diff --git a/examples/nvperf.jl b/examples/nvperf.jl
new file mode 100644
index 0000000000..8b4e8ce78b
--- /dev/null
+++ b/examples/nvperf.jl
@@ -0,0 +1,13 @@
+using CUDA
+
+NVPERF.initialize()
+CUPTI.initialize_profiler()
+
+avail = CUPTI.counter_availability()
+chip = first(NVPERF.supported_chips())
+
+evaluator = NVPERF.CUDAMetricsEvaluator(chip, avail)
+
+NVPERF.list_metrics(evaluator)
+
+@show NVPERF.MetricEvalRequest(me, "dram__bytes.sum.per_second")
\ No newline at end of file
diff --git a/lib/cupti/wrappers.jl b/lib/cupti/wrappers.jl
index b2456f83ba..009f54fe82 100644
--- a/lib/cupti/wrappers.jl
+++ b/lib/cupti/wrappers.jl
@@ -4,8 +4,34 @@ function version()
     VersionNumber(version_ref[])
 end
 
-# function counter_availabiltiy(ctx = context())
-#     params = Ref(CUpti_Profiler_GetCounterAvailability_Params(
-#         CUpti_Profiler_GetCounterAvailability_Params_STRUCT_SIZE,
-#         C_NULL, ctx, 0
-#     ))
+    # params = Ref(CUpti_Profiler_CounterDataImage_CalculateSize_Params(
+    #     CUpti_Profiler_CounterDataImage_CalculateSize_Params_STRUCT_SIZE,
+    #     C_NULL, 0, C_NULL, 0))
+    # cuptiProfilerCounterDataImageCalculateSize(params)
+
+
+function initialize_profiler()
+    params = Ref(CUpti_Profiler_Initialize_Params(
+        CUpti_Profiler_Initialize_Params_STRUCT_SIZE,
+        C_NULL))
+    cuptiProfilerInitialize(params)
+end
+
+function counter_availability(ctx = context())
+    # 1. Query size
+    params = Ref(CUpti_Profiler_GetCounterAvailability_Params(
+        CUpti_Profiler_GetCounterAvailability_Params_STRUCT_SIZE,
+        C_NULL, ctx.handle, 0, C_NULL))
+    cuptiProfilerGetCounterAvailability(params)
+
+    sz = params[].counterAvailabilityImageSize
+    buffer = Vector{UInt8}(undef, sz)
+
+    GC.@preserve buffer begin
+        params = Ref(CUpti_Profiler_GetCounterAvailability_Params(
+        CUpti_Profiler_GetCounterAvailability_Params_STRUCT_SIZE,
+        C_NULL, ctx.handle, sz, pointer(buffer)))
+        cuptiProfilerGetCounterAvailability(params)
+    end
+    return buffer
+end
diff --git a/lib/nvperf/wrappers.jl b/lib/nvperf/wrappers.jl
index 46c32a185f..d2ef392fa6 100644
--- a/lib/nvperf/wrappers.jl
+++ b/lib/nvperf/wrappers.jl
@@ -16,17 +16,132 @@ function supported_chips()
     return names
 end
 
-function scratch_buffer(chipName)
-    GC.@preserve chipName begin
+function scratch_buffer(chipName, counter_availability)
+    GC.@preserve chipName counter_availability begin
         params = Ref(NVPW_CUDA_MetricsEvaluator_CalculateScratchBufferSize_Params(
             NVPW_CUDA_MetricsEvaluator_CalculateScratchBufferSize_Params_STRUCT_SIZE,
-            C_NULL, pointer(chipName), C_NULL, 0
+            C_NULL, pointer(chipName), pointer(counter_availability), 0
         ))
         NVPW_CUDA_MetricsEvaluator_CalculateScratchBufferSize(params)
-        params[].scrachBufferSize
+        sz = params[].scratchBufferSize
     end
+    return Vector{UInt8}(undef, sz)
 end
 
-# function list_metrics(chipName)
+abstract type MetricsEvaluator end
 
-# end
\ No newline at end of file
+mutable struct CUDAMetricsEvaluator <: MetricsEvaluator
+    handle::Ptr{NVPW_MetricsEvaluator}
+    scratch::Vector{UInt8}
+    availability::Vector{UInt8}
+    chip::String
+
+    function CUDAMetricsEvaluator(chip, availability)
+        scratch = scratch_buffer(chip, availability)
+
+        GC.@preserve chip availability scratch begin
+            params = Ref(NVPW_CUDA_MetricsEvaluator_Initialize_Params(
+                NVPW_CUDA_MetricsEvaluator_Initialize_Params_STRUCT_SIZE,
+                C_NULL, pointer(scratch), length(scratch), pointer(chip),
+                pointer(availability), C_NULL, 0, C_NULL))
+            
+            NVPW_CUDA_MetricsEvaluator_Initialize(params)
+            this =  new(params[].pMetricsEvaluator, scratch, availability, chip)
+        end
+        finalizer(destroy, this)
+        return this
+    end
+end
+Base.unsafe_convert(::Type{Ptr{NVPW_MetricsEvaluator}}, me::CUDAMetricsEvaluator) = me.handle
+
+
+function destroy(me::MetricsEvaluator)
+    GC.@preserve me begin
+        params = Ref(NVPW_MetricsEvaluator_Destroy_Params(
+            NVPW_MetricsEvaluator_Destroy_Params_STRUCT_SIZE,
+            C_NULL, Base.unsafe_convert(Ptr{NVPW_MetricsEvaluator}, me)
+        ))
+        NVPW_MetricsEvaluator_Destroy(params)
+    end
+    return nothing
+end
+
+struct MetricsIterator
+    me::MetricsEvaluator
+    type::NVPW_MetricType
+    names::Ptr{Cchar}
+    indices::Ptr{Csize_t}
+    numMetrics::Csize_t
+
+    function MetricsIterator(me, type)
+        GC.@preserve me begin
+            params = Ref(NVPW_MetricsEvaluator_GetMetricNames_Params(
+                NVPW_MetricsEvaluator_GetMetricNames_Params_STRUCT_SIZE,
+                C_NULL, Base.unsafe_convert(Ptr{NVPW_MetricsEvaluator}, me), type, C_NULL, C_NULL, 0))
+            NVPW_MetricsEvaluator_GetMetricNames(params)
+
+            names = Ptr{Cchar}(params[].pMetricNames)
+            indices = params[].pMetricNameBeginIndices
+
+            return new(me, type, names, indices, params[].numMetrics)
+        end
+    end
+end
+
+Base.length(metrics::MetricsIterator) = metrics.numMetrics
+Base.eltype(::MetricsIterator) = String
+
+function Base.iterate(metrics::MetricsIterator, state=1)
+    if state <= metrics.numMetrics
+        name = unsafe_string(metrics.names + unsafe_load(metrics.indices, state))
+        return (name, state+1)
+    else
+        return nothing
+    end
+end
+
+function list_metrics(me::MetricsEvaluator)
+    for i in 0:(NVPW_METRIC_TYPE__COUNT-1)
+        type = NVPW_MetricType(i)
+
+        for metric in MetricsIterator(me, type)
+            @show metric
+        end
+    end
+end
+
+function submetrics(me::MetricsEvaluator, type)
+    GC.@preserve me begin
+        params = Ref(NVPW_MetricsEvaluator_GetSupportedSubmetrics_Params(
+            NVPW_MetricsEvaluator_GetSupportedSubmetrics_Params_STRUCT_SIZE,
+            C_NULL, Base.unsafe_convert(Ptr{NVPW_MetricsEvaluator}, me), type, C_NULL, 0))
+        NVPW_MetricsEvaluator_GetSupportedSubmetrics(params)
+        unsafe_wrap(Array, params[].pSupportedSubmetrics, params[].numSupportedSubmetrics)
+    end
+end
+
+# TODO rollup to string
+# TODO submetric to string
+
+# function submetric(m)
+#     if m == NVPW_SUBMETRIC_PEAK_SUSTAINED
+#         return ".peak_sustained"
+#     elseif 
+
+# MetricTypeAndIndex
+
+struct MetricEvalRequest
+    data::NVPW_MetricEvalRequest
+
+    function MetricEvalRequest(me::MetricsEvaluator, name)
+        eval_request = Ref{NVPW_MetricEvalRequest}()
+        GC.@preserve me name eval_request begin
+            params = Ref(NVPW_MetricsEvaluator_ConvertMetricNameToMetricEvalRequest_Params(
+                NVPW_MetricsEvaluator_ConvertMetricNameToMetricEvalRequest_Params_STRUCT_SIZE,
+                C_NULL, Base.unsafe_convert(Ptr{NVPW_MetricsEvaluator}, me), pointer(name),
+                Base.unsafe_convert(Ptr{NVPW_MetricEvalRequest}, eval_request), NVPW_MetricEvalRequest_STRUCT_SIZE))
+            NVPW_MetricsEvaluator_ConvertMetricNameToMetricEvalRequest(params)
+            return new(eval_request[])
+        end
+    end
+end