OpenCL ROCm 2.0 vs. AMDGPU-PRO Linux
Radeon RX Vega 64 ROCm 2.0 OpenCL versus PAL OpenCL driver in AMDGPU-PRO 18.50. Benchmarks by Michael Larabel for a future article on Phoronix.com.


ROCm 2.0: 

	Processor: AMD Ryzen Threadripper 2990WX 32-Core @ 3.00GHz (32 Cores / 64 Threads), Motherboard: ASUS ROG ZENITH EXTREME (1601 BIOS), Chipset: AMD Family 17h, Memory: 32768MB, Disk: 16GB Voyager 3.0 + Samsung SSD 970 EVO 500GB, Graphics: AMD Radeon RX Vega 8GB (1630/945MHz), Audio: Realtek ALC1220, Monitor: ASUS VP28U, Network: Intel I211 + Qualcomm Atheros QCA6174 802.11ac + Wilocity Wil6200 802.11ad

	OS: Ubuntu 18.04, Kernel: 4.15.0-43-generic (x86_64), Desktop: GNOME Shell 3.28.3, Display Server: X Server 1.19.6, Display Driver: amdgpu 18.0.1, OpenGL: 4.5 Mesa 18.0.5 (LLVM 6.0.0), Compiler: GCC 7.3.0, File-System: ext4, Screen Resolution: 3840x2160

AMDGPU-PRO 18.50 PAL: 

	Processor: AMD Ryzen Threadripper 2990WX 32-Core @ 3.00GHz (32 Cores / 64 Threads), Motherboard: ASUS ROG ZENITH EXTREME (1601 BIOS), Chipset: AMD Family 17h, Memory: 32768MB, Disk: 16GB Voyager 3.0 + Samsung SSD 970 EVO 500GB, Graphics: AMD Radeon RX Vega 8GB (1630/945MHz), Audio: Realtek ALC1220, Monitor: ASUS VP28U, Network: Intel I211 + Qualcomm Atheros QCA6174 802.11ac + Wilocity Wil6200 802.11ad

	OS: Ubuntu 18.04, Kernel: 4.15.0-43-generic (x86_64), Desktop: GNOME Shell 3.28.3, Display Server: X Server 1.19.6, Display Driver: amdgpu 18.1.99, OpenGL: 4.6.13542, Compiler: GCC 7.3.0, File-System: ext4, Screen Resolution: 3840x2160


PlaidML
FP16: No - Mode: Inference - Network: IMDB LSTM - Device: OpenCL
Examples Per Second > Higher Is Better
ROCm 2.0 ............. 252 |=================================================
AMDGPU-PRO 18.50 PAL . 263 |===================================================


PlaidML
FP16: No - Mode: Inference - Network: Mobilenet - Device: OpenCL
Examples Per Second > Higher Is Better
ROCm 2.0 ............. 479 |===================================================
AMDGPU-PRO 18.50 PAL . 479 |===================================================


PlaidML
FP16: No - Mode: Inference - Network: ResNet 50 - Device: OpenCL
Examples Per Second > Higher Is Better
ROCm 2.0 ............. 225 |=================================================
AMDGPU-PRO 18.50 PAL . 233 |===================================================


PlaidML
FP16: No - Mode: Inference - Network: Inception V3 - Device: OpenCL
Examples Per Second > Higher Is Better
ROCm 2.0 ............. 113.88 |========================================
AMDGPU-PRO 18.50 PAL . 136.38 |================================================


SHOC Scalable HeterOgeneous Computing 2015-11-10
Target: OpenCL - Benchmark: Triad
GB/s > Higher Is Better
ROCm 2.0 ............. 6.69 |==================================================
AMDGPU-PRO 18.50 PAL . 6.07 |=============================================


SHOC Scalable HeterOgeneous Computing 2015-11-10
Target: OpenCL - Benchmark: Bus Speed Download
GB/s > Higher Is Better
ROCm 2.0 ............. 7.14 |==================================================
AMDGPU-PRO 18.50 PAL . 7.14 |==================================================


SHOC Scalable HeterOgeneous Computing 2015-11-10
Target: OpenCL - Benchmark: Bus Speed Readback
GB/s > Higher Is Better
ROCm 2.0 ............. 7.16 |==================================================
AMDGPU-PRO 18.50 PAL . 7.16 |==================================================


SHOC Scalable HeterOgeneous Computing 2015-11-10
Target: OpenCL - Benchmark: Texture Read Bandwidth
GB/s > Higher Is Better
ROCm 2.0 ............. 441 |===================================================
AMDGPU-PRO 18.50 PAL . 424 |=================================================


cl-mem 2017-01-13
Benchmark: Copy
GB/s > Higher Is Better
ROCm 2.0 ............. 221 |===============================
AMDGPU-PRO 18.50 PAL . 364 |===================================================


cl-mem 2017-01-13
Benchmark: Read
GB/s > Higher Is Better
ROCm 2.0 ............. 160 |=====================
AMDGPU-PRO 18.50 PAL . 398 |===================================================


cl-mem 2017-01-13
Benchmark: Write
GB/s > Higher Is Better
ROCm 2.0 ............. 384 |===================================================
AMDGPU-PRO 18.50 PAL . 379 |==================================================


clpeak
OpenCL Test: Global Memory Bandwidth
GBPS > Higher Is Better
ROCm 2.0 ............. 362 |===================================================
AMDGPU-PRO 18.50 PAL . 361 |===================================================


clpeak
OpenCL Test: Transfer Bandwidth enqueueReadBuffer
GBPS > Higher Is Better
ROCm 2.0 ............. 17.06 |=================================================
AMDGPU-PRO 18.50 PAL . 10.92 |===============================


clpeak
OpenCL Test: Transfer Bandwidth enqueueWriteBuffer
GBPS > Higher Is Better
ROCm 2.0 ............. 45.44 |=================================================
AMDGPU-PRO 18.50 PAL . 22.49 |========================


SHOC Scalable HeterOgeneous Computing 2015-11-10
Target: OpenCL - Benchmark: FFT SP
GFLOPS > Higher Is Better
ROCm 2.0 ............. 1075 |==================================================
AMDGPU-PRO 18.50 PAL . 863  |========================================


clpeak
OpenCL Test: Single-Precision Float
GFLOPS > Higher Is Better
ROCm 2.0 ............. 13053 |=================================================
AMDGPU-PRO 18.50 PAL . 12528 |===============================================


clpeak
OpenCL Test: Double-Precision Double
GFLOPS > Higher Is Better
ROCm 2.0 ............. 833 |===================================================
AMDGPU-PRO 18.50 PAL . 828 |===================================================


SHOC Scalable HeterOgeneous Computing 2015-11-10
Target: OpenCL - Benchmark: MD5 Hash
GHash/s > Higher Is Better
ROCm 2.0 ............. 16.46 |==============================================
AMDGPU-PRO 18.50 PAL . 17.63 |=================================================


clpeak
OpenCL Test: Integer Compute INT
GIOPS > Higher Is Better
ROCm 2.0 ............. 2497 |==================================================
AMDGPU-PRO 18.50 PAL . 2486 |==================================================


LeelaChessZero 0.20.1
Backend: OpenCL
Nodes Per Second > Higher Is Better
ROCm 2.0 . 304 |===============================================================


JuliaGPU 1.2pts1
OpenCL Device: GPU
Samples/sec > Higher Is Better
ROCm 2.0 ............. 166858816 |===============================
AMDGPU-PRO 18.50 PAL . 240293469 |=============================================


Rodinia 2.4
Test: OpenCL Heartwall
Seconds < Lower Is Better
ROCm 2.0 ............. 4.03 |==================================================
AMDGPU-PRO 18.50 PAL . 3.80 |===============================================


Darktable 2.4.2
Test: Boat - Acceleration: OpenCL
Seconds < Lower Is Better
ROCm 2.0 ............. 4.48 |==========================
AMDGPU-PRO 18.50 PAL . 8.50 |==================================================


Darktable 2.4.2
Test: Masskrug - Acceleration: OpenCL
Seconds < Lower Is Better
ROCm 2.0 ............. 5.70 |==================================================
AMDGPU-PRO 18.50 PAL . 5.47 |================================================


Darktable 2.4.2
Test: Server Rack - Acceleration: OpenCL
Seconds < Lower Is Better
ROCm 2.0 ............. 0.23 |==================================================
AMDGPU-PRO 18.50 PAL . 0.13 |============================


Darktable 2.4.2
Test: Server Room - Acceleration: OpenCL
Seconds < Lower Is Better
ROCm 2.0 ............. 1.97 |======================================
AMDGPU-PRO 18.50 PAL . 2.57 |==================================================


clpeak
OpenCL Test: Kernel Latency
us < Lower Is Better
ROCm 2.0 ............. 10.56 |============
AMDGPU-PRO 18.50 PAL . 41.75 |=================================================