Skip to content

Working with simple Rationals like 2 // 10 is slow #2985

New issue
New issue
Open
Open
Working with simple Rationals like 2 // 10 is slow#2985
Labels
bugSomething isn't workingperformanceHow fast can we go?
@efaulhaber

Description

@efaulhaber
efaulhaber
opened on Nov 25, 2025

MWE:

julia> CUDA.@time map(y -> y * 2 / 10, CUDA.ones(5_000_000));
  0.163260 seconds (285.00 k CPU allocations: 14.995 MiB) (2 GPU allocations: 38.147 MiB, 0.02% memmgmt time)

julia> CUDA.@time map(y -> 2 // 10 * y, CUDA.ones(5_000_000));
  0.282400 seconds (322.23 k CPU allocations: 17.699 MiB) (2 GPU allocations: 38.147 MiB, 0.01% memmgmt time)

Note that this has nothing to do with FP64. Here is the FP64 version:

julia> CUDA.@time map(y -> 2 / 10 * y, CUDA.ones(5_000_000));
  0.159753 seconds (279.26 k CPU allocations: 14.702 MiB) (2 GPU allocations: 57.220 MiB, 0.13% memmgmt time)

Note that I get similar problems in larger kernels. Just using a simple Rational like this makes the whole kernel 2x slower.

LLVM for y * 2 / 10 
julia> CUDA.@device_code_llvm map(y -> 2 * y / 10, CUDA.ones(5_000_000));
; PTX CompilerJob of MethodInstance for (::GPUArrays.var"#gpu_broadcast_kernel_linear#_copyto!##1")(::KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicCheck, Nothing, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, KernelAbstractions.NDIteration.NDRange{1, KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicSize, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}}}, ::CuDeviceVector{Float32, 1}, ::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{1, CUDA.DeviceMemory}, Tuple{Base.OneTo{Int64}}, var"#424#425", Tuple{Base.Broadcast.Extruded{CuDeviceVector{Float32, 1}, Tuple{Bool}, Tuple{Int64}}}}) for sm_86
;  @ none within `gpu_broadcast_kernel_linear`
define ptx_kernel void @_Z27gpu_broadcast_kernel_linear16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILi1E5TupleI5OneToI5Int64EEE7NDRangeILi1ES0_S0_S8_S8_EE13CuDeviceArrayI7Float32Li1ELi1EE11BroadcastedI12CuArrayStyleILi1E12DeviceMemoryES7_4_424S3_I8ExtrudedISE_S3_I4BoolES3_IS5_EEEE({ ptr, i32 } %state, { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", { ptr addrspace(1), i64, [1 x i64], i64 } %"dest::CuDeviceArray", { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted") local_unnamed_addr {
conversion:
%"__ctx__::CompilerMetadata.fca.0.0.0.0.extract" = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", 0, 0, 0, 0
%"__ctx__::CompilerMetadata.fca.1.0.0.0.0.extract" = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", 1, 0, 0, 0, 0
%"__ctx__::CompilerMetadata.fca.1.1.0.0.0.extract" = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", 1, 1, 0, 0, 0
%0 = call i32 @llvm.nvvm.read.ptx.sreg.ctaid.x()
%1 = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
%2 = add nuw nsw i32 %1, 1
%3 = icmp sgt i64 %"__ctx__::CompilerMetadata.fca.1.0.0.0.0.extract", 0
call void @llvm.assume(i1 %3)
%4 = icmp sgt i64 %"__ctx__::CompilerMetadata.fca.1.1.0.0.0.extract", 0
call void @llvm.assume(i1 %4)
%5 = zext nneg i32 %2 to i64
%6 = zext nneg i32 %0 to i64
%7 = mul i64 %"__ctx__::CompilerMetadata.fca.1.1.0.0.0.extract", %6
%8 = add i64 %7, %5
%9 = icmp slt i64 %8, 1
%10 = icmp sgt i64 %8, %"__ctx__::CompilerMetadata.fca.0.0.0.0.extract"
%.not13 = or i1 %9, %10
br i1 %.not13, label %L328, label %L129

L129:                                             ; preds = %conversion
%"dest::CuDeviceArray.fca.0.extract" = extractvalue { ptr addrspace(1), i64, [1 x i64], i64 } %"dest::CuDeviceArray", 0
; ┌ @ /home/efaulha2/.julia/packages/KernelAbstractions/lGrz7/src/macros.jl:324 within `macro expansion`
; │┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:177 within `setindex!`
; ││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:135 within `arrayset`
; │││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:142 within `arrayset_bits`
; ││││┌ @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/pointer.jl:88 within `unsafe_store!`
; │││││┌ @ none within `pointerset`
; ││││││┌ @ none within `macro expansion` @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/base.jl:39
       %11 = getelementptr float, ptr addrspace(1) %"dest::CuDeviceArray.fca.0.extract", i64 %8
       %12 = getelementptr float, ptr addrspace(1) %11, i64 -1
       %"bc::Broadcasted.fca.0.0.0.0.extract" = extractvalue { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted", 0, 0, 0, 0
       %"bc::Broadcasted.fca.0.0.1.0.extract" = extractvalue { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted", 0, 0, 1, 0
; │└└└└└└
; │┌ @ broadcast.jl:616 within `getindex`
; ││┌ @ broadcast.jl:620 within `_getindex`
; │││┌ @ broadcast.jl:671 within `_broadcast_getindex`
; ││││┌ @ broadcast.jl:696 within `_getindex`
; │││││┌ @ broadcast.jl:665 within `_broadcast_getindex`
; ││││││┌ @ broadcast.jl:595 within `newindex`
; │││││││┌ @ essentials.jl:799 within `ifelse`
        %13 = and i8 %"bc::Broadcasted.fca.0.0.1.0.extract", 1
        %.not = icmp eq i8 %13, 0
        %"bc::Broadcasted.fca.0.0.2.0.extract" = extractvalue { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted", 0, 0, 2, 0
        %14 = select i1 %.not, i64 %"bc::Broadcasted.fca.0.0.2.0.extract", i64 %8
; ││││││└└
; ││││││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:175 within `getindex`
; │││││││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:90 within `arrayref`
; ││││││││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:96 within `arrayref_bits`
; │││││││││┌ @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/pointer.jl:85 within `unsafe_load`
; ││││││││││┌ @ none within `pointerref`
; │││││││││││┌ @ none within `macro expansion` @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/base.jl:39
            %15 = getelementptr float, ptr addrspace(1) %"bc::Broadcasted.fca.0.0.0.0.extract", i64 %14
            %16 = getelementptr float, ptr addrspace(1) %15, i64 -1
            %17 = load float, ptr addrspace(1) %16, align 4
; ││││└└└└└└└└
; ││││ @ broadcast.jl:672 within `_broadcast_getindex`
; ││││┌ @ broadcast.jl:699 within `_broadcast_getindex_evalf`
; │││││┌ @ REPL[89]:1 within `#424`
; ││││││┌ @ promotion.jl:434 within `*` @ float.jl:497
       %18 = fmul float %17, 2.000000e+00
; ││││││└
; ││││││┌ @ promotion.jl:436 within `/` @ float.jl:498
       %19 = fdiv float %18, 1.000000e+01
; │└└└└└└
; │┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:177 within `setindex!`
; ││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:135 within `arrayset`
; │││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:142 within `arrayset_bits`
; ││││┌ @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/pointer.jl:88 within `unsafe_store!`
; │││││┌ @ none within `pointerset`
; ││││││┌ @ none within `macro expansion` @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/base.jl:39
       store float %19, ptr addrspace(1) %12, align 4
; │││└└└└
; │││ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:137 within `arrayset`
   br label %L328

L328:                                             ; preds = %L129, %conversion
; │└└
 ret void
; └
}
LLVM for 2 // 10 * y 
julia> CUDA.@device_code_llvm map(y -> 2 // 10 * y, CUDA.ones(5_000_000));
; PTX CompilerJob of MethodInstance for (::GPUArrays.var"#gpu_broadcast_kernel_linear#_copyto!##1")(::KernelAbstractions.CompilerMetadata{KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicCheck, Nothing, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, KernelAbstractions.NDIteration.NDRange{1, KernelAbstractions.NDIteration.DynamicSize, KernelAbstractions.NDIteration.DynamicSize, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}, CartesianIndices{1, Tuple{Base.OneTo{Int64}}}}}, ::CuDeviceVector{Float32, 1}, ::Base.Broadcast.Broadcasted{CUDA.CuArrayStyle{1, CUDA.DeviceMemory}, Tuple{Base.OneTo{Int64}}, var"#427#428", Tuple{Base.Broadcast.Extruded{CuDeviceVector{Float32, 1}, Tuple{Bool}, Tuple{Int64}}}}) for sm_86
;  @ none within `gpu_broadcast_kernel_linear`
define ptx_kernel void @_Z27gpu_broadcast_kernel_linear16CompilerMetadataI11DynamicSize12DynamicCheckv16CartesianIndicesILi1E5TupleI5OneToI5Int64EEE7NDRangeILi1ES0_S0_S8_S8_EE13CuDeviceArrayI7Float32Li1ELi1EE11BroadcastedI12CuArrayStyleILi1E12DeviceMemoryES7_4_427S3_I8ExtrudedISE_S3_I4BoolES3_IS5_EEEE({ ptr, i32 } %state, { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", { ptr addrspace(1), i64, [1 x i64], i64 } %"dest::CuDeviceArray", { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted") local_unnamed_addr {
conversion:
%"__ctx__::CompilerMetadata.fca.0.0.0.0.extract" = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", 0, 0, 0, 0
%"__ctx__::CompilerMetadata.fca.1.0.0.0.0.extract" = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", 1, 0, 0, 0, 0
%"__ctx__::CompilerMetadata.fca.1.1.0.0.0.extract" = extractvalue { [1 x [1 x [1 x i64]]], [2 x [1 x [1 x [1 x i64]]]] } %"__ctx__::CompilerMetadata", 1, 1, 0, 0, 0
%"dest::CuDeviceArray.fca.0.extract" = extractvalue { ptr addrspace(1), i64, [1 x i64], i64 } %"dest::CuDeviceArray", 0
%sret_box = alloca [2 x i64], align 8
%0 = call i32 @llvm.nvvm.read.ptx.sreg.ctaid.x()
%1 = call i32 @llvm.nvvm.read.ptx.sreg.tid.x()
%2 = add nuw nsw i32 %1, 1
%3 = icmp sgt i64 %"__ctx__::CompilerMetadata.fca.1.0.0.0.0.extract", 0
call void @llvm.assume(i1 %3)
%4 = icmp sgt i64 %"__ctx__::CompilerMetadata.fca.1.1.0.0.0.extract", 0
call void @llvm.assume(i1 %4)
%5 = zext nneg i32 %2 to i64
%6 = zext nneg i32 %0 to i64
%7 = mul i64 %"__ctx__::CompilerMetadata.fca.1.1.0.0.0.extract", %6
%8 = add i64 %7, %5
%9 = icmp slt i64 %8, 1
%10 = icmp sgt i64 %8, %"__ctx__::CompilerMetadata.fca.0.0.0.0.extract"
%.not18 = or i1 %9, %10
br i1 %.not18, label %L360, label %L129

L129:                                             ; preds = %conversion
%"bc::Broadcasted.fca.0.0.2.0.extract" = extractvalue { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted", 0, 0, 2, 0
%"bc::Broadcasted.fca.0.0.1.0.extract" = extractvalue { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted", 0, 0, 1, 0
%"bc::Broadcasted.fca.0.0.0.0.extract" = extractvalue { [1 x { { ptr addrspace(1), i64, [1 x i64], i64 }, [1 x i8], [1 x i64] }], [1 x [1 x i64]] } %"bc::Broadcasted", 0, 0, 0, 0
; ┌ @ /home/efaulha2/.julia/packages/KernelAbstractions/lGrz7/src/macros.jl:324 within `macro expansion`
; │┌ @ broadcast.jl:616 within `getindex`
; ││┌ @ broadcast.jl:620 within `_getindex`
; │││┌ @ broadcast.jl:671 within `_broadcast_getindex`
; ││││┌ @ broadcast.jl:696 within `_getindex`
; │││││┌ @ broadcast.jl:665 within `_broadcast_getindex`
; ││││││┌ @ broadcast.jl:595 within `newindex`
; │││││││┌ @ essentials.jl:799 within `ifelse`
        %11 = and i8 %"bc::Broadcasted.fca.0.0.1.0.extract", 1
        %.not = icmp eq i8 %11, 0
        %12 = select i1 %.not, i64 %"bc::Broadcasted.fca.0.0.2.0.extract", i64 %8
; ││││││└└
; ││││││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:175 within `getindex`
; │││││││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:90 within `arrayref`
; ││││││││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:96 within `arrayref_bits`
; │││││││││┌ @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/pointer.jl:85 within `unsafe_load`
; ││││││││││┌ @ none within `pointerref`
; │││││││││││┌ @ none within `macro expansion` @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/base.jl:39
            %13 = getelementptr float, ptr addrspace(1) %"bc::Broadcasted.fca.0.0.0.0.extract", i64 %12
            %14 = getelementptr float, ptr addrspace(1) %13, i64 -1
            %15 = load float, ptr addrspace(1) %14, align 4
; ││││└└└└└└└└
; ││││ @ broadcast.jl:672 within `_broadcast_getindex`
; ││││┌ @ broadcast.jl:699 within `_broadcast_getindex_evalf`
; │││││┌ @ REPL[90]:1 within `#427`
; ││││││┌ @ rational.jl:91 within `//`
; │││││││┌ @ rational.jl:48 within `Rational` @ rational.jl:43
        call fastcc void @julia_divgcd_100531({ ptr, i32 } %state, ptr %sret_box, i64 2, i64 10)
; ││││││││┌ @ tuple.jl:162 within `indexed_iterate`
         %"sret_box[2]_ptr" = getelementptr inbounds i8, ptr %sret_box, i64 8
; ││││││││└
; ││││││││ @ rational.jl:48 within `Rational` @ rational.jl:44
; ││││││││┌ @ rational.jl:21 within `checked_den`
; │││││││││┌ @ int.jl:139 within `signbit`
; ││││││││││┌ @ int.jl:83 within `<`
           %"sret_box[2]_ptr.unbox" = load i64, ptr %"sret_box[2]_ptr", align 8
           %16 = icmp sgt i64 %"sret_box[2]_ptr.unbox", -1
; │││││││││└└
         br i1 %16, label %L308.L330_crit_edge, label %L313

L308.L330_crit_edge:                              ; preds = %L129
         %value_phi.in.sroa.speculate.load.L308.L330_crit_edge = load i64, ptr %sret_box, align 8
         br label %L336

L313:                                             ; preds = %L129
; │││││││││ @ rational.jl:22 within `checked_den`
; │││││││││┌ @ checked.jl:95 within `checked_neg`
; ││││││││││┌ @ checked.jl:231 within `checked_sub`
; │││││││││││┌ @ checked.jl:203 within `sub_with_overflow`
            %17 = call { i64, i1 } @llvm.ssub.with.overflow.i64(i64 0, i64 %"sret_box[2]_ptr.unbox")
            %18 = extractvalue { i64, i1 } %17, 0
            %19 = extractvalue { i64, i1 } %17, 1
; │││││││││││└
; │││││││││││ @ checked.jl:232 within `checked_sub`
           br i1 %19, label %L317, label %L321

L317:                                             ; preds = %L313
           call fastcc void @julia_throw_overflowerr_binaryop_100524({ ptr, i32 } %state)
           unreachable

L321:                                             ; preds = %L313
; │││││││││└└
; │││││││││ @ rational.jl:23 within `checked_den`
; │││││││││┌ @ checked.jl:95 within `checked_neg`
; ││││││││││┌ @ checked.jl:231 within `checked_sub`
; │││││││││││┌ @ checked.jl:203 within `sub_with_overflow`
            %sret_box.unbox = load i64, ptr %sret_box, align 8
            %20 = call { i64, i1 } @llvm.ssub.with.overflow.i64(i64 0, i64 %sret_box.unbox)
            %21 = extractvalue { i64, i1 } %20, 1
; │││││││││││└
; │││││││││││ @ checked.jl:232 within `checked_sub`
           br i1 %21, label %L325, label %L329

L325:                                             ; preds = %L321
           call fastcc void @julia_throw_overflowerr_binaryop_100524({ ptr, i32 } %state)
           unreachable

L329:                                             ; preds = %L321
; │││││││││││ @ checked.jl:231 within `checked_sub`
; │││││││││││┌ @ checked.jl:203 within `sub_with_overflow`
            %22 = extractvalue { i64, i1 } %20, 0
; │││││││││││└
; │││││││││││ @ checked.jl:232 within `checked_sub`
           br label %L336

L336:                                             ; preds = %L329, %L308.L330_crit_edge
           %value_phi.in.sroa.speculated = phi i64 [ %22, %L329 ], [ %value_phi.in.sroa.speculate.load.L308.L330_crit_edge, %L308.L330_crit_edge ]
           %value_phi46 = phi i64 [ %18, %L329 ], [ %"sret_box[2]_ptr.unbox", %L308.L330_crit_edge ]
; ││││││└└└└└
; ││││││┌ @ promotion.jl:434 within `*`
; │││││││┌ @ promotion.jl:404 within `promote`
; ││││││││┌ @ promotion.jl:379 within `_promote`
; │││││││││┌ @ number.jl:7 within `convert`
; ││││││││││┌ @ rational.jl:164 within `AbstractFloat`
; │││││││││││┌ @ number.jl:7 within `convert`
; ││││││││││││┌ @ float.jl:245 within `Float32`
             %23 = sitofp i64 %value_phi.in.sroa.speculated to float
             %24 = sitofp i64 %value_phi46 to float
; │││││││││││└└
; │││││││││││┌ @ float.jl:498 within `/`
            %25 = fdiv float %23, %24
; │││││││└└└└└
; │││││││ @ promotion.jl:434 within `*` @ float.jl:497
       %26 = fmul float %15, %25
; │└└└└└└
; │┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:177 within `setindex!`
; ││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:135 within `arrayset`
; │││┌ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:142 within `arrayset_bits`
; ││││┌ @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/pointer.jl:88 within `unsafe_store!`
; │││││┌ @ none within `pointerset`
; ││││││┌ @ none within `macro expansion` @ /home/efaulha2/.julia/packages/LLVM/iza6e/src/interop/base.jl:39
       %27 = getelementptr float, ptr addrspace(1) %"dest::CuDeviceArray.fca.0.extract", i64 %8
       %28 = getelementptr float, ptr addrspace(1) %27, i64 -1
       store float %26, ptr addrspace(1) %28, align 4
; │││└└└└
; │││ @ /home/efaulha2/.julia/packages/CUDA/UurkZ/src/device/array.jl:137 within `arrayset`
   br label %L360

L360:                                             ; preds = %L336, %conversion
; │└└
 ret void
; └
}

Here is my versioninfo:

julia> CUDA.versioninfo()
CUDA toolchain: 
- runtime 12.5, artifact installation
- driver 555.58.2 for 12.5
- compiler 12.9

CUDA libraries: 
- CUBLAS: 12.5.3
- CURAND: 10.3.6
- CUFFT: 11.2.3
- CUSOLVER: 11.6.3
- CUSPARSE: 12.5.1
- CUPTI: 2024.2.1 (API 12.5.0)
- NVML: 12.0.0+555.58.2

Julia packages: 
- CUDA: 5.9.2
- CUDA_Driver_jll: 13.0.2+0
- CUDA_Compiler_jll: 0.3.0+0
- CUDA_Runtime_jll: 0.19.2+0

Toolchain:
- Julia: 1.12.1
- LLVM: 18.1.7

Preferences:
- CUDA_Runtime_jll.version: 12.5

2 devices:
  0: NVIDIA RTX A4500 (sm_86, 19.375 MiB / 19.990 GiB available)
  1: NVIDIA RTX A4500 (sm_86, 19.593 GiB / 19.990 GiB available)

Metadata

Metadata

Assignees

No one assigned

    Labels

    bugSomething isn't workingperformanceHow fast can we go?

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions