml |     11.1 s | 8.05%
+   tomlkit |       63 s | 1.41%
 ```
 
 The parsers are ordered from fastest to slowest, using the fastest parser as baseline.
diff --git a/fuzzer/requirements.txt b/fuzzer/requirements.txt
index b4d78cc..ab854b4 100644
--- a/fuzzer/requirements.txt
+++ b/fuzzer/requirements.txt
@@ -1,4 +1,4 @@
 # sudo apt-get install clang
 wheel
-atheris==2.0.3
+atheris==2.0.7
 tomli_w>=0.2.2
diff --git a/profiler/profiler_script.py b/profiler/profiler_script.py
new file mode 100644
index 0000000..7264dc7
--- /dev/null
+++ b/profiler/profiler_script.py
@@ -0,0 +1,17 @@
+"""A script for profiling.
+
+To generate and read results:
+  - `tox -e profile`
+  - `firefox .tox/prof/output.svg`
+"""
+from pathlib import Path
+
+import tomli
+
+benchmark_toml = (
+    (Path(__file__).parent.parent / "benchmark" / "data.toml").read_bytes().decode()
+)
+
+# Run this a few times to emphasize over imports and other overhead above.
+for _ in range(1000):
+    tomli.loads(benchmark_toml)
diff --git a/profiler/requirements.txt b/profiler/requirements.txt
index c65a7f7..9d501fb 100644
--- a/profiler/requirements.txt
+++ b/profiler/requirements.txt
@@ -1 +1 @@
-pytest-profiling
+gprof2dot
diff --git a/profiler/test_for_profiler.py b/profiler/test_for_profiler.py
deleted file mode 100644
index 0d5e079..0000000
--- a/profiler/test_for_profiler.py
+++ /dev/null
@@ -1,24 +0,0 @@
-"""Test for profiling.
-
-This test can be useful for profiling, as most of the execution time
-will be spent parsing and rendering instead of managing pytest execution
-environment. To get and read profiler results:
-  - `tox -e profile`
-  - `firefox .tox/prof/combined.svg`
-"""
-import os
-from pathlib import Path
-
-import tomli
-
-
-def test_for_profiler() -> None:
-    path = Path(__file__).parent.parent / "benchmark" / "data.toml"
-    benchmark_toml = path.read_bytes().decode()
-    # increase the count here to reduce the impact of
-    # setting up pytest execution environment. Let's keep
-    # the count low by default because this is part of the
-    # standard test suite.
-    iterations = int(os.environ.get("PROFILER_ITERATIONS", 1))
-    for _ in range(iterations):
-        tomli.loads(benchmark_toml)
diff --git a/pyproject.toml b/pyproject.toml
index adcf366..bbc1e2d 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -60,14 +60,18 @@ commands =
     python -m unittest {posargs}
 
 [testenv:profile]
-description = run profiler (use e.g. `firefox .tox/prof/combined.svg` to open)
-setenv =
-    PROFILER_ITERATIONS=1000
+description = run profiler (use e.g. `firefox .tox/prof/output.svg` to open)
 deps =
     -r profiler/requirements.txt
+allowlist_externals =
+    mkdir
+    dot
 commands =
-    pytest profiler/test_for_profiler.py --profile-svg --pstats-dir "{toxworkdir}/prof"
-    python -c 'import pathlib; print("profiler svg output under file://\{0\}".format(pathlib.Path(r"{toxworkdir}") / "prof" / "combined.svg"))'
+    mkdir -p "{toxworkdir}/prof"
+    python -m cProfile -o "{toxworkdir}/prof/output.pstats" profiler/profiler_script.py
+    gprof2dot -f pstats -o "{toxworkdir}/prof/output.dot" "{toxworkdir}/prof/output.pstats"
+    dot -Tsvg -o "{toxworkdir}/prof/output.svg" "{toxworkdir}/prof/output.dot"
+    python -c 'import pathlib; print("profiler svg output under file://\{0\}".format(pathlib.Path(r"{toxworkdir}") / "prof" / "output.svg"))'
 
 [testenv:pre-commit]
 description = run linters