Radeon Open-Source vs. Catalyst With OpenCL CLPEAK

[phoronix]

Phoronix: Radeon Open-Source vs. Catalyst With OpenCL CLPEAK

With yesterday having started to run some fresh basic OpenCL benchmarks on the open-source Radeon driver given the interesting remarks by some super-computing researchers about having more hope for the open-source drivers than the proprietary Catalyst, here are some results comparing the latest open-source AMD Radeon Linux driver code to the proprietary driver.

http://www.phoronix.com/vr.php?view=22784

[Pontostroy]

i added hd 7970 result (my single and double result much higher than even 290x, but bandwidth much slower )

http://openbenchmarking.org/result/1602026-GA-OPENCLAMD14,1602036-HA-HD797015261

[thelongdivider]

Holy crap. Those Russian guys were right to be concerned. AMD should really work on opencl support to compete with Nvidia in more enterprise scenarios.

[boffo]

Is Mesa OpenCL still being developed?

[pete910]

Here's my 290x, just the cat driver though http://openbenchmarking.org/result/1...GA-1602026GA07
Is there normally that much difference between the 290 - 290x in general ?

micheal The new look + updates have certainly improved

Just need to remember to login before I upload/run the benchmarks that I do

[Nille_kungen]

Holy crap. Those Russian guys were right to be concerned. AMD should really work on opencl support to compete with Nvidia in more enterprise scenarios.

I don't see any nvidia numbers here so maybe you can add them if you have them.
As far as i know it was the drivers that wasn't stable that they were concerned about.

[Veerappan]

Is Mesa OpenCL still being developed?

Yes, there's still improvements being made. There's just not that many people working on it at the moment, and most of us only have a small amount of time to dedicate to it.

Also, most of the benchmarks that Michael uses that don't run on clover are benchmarks that use optional features of OpenCL that aren't implemented yet in Mesa without checking for them first.

[s_j_newbury]

I don't see any nvidia numbers here so maybe you can add them if you have them.
As far as i know it was the drivers that wasn't stable that they were concerned about.

I posted my numbers in the other OpenCL comments.

I just re-ran the test with my kernel build from yesterday:

Dell Latitude E6530 (custom build) i7-3840QM

Beignet git master from yesterday, NVIDIA binary driver only used for OpenCL (not currently supported by Clover)

This is a very weak GPU, a fairly low end Fermi with 64bit memory, amusingly enough, in this test it's pretty well matched to the the Intel HD4000 GT2!

Veerappan, do you have any idea why AMD+Clover has such poor latency compared Catalyst and to below? Too many optimizer passes?


Platform: Intel Gen OCL Driver
Device: Intel(R) HD Graphics IvyBridge M GT2
Driver version : 1.2 (Linux x64)
Compute units : 16
Clock frequency : 1000 MHz

Global memory bandwidth (GBPS)
float : 21.18
float2 : 20.97
float4 : 21.16
float8 : 22.22
float16 : 19.52

Single-precision compute (GFLOPS)
float : 217.48
float2 : 216.55
float4 : 212.56
float8 : 210.17
float16 : 226.62

No double precision support! Skipped

Transfer bandwidth (GBPS)
enqueueWriteBuffer : 9.09
enqueueReadBuffer : 6.38
enqueueMapBuffer(for read) : 335544.31
memcpy from mapped ptr : 6.38
enqueueUnmap(after write) : 536870.94
memcpy to mapped ptr : 6.38

Kernel launch latency : 14.22 us


Platform: NVIDIA CUDA
Device: NVS 5200M
Driver version : 361.18 (Linux x64)
Compute units : 2
Clock frequency : 1344 MHz

Global memory bandwidth (GBPS)
float : 21.03
float2 : 21.71
float4 : 22.18
float8 : 19.40
float16 : 11.61

Single-precision compute (GFLOPS)
float : 170.23
float2 : 250.14
float4 : 254.22
float8 : 245.37
float16 : 237.28

Double-precision compute (GFLOPS)
double : 21.44
double2 : 21.42
double4 : 21.38
double8 : 21.29
double16 : 21.13

Integer compute (GIOPS)
int : 85.64
int2 : 85.63
int4 : 85.63
int8 : 85.62
int16 : 85.56

Transfer bandwidth (GBPS)
enqueueWriteBuffer : 5.50
enqueueReadBuffer : 5.48
enqueueMapBuffer(for read) : 5.28
memcpy from mapped ptr : 6.70
enqueueUnmap(after write) : 5.68
memcpy to mapped ptr : 6.77

Kernel launch latency : 5.53 us