Intel UHD GPU: Only reaching half the peak GFlops for GEMM

[EJainDev]

I tuned the GEMM kernel twice for my GPU (one with default and one with fraction set to 100) and I am getting about half the peak performance of my device. About 258 GFlops when the maximum is 530 GFlops. Why is the performance so low?

[CNugteren]

Can you add some more information about the device (e.g. from the CLBlast diagnostics tool)? And are you talking about the performance reported by the tuners or the final performance of the actual GEMM routine?

[EJainDev]

I'm talking about the tuner reported performance of the regular gemm kernel.

This is the information that the tuners store in the json:
"clblast_device_type": "GPU",
"clblast_device_vendor": "Intel",
"clblast_device_architecture": "",
"clblast_device_name": "Intel(R) UHD Graphics",
"device": "Intel(R) UHD Graphics",
"platform_vendor": "Intel(R) Corporation",
"platform_version": "OpenCL 3.0 ",
"device_vendor": "Intel(R) Corporation",
"device_type": "GPU",
"device_core_clock": "1050",
"device_compute_units": "32",
"device_extra_info": "",

And this is all relevant device info from the clinfo binary:

Platform Name Intel(R) OpenCL HD Graphics
Device Name Intel(R) UHD Graphics
Device Vendor Intel(R) Corporation
Device Vendor ID 0x8086
Device Version OpenCL 3.0 NEO
Device Numeric Version 0xc00000 (3.0.0)
Driver Version 31.0.101.2134
Device OpenCL C Version OpenCL C 1.2
Device OpenCL C all versions OpenCL C 0x400000 (1.0.0)
OpenCL C 0x401000 (1.1.0)
OpenCL C 0x402000 (1.2.0)
OpenCL C 0xc00000 (3.0.0)
Device OpenCL C features __opencl_c_int64 0xc00000 (3.0.0)
__opencl_c_3d_image_writes 0xc00000 (3.0.0)
__opencl_c_images 0xc00000 (3.0.0)
__opencl_c_read_write_images 0xc00000 (3.0.0)
__opencl_c_atomic_order_acq_rel 0xc00000 (3.0.0)
__opencl_c_atomic_order_seq_cst 0xc00000 (3.0.0)
__opencl_c_atomic_scope_all_devices 0xc00000 (3.0.0)
__opencl_c_atomic_scope_device 0xc00000 (3.0.0)
__opencl_c_generic_address_space 0xc00000 (3.0.0)
__opencl_c_program_scope_global_variables 0xc00000 (3.0.0)
__opencl_c_work_group_collective_functions 0xc00000 (3.0.0)
__opencl_c_subgroups 0xc00000 (3.0.0)
__opencl_c_pipes 0xc00000 (3.0.0)
Latest conformance test passed v2021-06-16-00
Device Type GPU
Device Profile FULL_PROFILE
Device Available Yes
Compiler Available Yes
Linker Available Yes
Max compute units 32
Max clock frequency 1050MHz
Slices (Intel) 1
Sub-slices per slice (Intel) 4
EUs per sub-slice (Intel) 8
Threads per EU (Intel) 7
Max work item dimensions 3
Max work item sizes 256x256x256
Max work group size 256
Preferred work group size multiple (device) 32
Preferred work group size multiple (kernel) 32
Max sub-groups per work group 32
Sub-group sizes (Intel) 8, 16, 32
Preferred / native vector sizes
char 16 / 16
short 8 / 8
int 4 / 4
long 1 / 1
half 8 / 8 (cl_khr_fp16)
float 1 / 1
double 1 / 1 (n/a)
Single-precision Floating-point support (core)
Denormals Yes
Infinity and NANs Yes
Round to nearest Yes
Round to zero Yes
Round to infinity Yes
IEEE754-2008 fused multiply-add Yes
Support is emulated in software No
Correctly-rounded divide and sqrt operations No
Address bits 64, Little-Endian
Global memory size 5054722048 (4.708GiB)
Error Correction support No
Max memory allocation 2527361024 (2.354GiB)
Minimum alignment for any data type 128 bytes
Alignment of base address 1024 bits (128 bytes)
Max size for global variable 65536 (64KiB)
Preferred total size of global vars 2527361024 (2.354GiB)
Global Memory cache type Read/Write
Global Memory cache size 2359296 (2.25MiB)
Global Memory cache line size 64 bytes
Local memory type Local
Local memory size 65536 (64KiB)
Max number of constant args 8
Max constant buffer size 2527361024 (2.354GiB)
Generic address space support Yes
Max size of kernel argument 2048 (2KiB)
Queue properties (on host)
Out-of-order execution Yes
Profiling Yes
Execution capabilities
Run OpenCL kernels Yes
Run native kernels No
Non-uniform work-groups Yes
Work-group collective functions Yes
Sub-group independent forward progress Yes