diff --git a/benchmark/Manifest.toml b/benchmark/Manifest.toml
new file mode 100644
index 0000000..f3b6d51
--- /dev/null
+++ b/benchmark/Manifest.toml
@@ -0,0 +1,104 @@
+[[Base64]]
+uuid = "2a0f44e3-6c83-55bd-87e4-b1978d98bd5f"
+
+[[BenchmarkTools]]
+deps = ["JSON", "Printf", "Statistics", "Test"]
+git-tree-sha1 = "e686f1754227e4748259f400839b83a1e8773e02"
+uuid = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
+version = "0.4.1"
+
+[[Dates]]
+deps = ["Printf"]
+uuid = "ade2ca70-3891-5945-98fb-dc099432e06a"
+
+[[Distributed]]
+deps = ["Random", "Serialization", "Sockets"]
+uuid = "8ba89e20-285c-5b6f-9357-94700520ee1b"
+
+[[InteractiveUtils]]
+deps = ["Markdown"]
+uuid = "b77e0a4c-d291-57a0-90e8-8db25a27a240"
+
+[[JSON]]
+deps = ["Dates", "Distributed", "Mmap", "Sockets", "Test", "Unicode"]
+git-tree-sha1 = "1f7a25b53ec67f5e9422f1f551ee216503f4a0fa"
+uuid = "682c06a0-de6a-54ab-a142-c8b1cf79cde6"
+version = "0.20.0"
+
+[[LibGit2]]
+uuid = "76f85450-5226-5b5a-8eaa-529ad045b433"
+
+[[Libdl]]
+uuid = "8f399da3-3557-5675-b5ff-fb832c97cbdb"
+
+[[LinearAlgebra]]
+deps = ["Libdl"]
+uuid = "37e2e46d-f89d-539d-b4ee-838fcccc9c8e"
+
+[[Logging]]
+uuid = "56ddb016-857b-54e1-b83d-db4d58db5568"
+
+[[Markdown]]
+deps = ["Base64"]
+uuid = "d6f4376e-aef5-505a-96c1-9c027394607a"
+
+[[Mmap]]
+uuid = "a63ad114-7e13-5084-954f-fe012c677804"
+
+[[Pkg]]
+deps = ["Dates", "LibGit2", "Markdown", "Printf", "REPL", "Random", "SHA", "UUIDs"]
+uuid = "44cfe95a-1eb2-52ea-b672-e2afdf69b78f"
+
+[[PkgBenchmark]]
+deps = ["BenchmarkTools", "Dates", "InteractiveUtils", "JSON", "LibGit2", "Pkg", "Printf", "ProgressMeter", "Random", "Statistics", "Test"]
+git-tree-sha1 = "717f93d4cfcbd5b5dadbbb74071541c725ba6552"
+repo-rev = "master"
+repo-url = "https://github.com/JuliaCI/PkgBenchmark.jl.git"
+uuid = "32113eaa-f34f-5b0d-bd6c-c81e245fc73d"
+version = "0.2.0+"
+
+[[Printf]]
+deps = ["Unicode"]
+uuid = "de0858da-6303-5e67-8744-51eddeeeb8d7"
+
+[[ProgressMeter]]
+deps = ["Distributed", "Printf", "Random", "Test"]
+git-tree-sha1 = "48058bc11607676e5bbc0b974af79106c6200787"
+uuid = "92933f4c-e287-5a05-a399-4b506db050ca"
+version = "0.9.0"
+
+[[REPL]]
+deps = ["InteractiveUtils", "Markdown", "Sockets"]
+uuid = "3fa0cd96-eef1-5676-8a61-b3b8758bbffb"
+
+[[Random]]
+deps = ["Serialization"]
+uuid = "9a3f8284-a2c9-5f02-9a11-845980a1fd5c"
+
+[[SHA]]
+uuid = "ea8e919c-243c-51af-8825-aaa63cd721ce"
+
+[[Serialization]]
+uuid = "9e88b42a-f829-5b0c-bbe9-9e923198166b"
+
+[[Sockets]]
+uuid = "6462fe0b-24de-5631-8697-dd941f90decc"
+
+[[SparseArrays]]
+deps = ["LinearAlgebra", "Random"]
+uuid = "2f01184e-e22b-5df5-ae63-d93ebab69eaf"
+
+[[Statistics]]
+deps = ["LinearAlgebra", "SparseArrays"]
+uuid = "10745b16-79ce-11e8-11f9-7d13ad32a3b2"
+
+[[Test]]
+deps = ["Distributed", "InteractiveUtils", "Logging", "Random"]
+uuid = "8dfed614-e22c-5e08-85e1-65c5234f0b40"
+
+[[UUIDs]]
+deps = ["Random", "SHA"]
+uuid = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
+
+[[Unicode]]
+uuid = "4ec0a83e-493e-50e2-b9ac-8f72acf5a8f5"
diff --git a/benchmark/Project.toml b/benchmark/Project.toml
new file mode 100644
index 0000000..16acc49
--- /dev/null
+++ b/benchmark/Project.toml
@@ -0,0 +1,3 @@
+[deps]
+BenchmarkTools = "6e4b80f9-dd63-53aa-95a3-0cdb28fa8baf"
+PkgBenchmark = "32113eaa-f34f-5b0d-bd6c-c81e245fc73d"
diff --git a/benchmark/REQUIRE b/benchmark/REQUIRE
new file mode 100644
index 0000000..bf369b9
--- /dev/null
+++ b/benchmark/REQUIRE
@@ -0,0 +1 @@
+BenchmarkTools
diff --git a/benchmark/benchmarks.jl b/benchmark/benchmarks.jl
new file mode 100644
index 0000000..7c60957
--- /dev/null
+++ b/benchmark/benchmarks.jl
@@ -0,0 +1,90 @@
+# Decimal Representation Comparisons
+#
+# This benchmark compares the performance of several numeric representations, over various
+# numeric operations (+,-,*,/,÷...) on large arrays of numbers, in order to guide
+# decision-making about how to represent fixed-decimal numbers.
+#
+# It compares fixed-decimal types against the builtin Int and Float types of various sizes.
+# The output is written to a .csv file in the same directory as this file.
+
+# TODO: remove this file once BenchmarkTools has a built-in solution for diffing two
+# @benchmarkable runs
+using Pkg
+Pkg.activate(@__DIR__)
+
+using FixedPointDecimals
+using Random
+using BenchmarkTools, Statistics
+
+# Define a parent BenchmarkGroup to contain our suite
+const SUITE = BenchmarkGroup()
+const N = parse(Int, get(ENV, "BENCH_NUM_ITERS", "1000"))
+
+benchtypes = [
+    FixedPointDecimals.FixedDecimal{Int32,  2},
+    FixedPointDecimals.FixedDecimal{Int64,  2},
+    FixedPointDecimals.FixedDecimal{Int128, 2},
+]
+
+identity1(a,_) = a
+allops = (*, /, +, ÷, identity1)
+
+prettytype(::Type{FixedPointDecimals.FixedDecimal{T,f}}) where {T,f} = "FD{$T,$f}"
+prettytype(::Type{FixedPointDecimals.FixedDecimal{T,f}}) where {T<:Union{Int32,Int64},f} = "FD{ $T,$f}"
+opname(f) = string(Symbol(f))
+opname(f::typeof(identity1)) = "identity"
+
+# --------- Define benchmark functions -------------
+# Some care is taken here to prevent the compiler from optimizing away the operations:
+#  - Marked @noinline so the constants we pass in aren't available to the optimizer.
+#  - We take `a` and `out` as parameters so that their values aren't available when
+#      compiling this function.
+#  - `out` is a Ref{T} so that this function will have side effects. We use an output
+#      parameter instead of returning the value directly  so that it will play nicely with
+#      the `@benchmark` macro which returns the benchmark results as an object.
+#  - `T` and `op` _should_ be available as compile-time constants, since we don't want to be
+#      measuring the time it takes to read from global variables.
+@noinline function benchmark(::Type{T}, op, a::T, n, out::Ref{T}) where {T}
+    for _ in 1:n
+        tmp = op(a,a)
+        out[] += tmp
+        a += one(T)
+    end
+end
+
+@noinline function baseline(::Type{T}, a::T, n, out::Ref{T}) where {T}
+    for _ in 1:n
+        tmp = a
+        out[] += tmp
+        a += one(T)
+    end
+end
+
+# Define the benchmark structure
+for op in allops
+    SUITE[opname(op)] = BenchmarkGroup()
+    for T in benchtypes
+        SUITE[opname(op)][prettytype(T)] = BenchmarkGroup(["base", "bench"])
+    end
+end
+
+for op in allops
+    println()
+    println("$op")
+    for T in benchtypes
+        print("$T ")
+
+        initial_value = zero(T)
+        a = one(T)
+
+        # For some reason this is necessary to eliminate mysterious "1 allocation"
+        fbase = @eval (out::Ref{$T})->baseline($T, $a, $N, out)
+        fbench = @eval (out::Ref{$T})->benchmark($T, $op, $a, $N, out)
+
+        # Run the benchmark
+        outbase = Ref(initial_value)
+        SUITE[opname(op)][prettytype(T)]["base"] = @benchmarkable $fbase($outbase) evals=1 setup=($outbase[]=$initial_value)
+        outbench = Ref(initial_value)
+        SUITE[opname(op)][prettytype(T)]["bench"] = @benchmarkable $fbench($outbench) evals=1 setup=($outbench[]=$initial_value)
+    end
+end
diff --git a/benchmark/runbench.jl b/benchmark/runbench.jl
new file mode 100644
index 0000000..45658c8
--- /dev/null
+++ b/benchmark/runbench.jl
@@ -0,0 +1,85 @@
+module FixedPointDecimals_RunBench
+
+using Pkg
+
+Pkg.activate(@__DIR__)
+using PkgBenchmark, BenchmarkTools, Statistics
+
+const N = 1_000
+
+import Base: -, /
+function -(a::BenchmarkTools.TrialEstimate, b::BenchmarkTools.TrialEstimate)
+    ttol = max(params(a).time_tolerance, params(b).time_tolerance)
+    mtol = max(params(a).memory_tolerance, params(b).memory_tolerance)
+    p = BenchmarkTools.Parameters(params(a); time_tolerance = ttol, memory_tolerance = mtol)
+    return BenchmarkTools.TrialEstimate(p, -(time(a), time(b)), -(gctime(a), gctime(b)),
+                                        -(memory(a), memory(b)), -(allocs(a), allocs(b)))
+end
+function /(a::BenchmarkTools.TrialEstimate, b::Int)
+    ttol = params(a).time_tolerance / b
+    mtol = params(a).memory_tolerance / b
+    p = BenchmarkTools.Parameters(params(a); time_tolerance = ttol, memory_tolerance = mtol)
+    return BenchmarkTools.TrialEstimate(p, time(a)/b, gctime(a)/b,
+                                        round(memory(a)/b), round(allocs(a)/b))
+end
+
+function postprocess(results::BenchmarkGroup)
+    for (op, op_group) in results.data
+        op_results = op_group.data
+        for (type, type_group) in op_results
+            benchresults = type_group.data
+            if op == "identity"
+                # For :identity, bench and base are identical so we don't want to subtract.
+                op_results[type] = median(benchresults["bench"]) / N
+            else
+                op_results[type] = median(benchresults["bench"])/N - median(benchresults["base"])/N
+            end
+        end
+    end
+    results
+end
+function postprocess_no_div(results::BenchmarkGroup)
+    for (op, op_group) in results.data
+        op_results = op_group.data
+        for (type, type_group) in op_results
+            benchresults = type_group.data
+            if op == "identity"
+                # For :identity, bench and base are identical so we don't want to subtract.
+                op_results[type] = median(benchresults["bench"])
+            else
+                op_results[type] = median(benchresults["bench"]) - median(benchresults["base"])
+            end
+        end
+    end
+    results
+end
+
+
+function runbench()
+    rm(joinpath(@__DIR__, "tune.json"))  # Remove the existing tune.json file.
+    bench_results = withenv("BENCH_NUM_ITERS"=>string(N)) do
+        benchmarkpkg("FixedPointDecimals"; postprocess=postprocess)
+    end
+
+    export_markdown(joinpath(@__DIR__, "results.md"), bench_results)
+    return bench_results
+end
+
+function judgebench(target::Union{String, BenchmarkConfig}, baseline::Union{String, BenchmarkConfig},
+            postprocess_fn=postprocess_no_div)
+    try rm(joinpath(@__DIR__, "tune.json")) catch end  # Remove the existing tune.json file.
+    bench_results = withenv("BENCH_NUM_ITERS"=>string(N)) do
+        if postprocess_fn != nothing
+            judge("FixedPointDecimals", target, baseline; f=identity, postprocess=postprocess_fn)
+        else
+            judge("FixedPointDecimals", target, baseline)
+        end
+    end
+    export_markdown(joinpath(@__DIR__, "judge.md"), bench_results)
+    return bench_results
+end
+function judgebench(baseline::Union{String, BenchmarkConfig})
+    judgebench(BenchmarkConfig(), baseline)
+end
+
+end
diff --git a/benchmark/subtract-benchmarks.jl b/benchmark/subtract-benchmarks.jl
new file mode 100644
index 0000000..1338f7f
--- /dev/null
+++ b/benchmark/subtract-benchmarks.jl
@@ -0,0 +1,10 @@
+import PkgBenchmark
+function PkgBenchmark._run(b::PkgBenchmark.BenchmarkTools.BenchmarkDiff, p::PkgBenchmark.BenchmarkTools.Parameters = b.params;
+    prog = nothing, verbose::Bool = false, pad = "", hierarchy = [], kwargs...)
+    res = BenchmarkTools.run_result(b, p; kwargs...)[1]
+    if prog != nothing
+        indent = 0
+        ProgressMeter.next!(prog; showvalues = [map(id -> ("  "^(indent += 1) * "[$(id[2])/$(id[3])]", id[1]), hierarchy)...])
+    end
+    return res
+end