I recently bought new 2014 rMBP15 with 2.8gHz and GT750m upgrades. With amazing deals on those laptops I am debating the need for GT750M as BH sells 2.2 IrisPro for $1700 which is like $1000 less then my configuration. And to consider whether to keep it or not I put it through some real life tests. It lead me to some interesting findings of components importance for overall performance of the system. And give something back to community where I get so many information from.
Test rats:
2009 i970@3.84GHz OC'd, RAM DDR3 24GB@1600MHz OC'd, Radeon 6870 1GB, Samsung 840 SSD
2014 rMBP15 2.8Ghz, RAM DDR3 16GB@1600MHz, 512GB PCI-E SSD using:
- IrisPro only
- GT750M only
Test task:
- merge 9 TIFF images @ 128MB each into 360 panorama in AutoPano Giga 3 and render as full sphere 360x180 panorama
- content-aware fill of top black part of panorama of 630MB in size in PS6 (I shoot 360 in hand without nadir and bottom)
And here is the interesting piece:
Render1 is actual rendering time, Render2 includes 'writing file' time which I think corresponds with Geekbench performance numbers.
Scratching my head seeing the times I put the machines to syntetic tests of Cinebench R15 and Geekbench 2.1
Cinebench OpenGL test:
Hack/AMD - 72
rMBP/GT750m - 58
rMBP/IrisPro - 36
Geekbench 2.1 - overall/integer/floating/memory/mem bandwidth
Hack - 11800/9646/19030/6062/5511
rMBP - 15172/11900/23409/8845/10452
OpenCL test Luxmark 2.1
Radeon - 620
IrisPro - ~600
GT750M - ~140
(I am pulling Iris and GT numbers from my memory)
Based on synthetic benchmark my 5 year old hack should have been destroyed in this test - yet the real life performance of procedures which matter in my workflow shows the opposite. And the point of having dGPU in MBP when there is IrisPro - no idea. At least not when I consider productivity apps in OSX environment.
Off course, Hack's memory is slow and it shows in the writing portion of render test. But still I am quite surprised to see that the actual render time is so much faster on old desktop class GPU. Or is it because the Hack's cores are running @3.84 at full throttle while rMBP cores might go at 2.8 at best? And that Photoshop time wasn't that much different. No matter what GPU was used. Interesting.
Cheers
Test rats:
2009 i970@3.84GHz OC'd, RAM DDR3 24GB@1600MHz OC'd, Radeon 6870 1GB, Samsung 840 SSD
2014 rMBP15 2.8Ghz, RAM DDR3 16GB@1600MHz, 512GB PCI-E SSD using:
- IrisPro only
- GT750M only
Test task:
- merge 9 TIFF images @ 128MB each into 360 panorama in AutoPano Giga 3 and render as full sphere 360x180 panorama
- content-aware fill of top black part of panorama of 630MB in size in PS6 (I shoot 360 in hand without nadir and bottom)
And here is the interesting piece:
Code:
hack dgpu iris pro
render1 1:05 1:29 1:24
render2 1:55 2:18 2:15
PS6fill 1:09 58 58Render1 is actual rendering time, Render2 includes 'writing file' time which I think corresponds with Geekbench performance numbers.
Scratching my head seeing the times I put the machines to syntetic tests of Cinebench R15 and Geekbench 2.1
Cinebench OpenGL test:
Hack/AMD - 72
rMBP/GT750m - 58
rMBP/IrisPro - 36
Geekbench 2.1 - overall/integer/floating/memory/mem bandwidth
Hack - 11800/9646/19030/6062/5511
rMBP - 15172/11900/23409/8845/10452
OpenCL test Luxmark 2.1
Radeon - 620
IrisPro - ~600
GT750M - ~140
(I am pulling Iris and GT numbers from my memory)
Based on synthetic benchmark my 5 year old hack should have been destroyed in this test - yet the real life performance of procedures which matter in my workflow shows the opposite. And the point of having dGPU in MBP when there is IrisPro - no idea. At least not when I consider productivity apps in OSX environment.
Off course, Hack's memory is slow and it shows in the writing portion of render test. But still I am quite surprised to see that the actual render time is so much faster on old desktop class GPU. Or is it because the Hack's cores are running @3.84 at full throttle while rMBP cores might go at 2.8 at best? And that Photoshop time wasn't that much different. No matter what GPU was used. Interesting.
Cheers
