diff --git a/.buildkite/pipeline.yml b/.buildkite/pipeline.yml
index 6e229c6ea..7e2269723 100644
--- a/.buildkite/pipeline.yml
+++ b/.buildkite/pipeline.yml
@@ -16,7 +16,7 @@ steps:
       julia -e 'println("--- :julia: Developing CUDA")
                 using Pkg
                 url="https://github.com/christiangnrd/CUDA.jl"
-                rev="intrinsics"
+                rev="intrinsicsnew"
                 subdirs = ["CUDACore", "CUDATools", "lib/cusolver", "lib/curand", "lib/cublas", "lib/cufft", "lib/cusparse", "lib/cupti", "lib/nvml", "lib/custatevec", "lib/cudnn", "lib/cutensor", "lib/cutensornet"]
                 Pkg.add([(;url, rev); [(; url, rev, subdir) for subdir in subdirs]])'
       julia -e 'println("--- :julia: Instantiating project")
@@ -81,7 +81,7 @@ steps:
     command: |
       julia -e 'println("--- :julia: Developing Metal")
                 using Pkg
-                Pkg.add(url="https://github.com/JuliaGPU/Metal.jl", rev="kaintr")'
+                Pkg.add(url="https://github.com/JuliaGPU/Metal.jl", rev="kaintrnew")'
       julia -e 'println("--- :julia: Instantiating project")
                 using Pkg
                 Pkg.develop(; path=pwd())' || exit 3
@@ -113,8 +113,8 @@ steps:
     command: |
       julia -e 'println("--- :julia: Developing oneAPI")
                 using Pkg
-                Pkg.add(url="https://github.com/christiangnrd/oneAPI.jl", rev="intrinsics")
-                Pkg.add(url="https://github.com/christiangnrd/AcceleratedKernels.jl", rev="ka0.10simple")'
+                Pkg.add(url="https://github.com/christiangnrd/oneAPI.jl", rev="intrinsicsnew")
+                Pkg.add(url="https://github.com/JuliaGPU/AcceleratedKernels.jl", rev="main")'
       julia -e 'println("--- :julia: Instantiating project")
                 using Pkg
                 Pkg.develop(; path=pwd())' || exit 3
@@ -146,10 +146,10 @@ steps:
     command: |
       julia -e 'println("--- :julia: Developing AMDGPU")
                 using Pkg
-                Pkg.add(url="https://github.com/christiangnrd/AcceleratedKernels.jl", rev="ka0.10simple")'
+                Pkg.add(url="https://github.com/JuliaGPU/AcceleratedKernels.jl", rev="main")'
       julia -e '
                 using Pkg
-                Pkg.add(url="https://github.com/christiangnrd/AMDGPU.jl", rev="intrinsics")
+                Pkg.add(url="https://github.com/christiangnrd/AMDGPU.jl", rev="intrinsicsnew")
                 println("--- :julia: Instantiating project")
                 Pkg.develop(; path=pwd())' || exit 3
 
@@ -181,7 +181,7 @@ steps:
     command: |
       julia -e 'println("--- :julia: Developing OpenCL")
                 using Pkg
-                Pkg.add(url="https://github.com/christiangnrd/OpenCL.jl", rev="intrinsics")
+                Pkg.add(url="https://github.com/christiangnrd/OpenCL.jl", rev="intrinsicsnew")
                 Pkg.develop(; name="SPIRVIntrinsics")'
       julia -e 'println("--- :julia: Instantiating project")
                 using Pkg
diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index 64c294c97..229d78001 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -182,7 +182,7 @@ jobs:
         run: |
           julia -e 'println("--- :julia: Developing OpenCL")
                 using Pkg
-                Pkg.add(url="https://github.com/christiangnrd/OpenCL.jl", rev="intrinsics")
+                Pkg.add(url="https://github.com/christiangnrd/OpenCL.jl", rev="intrinsicsnew")
                 Pkg.develop(; name="SPIRVIntrinsics")'
       - name: "Instantiating project"
         run: |
diff --git a/src/interface.jl b/src/interface.jl
index f2c21d2cc..6748c227a 100644
--- a/src/interface.jl
+++ b/src/interface.jl
@@ -102,6 +102,78 @@ Returns the unique group ID.
 """
 function get_group_id end
 
+"""
+    get_sub_group_size()::UInt32
+
+Returns the number of work-items in the sub-group.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    @device_override get_sub_group_size()::UInt32
+    ```
+"""
+function get_sub_group_size end
+
+"""
+    get_max_sub_group_size()::UInt32
+
+Returns the maximum sub-group size for sub-groups in the current workgroup.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    @device_override get_max_sub_group_size()::UInt32
+    ```
+"""
+function get_max_sub_group_size end
+
+"""
+    get_num_sub_groups()::UInt32
+
+Returns the number of sub-groups in the current workgroup.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    @device_override get_num_sub_groups()::UInt32
+    ```
+"""
+function get_num_sub_groups end
+
+"""
+    get_sub_group_id()::UInt32
+
+Returns the sub-group ID within the work-group.
+
+!!! note
+    1-based.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    @device_override get_sub_group_id()::UInt32
+    ```
+"""
+function get_sub_group_id end
+
+"""
+    get_sub_group_local_id()::UInt32
+
+Returns the work-item ID within the current sub-group.
+
+!!! note
+    1-based.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    @device_override get_sub_group_local_id()::UInt32
+    ```
+"""
+function get_sub_group_local_id end
+
+
 """
     localmemory(::Type{T}, dims)
 
@@ -116,6 +188,40 @@ Declare memory that is local to a workgroup.
 """
 localmemory(::Type{T}, dims) where {T} = localmemory(T, Val(dims))
 
+"""
+    shfl_down(val::T, offset::Integer) where T
+
+Read `val` from a lane with higher id given by `offset`.
+
+!!! note
+    `shfl_down` must be encountered by all workitems of a sub-group executing the kernel or by none at all.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    @device_override shfl_down(val::T, offset::Integer) where T
+    ```
+    As well as the on-device functionality.
+
+    This implementation **must** be synchronizing.
+    That is, kernels using this function can safely assume that
+    they do **not** need a `sub_group_barrier` before calling
+    this function.
+"""
+function shfl_down end
+
+"""
+    shfl_down_types(::Backend)::Vector{DataType}
+
+Returns a vector of `DataType`s supported on `backend`
+
+!!! note
+    Backend implementations **must** implement this function
+    only if they support `shfl_down` for any types.
+"""
+shfl_down_types(::Backend) = DataType[]
+
+
 """
     barrier()
 
@@ -139,6 +245,29 @@ function barrier()
     error("Group barrier used outside kernel or not captured")
 end
 
+"""
+    sub_group_barrier()
+
+After a `sub_group_barrier()` call, all read and writes to global and local memory
+from each thread in the sub-group are visible in from all other threads in the
+sub-group.
+
+This does **not** guarantee that a write from a thread in a certain sub-group will
+be visible to a thread in a different sub-group.
+
+!!! note
+    `sub_group_barrier()` must be encountered by all workitems of a sub-group executing the kernel or by none at all.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    @device_override sub_group_barrier()
+    ```
+"""
+function sub_group_barrier()
+    error("Sub-group barrier used outside kernel or not captured")
+end
+
 """
     _print(args...)
 
@@ -220,6 +349,22 @@ kernel launch with too big a workgroup is attempted.
 """
 function max_work_group_size end
 
+"""
+    sub_group_size(backend)::Int
+
+Returns a reasonable sub-group size supported by the currently
+active device for the specified backend. This would typically
+be 32, or 64 for devices that don't support 32.
+
+!!! note
+    Backend implementations **must** implement:
+    ```
+    sub_group_size(backend::NewBackend)::Int
+    ```
+    As well as the on-device functionality.
+"""
+function sub_group_size end
+
 """
     multiprocessor_count(backend::NewBackend)::Int
 
diff --git a/src/pocl/backend.jl b/src/pocl/backend.jl
index 3b0fdb3fa..a72eae309 100644
--- a/src/pocl/backend.jl
+++ b/src/pocl/backend.jl
@@ -9,6 +9,8 @@ using SPIRV_LLVM_Backend_jll, SPIRV_Tools_jll
 import KernelAbstractions as KA
 import KernelAbstractions.KernelInterface as KI
 
+import SPIRVIntrinsics
+
 import StaticArrays
 
 import Adapt
@@ -227,10 +229,39 @@ end
 function KI.max_work_group_size(::POCLBackend)::Int
     return Int(device().max_work_group_size)
 end
+function KI.sub_group_size(::POCLBackend)::Int
+    # POCL can technically support any sub_group size.
+    #  Check for common values used on GPUs then
+    #  return 1 otherwise
+    sg_sizes = cl.device().sub_group_sizes
+    if 32 in sg_sizes
+        return 32
+    elseif 64 in sg_sizes
+        return 64
+    elseif 16 in sg_sizes
+        return 16
+    else
+        return 1
+    end
+end
 function KI.multiprocessor_count(::POCLBackend)::Int
     return Int(device().max_compute_units)
 end
 
+function KI.shfl_down_types(::POCLBackend)
+    res = copy(SPIRVIntrinsics.gentypes)
+
+    backend_extensions = cl.device().extensions
+    if "cl_khr_fp64" ∉ backend_extensions
+        res = setdiff(res, [Float64])
+    end
+    if "cl_khr_fp16" ∉ backend_extensions
+        res = setdiff(res, [Float16])
+    end
+
+    return res
+end
+
 ## Indexing Functions
 
 @device_override @inline function KI.get_local_id()
@@ -257,6 +288,16 @@ end
     return (; x = Int(get_global_size(1)), y = Int(get_global_size(2)), z = Int(get_global_size(3)))
 end
 
+@device_override KI.get_sub_group_size() = get_sub_group_size()
+
+@device_override KI.get_max_sub_group_size() = get_max_sub_group_size()
+
+@device_override KI.get_num_sub_groups() = get_num_sub_groups()
+
+@device_override KI.get_sub_group_id() = get_sub_group_id()
+
+@device_override KI.get_sub_group_local_id() = get_sub_group_local_id()
+
 @device_override @inline function KA.__validindex(ctx)
     if KA.__dynamic_checkbounds(ctx)
         I = @inbounds KA.expand(KA.__iterspace(ctx), get_group_id(1), get_local_id(1))
@@ -285,6 +326,14 @@ end
     work_group_barrier(POCL.LOCAL_MEM_FENCE | POCL.GLOBAL_MEM_FENCE)
 end
 
+@device_override @inline function KI.sub_group_barrier()
+    sub_group_barrier(POCL.LOCAL_MEM_FENCE | POCL.GLOBAL_MEM_FENCE)
+end
+
+@device_override function KI.shfl_down(val::T, offset::Integer) where {T}
+    sub_group_shuffle(val, get_sub_group_local_id() + offset)
+end
+
 @device_override @inline function KI._print(args...)
     POCL._print(args...)
 end
diff --git a/src/pocl/compiler/compilation.jl b/src/pocl/compiler/compilation.jl
index 8aee85447..e5eefd31c 100644
--- a/src/pocl/compiler/compilation.jl
+++ b/src/pocl/compiler/compilation.jl
@@ -1,6 +1,10 @@
 ## gpucompiler interface
 
-struct OpenCLCompilerParams <: AbstractCompilerParams end
+Base.@kwdef struct OpenCLCompilerParams <: AbstractCompilerParams
+    # request a fixed sub-group width via `intel_reqd_sub_group_size`
+    sub_group_size::Union{Nothing, Int} = nothing
+end
+
 const OpenCLCompilerConfig = CompilerConfig{SPIRVCompilerTarget, OpenCLCompilerParams}
 const OpenCLCompilerJob = CompilerJob{SPIRVCompilerTarget, OpenCLCompilerParams}
 
@@ -31,6 +35,11 @@ function GPUCompiler.finish_module!(
         job, mod, entry
     )
 
+    sg_size = job.config.params.sub_group_size
+    if sg_size !== nothing
+        metadata(entry)["intel_reqd_sub_group_size"] = MDNode([ConstantInt(Int32(sg_size))])
+    end
+
     # if this kernel uses our RNG, we should prime the shared state.
     # XXX: these transformations should really happen at the Julia IR level...
     if haskey(functions(mod), "julia.opencl.random_keys") && job.config.kernel
@@ -130,14 +139,17 @@ function compiler_config(dev::cl.Device; kwargs...)
     end
     return config
 end
-@noinline function _compiler_config(dev; kernel = true, name = nothing, always_inline = false, kwargs...)
+@noinline function _compiler_config(dev; kernel = true, name = nothing, always_inline = false, sub_group_size::Union{Nothing, Int} = 32, kwargs...)
     supports_fp16 = "cl_khr_fp16" in dev.extensions
     supports_fp64 = "cl_khr_fp64" in dev.extensions
 
+    if sub_group_size !== nothing && sub_group_size ∉ dev.sub_group_sizes
+        error("$sub_group_size is not a valid sub-group size for this device.")
+    end
 
     # create GPUCompiler objects
     target = SPIRVCompilerTarget(; supports_fp16, supports_fp64, validate = true, kwargs...)
-    params = OpenCLCompilerParams()
+    params = OpenCLCompilerParams(; sub_group_size)
     return CompilerConfig(target, params; kernel, name, always_inline)
 end
 
diff --git a/src/pocl/compiler/execution.jl b/src/pocl/compiler/execution.jl
index 595e7ff1a..fd22b57c8 100644
--- a/src/pocl/compiler/execution.jl
+++ b/src/pocl/compiler/execution.jl
@@ -4,7 +4,7 @@ export @opencl, clfunction, clconvert
 ## high-level @opencl interface
 
 const MACRO_KWARGS = [:launch]
-const COMPILER_KWARGS = [:kernel, :name, :always_inline, :validate]
+const COMPILER_KWARGS = [:kernel, :name, :always_inline, :validate, :sub_group_size]
 const LAUNCH_KWARGS = [:global_size, :local_size, :queue]
 
 macro opencl(ex...)
diff --git a/src/pocl/nanoOpenCL.jl b/src/pocl/nanoOpenCL.jl
index e7901ad07..eec955990 100644
--- a/src/pocl/nanoOpenCL.jl
+++ b/src/pocl/nanoOpenCL.jl
@@ -389,6 +389,8 @@ const CL_KERNEL_EXEC_INFO_SVM_PTRS = 0x11b6
 
 const CL_KERNEL_EXEC_INFO_SVM_FINE_GRAIN_SYSTEM = 0x11b7
 
+const CL_DEVICE_SUB_GROUP_SIZES_INTEL = 0x4108
+
 struct CLError <: Exception
     code::Cint
 end
@@ -951,6 +953,14 @@ devices(p::Platform) = devices(p, CL_DEVICE_TYPE_ALL)
         return tuple([Int(r) for r in result]...)
     end
 
+    if s == :sub_group_sizes
+        res_size = Ref{Csize_t}()
+        clGetDeviceInfo(d, CL_DEVICE_SUB_GROUP_SIZES_INTEL, C_NULL, C_NULL, res_size)
+        result = Vector{Csize_t}(undef, res_size[] ÷ sizeof(Csize_t))
+        clGetDeviceInfo(d, CL_DEVICE_SUB_GROUP_SIZES_INTEL, sizeof(result), result, C_NULL)
+        return tuple([Int(r) for r in result]...)
+    end
+
     if s == :max_image2d_shape
         width = Ref{Csize_t}()
         height = Ref{Csize_t}()
diff --git a/src/pocl/pocl.jl b/src/pocl/pocl.jl
index de1bd4b32..681683e70 100644
--- a/src/pocl/pocl.jl
+++ b/src/pocl/pocl.jl
@@ -42,6 +42,7 @@ end
 
 using GPUCompiler
 using LLVM, LLVM.Interop
+import LLVM: LLVM, MDNode, ConstantInt, metadata
 using SPIRV_LLVM_Backend_jll, SPIRV_Tools_jll
 using Adapt
 
diff --git a/test/interface.jl b/test/interface.jl
index a6fdd05bc..416ddcd1e 100644
--- a/test/interface.jl
+++ b/test/interface.jl
@@ -23,6 +23,65 @@ function test_interface_kernel(results)
     end
     return
 end
+struct SubgroupData
+    sub_group_size::UInt32
+    max_sub_group_size::UInt32
+    num_sub_groups::UInt32
+    sub_group_id::UInt32
+    sub_group_local_id::UInt32
+end
+function test_subgroup_kernel(results)
+    i = KI.get_global_id().x
+
+    if i <= length(results)
+        @inbounds results[i] = SubgroupData(
+            KI.get_sub_group_size(),
+            KI.get_max_sub_group_size(),
+            KI.get_num_sub_groups(),
+            KI.get_sub_group_id(),
+            KI.get_sub_group_local_id()
+        )
+    end
+    return
+end
+
+function shfl_down_test_kernel(a, b, ::Val{N}) where {N}
+    idx = KI.get_sub_group_local_id()
+
+    val = a[idx]
+
+    offset = 0x00000001
+    while offset < N
+        val += KI.shfl_down(val, offset)
+        offset <<= 1
+    end
+
+    KI.sub_group_barrier()
+
+    if idx == 1
+        b[idx] = val
+    end
+    return
+end
+
+function shfl_down_test_kernel(a, b, ::Val{N}) where {N}
+    idx = KI.get_sub_group_local_id()
+
+    val = a[idx]
+
+    offset = 0x00000001
+    while offset < N
+        val += KI.shfl_down(val, offset)
+        offset <<= 1
+    end
+
+    KI.sub_group_barrier()
+
+    if idx == 1
+        b[idx] = val
+    end
+    return
+end
 
 function interface_testsuite(backend, AT)
     @testset "KernelInterface Tests" begin
@@ -122,6 +181,60 @@ function interface_testsuite(backend, AT)
                 @test k_data.local_id == expected_local
             end
         end
-    end
+
+        # Used as a proxy for sub-group support
+        if !isempty(KI.shfl_down_types(backend()))
+            @testset "Sub-groups" begin
+                @test KI.sub_group_size(backend()) isa Int
+
+                # Test with small kernel
+                sg_size = KI.sub_group_size(backend())
+                sg_n = 2
+                workgroupsize = sg_size * sg_n
+                numworkgroups = 2
+                N = workgroupsize * numworkgroups
+
+                results = AT(Vector{SubgroupData}(undef, N))
+                kernel = KI.@kernel backend() launch = false test_subgroup_kernel(results)
+
+                kernel(results; workgroupsize, numworkgroups)
+                KernelAbstractions.synchronize(backend())
+
+                host_results = Array(results)
+
+                # Verify results make sense
+                for (i, sg_data) in enumerate(host_results)
+                    @test sg_data.sub_group_size == sg_size
+                    @test sg_data.max_sub_group_size == sg_size
+                    @test sg_data.num_sub_groups == sg_n
+
+                    # Group ID should be 1-based
+                    expected_sub_group = div(((i - 1) % workgroupsize), sg_size) + 1
+                    @test sg_data.sub_group_id == expected_sub_group
+
+                    # Local ID should be 1-based within group
+                    expected_sg_local = ((i - 1) % sg_size) + 1
+                    @test sg_data.sub_group_local_id == expected_sg_local
+                end
+            end
+            @testset "shfl_down" begin
+                @test !isempty(KI.shfl_down_types(backend()))
+                types_to_test = setdiff(KI.shfl_down_types(backend()), [Bool])
+                @testset "$T" for T in types_to_test
+                    N = KI.sub_group_size(backend())
+                    a = zeros(T, N)
+                    rand!(a, (0:1))
+
+                    dev_a = AT(a)
+                    dev_b = AT(zeros(T, N))
+
+                    KI.@kernel backend() workgroupsize = N shfl_down_test_kernel(dev_a, dev_b, Val(N))
+
+                    b = Array(dev_b)
+                    @test sum(a) ≈ b[1]
+                end
+            end
+        end
+    end # @testset "KernelInterface Tests" begin
     return nothing
 end