diff --git a/.circleci/config.yml b/.circleci/config.yml
index 8f21c1aa..5c6e6e0e 100644
--- a/.circleci/config.yml
+++ b/.circleci/config.yml
@@ -24,12 +24,6 @@ jobs:
       - run:
           name: "Build infos"
           command: go version
-      - run:
-          name: "Run benchmark tests"
-          command: ginkgo -randomizeAllSpecs -trace benchmark -- -size=10
-      - run:
-          name: "Run benchmark tests with race detector"
-          command: ginkgo -race -randomizeAllSpecs -trace benchmark -- -size=5
       - run:
           name: "Run tools tests"
           command: ginkgo -race -r -v -randomizeAllSpecs -trace integrationtests/tools
diff --git a/.github/workflows/unit.yml b/.github/workflows/unit.yml
index 1985081c..4b30f747 100644
--- a/.github/workflows/unit.yml
+++ b/.github/workflows/unit.yml
@@ -22,18 +22,18 @@ jobs:
       - name: Run tests
         env:
           TIMESCALE_FACTOR: 10
-        run: ginkgo -r -v -cover -randomizeAllSpecs -randomizeSuites -trace -skipPackage integrationtests,benchmark
+        run: ginkgo -r -v -cover -randomizeAllSpecs -randomizeSuites -trace -skipPackage integrationtests
       - name: Run tests (32 bit)
         if: ${{ matrix.os != 'macos' }} # can't run 32 bit tests on OSX.
         env:
           TIMESCALE_FACTOR: 10
           GOARCH: 386
-        run: ginkgo -r -v -cover -coverprofile coverage.txt -outputdir . -randomizeAllSpecs -randomizeSuites -trace -skipPackage integrationtests,benchmark
+        run: ginkgo -r -v -cover -coverprofile coverage.txt -outputdir . -randomizeAllSpecs -randomizeSuites -trace -skipPackage integrationtests
       - name: Run tests with race detector
         if: ${{ matrix.os == 'ubuntu' }} # speed things up. Windows and OSX VMs are slow
         env:
           TIMESCALE_FACTOR: 20
-        run: ginkgo -r -v -race -randomizeAllSpecs -randomizeSuites -trace -skipPackage integrationtests,benchmark
+        run: ginkgo -r -v -race -randomizeAllSpecs -randomizeSuites -trace -skipPackage integrationtests
       - name: Upload coverage to Codecov
         uses: codecov/codecov-action@v1
         with:
diff --git a/benchmark/benchmark_suite_test.go b/benchmark/benchmark_suite_test.go
deleted file mode 100644
index d2334429..00000000
--- a/benchmark/benchmark_suite_test.go
+++ /dev/null
@@ -1,27 +0,0 @@
-package benchmark
-
-import (
-	"flag"
-	"math/rand"
-	"testing"
-
-	. "github.com/onsi/ginkgo"
-	. "github.com/onsi/gomega"
-)
-
-func TestBenchmark(t *testing.T) {
-	RegisterFailHandler(Fail)
-	RunSpecs(t, "Benchmark Suite")
-}
-
-var size int // file size in MB, will be read from flags
-
-func init() {
-	flag.IntVar(&size, "size", 50, "data length (in MB)")
-}
-
-var _ = BeforeSuite(func() {
-	rand.Seed(GinkgoRandomSeed())
-
-	flag.Parse()
-})
diff --git a/benchmark/benchmark_test.go b/benchmark/benchmark_test.go
deleted file mode 100644
index 451e6dbb..00000000
--- a/benchmark/benchmark_test.go
+++ /dev/null
@@ -1,92 +0,0 @@
-package benchmark
-
-import (
-	"bytes"
-	"context"
-	"crypto/tls"
-	"fmt"
-	"io"
-	"math/rand"
-	"net"
-
-	quic "github.com/lucas-clemente/quic-go"
-	"github.com/lucas-clemente/quic-go/internal/protocol"
-	"github.com/lucas-clemente/quic-go/internal/testdata"
-
-	. "github.com/onsi/ginkgo"
-	. "github.com/onsi/gomega"
-)
-
-var _ = Describe("Benchmarks", func() {
-	for i := range protocol.SupportedVersions {
-		version := protocol.SupportedVersions[i]
-
-		Context(fmt.Sprintf("with version %s", version), func() {
-			var data []byte
-			var dataLen int
-
-			BeforeEach(func() {
-				dataLen = size * /* MB */ 1e6
-				data = make([]byte, dataLen)
-				rand.Read(data) // no need to check for an error. math.Rand.Read never errors
-			})
-
-			Measure("transferring a file", func(b Benchmarker) {
-				var ln quic.Listener
-				serverAddr := make(chan net.Addr)
-				handshakeChan := make(chan struct{})
-				// start the server
-				go func() {
-					defer GinkgoRecover()
-					var err error
-					tlsConf := testdata.GetTLSConfig()
-					tlsConf.NextProtos = []string{"benchmark"}
-					ln, err = quic.ListenAddr(
-						"localhost:0",
-						tlsConf,
-						&quic.Config{Versions: []protocol.VersionNumber{version}},
-					)
-					Expect(err).ToNot(HaveOccurred())
-					serverAddr <- ln.Addr()
-					sess, err := ln.Accept(context.Background())
-					Expect(err).ToNot(HaveOccurred())
-					// wait for the client to complete the handshake before sending the data
-					// this should not be necessary, but due to timing issues on the CIs, this is necessary to avoid sending too many undecryptable packets
-					<-handshakeChan
-					str, err := sess.OpenStream()
-					Expect(err).ToNot(HaveOccurred())
-					_, err = str.Write(data)
-					Expect(err).ToNot(HaveOccurred())
-					err = str.Close()
-					Expect(err).ToNot(HaveOccurred())
-				}()
-
-				// start the client
-				addr := <-serverAddr
-				sess, err := quic.DialAddr(
-					addr.String(),
-					&tls.Config{InsecureSkipVerify: true, NextProtos: []string{"benchmark"}},
-					&quic.Config{Versions: []protocol.VersionNumber{version}},
-				)
-				Expect(err).ToNot(HaveOccurred())
-				close(handshakeChan)
-				str, err := sess.AcceptStream(context.Background())
-				Expect(err).ToNot(HaveOccurred())
-
-				buf := &bytes.Buffer{}
-				// measure the time it takes to download the dataLen bytes
-				// note we're measuring the time for the transfer, i.e. excluding the handshake
-				runtime := b.Time("transfer time", func() {
-					_, err := io.Copy(buf, str)
-					Expect(err).NotTo(HaveOccurred())
-				})
-				Expect(buf.Bytes()).To(Equal(data))
-
-				b.RecordValue("transfer rate [MB/s]", float64(dataLen)/1e6/runtime.Seconds())
-
-				ln.Close()
-				sess.CloseWithError(0, "")
-			}, 3)
-		})
-	}
-})