AMD Details Radeon Pro Graphics in New 15-inch MacBook Pro Models

[tipoo]

I'd expect the opposite From what I've seen, AMD hardware is better at utilising its resources. Where AMD lacks is drivers — and thats IMO why we are seeing large improvements when Vulcan or DX12 is being used on AMD. The simpler, closer to the hardware model API partially removes the advantage Nvidia has with its driver optimisations.

I think the finer grained level of hardware asynchronous compute in AMD will prove an advantage as DX12 and Vulkan come along, but I'm talking more generally than that, AMD has been using smaller cores in higher numbers, Nvidia fewer in larger cores, which leads to higher Gflops numbers on AMD for generally equivalent cards, discounting big future advantages like Async.

I.e by the first guys example, Fury X had 4096 shaders, GM200 having 3072.

If Fury X bridges the gap in DX12, that's still in line with what I'm saying
33% more shaders = roughly 33% more paper Gflops with whatever clock differences, and lets be charitable to the Fury X and assume it's 100% equivalent performance in DX12. Presto, higher Gflops per unit performance for AMD, aka higher performance per Gflop for Nvidia.

It's not at all about which one is better at using resources, just how they set up their cores = what Gflops will look like on paper, regardless if I use a billion cores to achieve performance of 10, or 1 core to achieve performance of 10.

 

[koyoot]

I think the finer grained level of hardware asynchronous compute in AMD will prove an advantage as DX12 and Vulkan come along, but I'm talking more generally than that, AMD has been using smaller cores in higher numbers, Nvidia fewer in larger cores, which leads to higher Gflops numbers on AMD for generally equivalent cards, discounting big future advantages like Async.

I.e by the first guys example, Fury X had 4096 shaders, GM200 having 3072.

If Fury X bridges the gap in DX12, that's still in line with what I'm saying
33% more shaders = roughly 33% more paper Gflops with whatever clock differences, and lets be charitable to the Fury X and assume it's 100% equivalent performance in DX12. Presto, higher Gflops per unit performance for AMD, aka higher performance per Gflop for Nvidia.

It's not at all about which one is better at using resources, just how they set up their cores = what Gflops will look like on paper, regardless if I use a billion cores to achieve performance of 10, or 1 core to achieve performance of 10.

You are talking about games or compute?

Only advantage Nvidia in compute has is CUDA software. When you take out CUDA, and let AMD compete with Nvidia on brand agnostic software, AMD wins by few percent.

 

[tipoo]

You are talking about games or compute?

Only advantage Nvidia in compute has is CUDA software. When you take out CUDA, and let AMD compete with Nvidia on brand agnostic software, AMD wins by few percent.

Yeah, Nvidia doesn't give a **** about OpenCL driver optimization in favour of CUDA, which I bet is a good half of the reason Apple chose AMD. AMD cards being more optimized for OpenCL = FCPX runs faster than Premier on them.

 

[koyoot]

Yeah, Nvidia doesn't give a **** about OpenCL driver optimization in favour of CUDA, which I bet is a good half of the reason Apple chose AMD. AMD cards being more optimized for OpenCL = FCPX runs faster than Premier on them.

CUDA is Nvidia's proprietary compute API. It is locked ONLY to Nvidia GPUs. It was first compute API that gave so much optimization, thats why it got so much traction.

AMD has its own compute API, based on Mantle. Its called Radeon ProRender. It is a plug-in API for applications that can use it. And brings similar level of optimization for AMD GPUs.

What is more, it is not locked to only AMD GPUs. It is completely Open Source. Apple will not give a **** about it however. They have Metal. Which brings similar optimization of software for Apple hardware. Regardless of brand of GPUs that are used in the Apple computers.

 

[tipoo]

CUDA is Nvidia's proprietary compute API. It is locked ONLY to Nvidia GPUs. It was first compute API that gave so much optimization, thats why it got so much traction.

AMD has its own compute API, based on Mantle. Its called Radeon ProRender. It is a plug-in API for applications that can use it. And brings similar level of optimization for AMD GPUs.

What is more, it is not locked to only AMD GPUs. It is completely Open Source. Apple will not give a **** about it however. They have Metal. Which brings similar optimization of software for Apple hardware. Regardless of brand of GPUs that are used in the Apple computers.

I'm aware, at the same time despite Prorender AMD also does more OpenCL optimization than Nvidia, I don't think Nvidia even supports OpenCL 2.0 yet. Since Apple chose OpenCL to accelerate some of their applications and even parts of macOS, Nvidia somewhat forced their choice until they give a hoot about also optimizing their OpenCL driver.


From what I've seen, despite Metals lower level nature, right now it nets you slightly slower renders than OpenCL on AMD Macs. So Apple still very much cares about OpenCL speed despite Metal. Probably better than OpenCL on Nvidia Macs, haven't dug too deep there.

 

[koyoot]

I'm aware, at the same time despite Prorender AMD also does more OpenCL optimization than Nvidia, I don't think Nvidia even supports OpenCL 2.0 yet. Since Apple chose OpenCL to accelerate some of their applications and even parts of macOS, Nvidia somewhat forced their choice until they give a hoot about also optimizing their OpenCL driver.


From what I've seen, despite Metals lower level nature, right now it nets you slightly slower renders than OpenCL on AMD Macs. So Apple still very much cares about OpenCL speed despite Metal. Probably better than OpenCL on Nvidia Macs, haven't dug too deep there.

I have one word for Metal in its current form: it's rubbish. End of the story. I jumped ship from Macs to Windows some time ago, despite what people been saying and I see gigantic difference in simple application performance. Apple platform in general looks like "work in progress". Everything is started, nothing is finished, polished.

 

[tipoo]

I have one word for Metal in its current form: it's rubbish. End of the story. I jumped ship from Macs to Windows some time ago, despite what people been saying and I see gigantic difference in simple application performance. Apple platform in general looks like "work in progress". Everything is started, nothing is finished, polished.

No doubt. On Boot Camp on the same systems I see a 30-50% speedup on my native apps on both GPU compute and graphics. Native, not performance sapping wrappers. And DX12/Vulkan/Metal are too early right now.

Makes me a bit sad. I really like macOS, there's nowhere else to get a well supported FreeBSD based workstation really, but not only does their hardware ship with ho-hum GPUs, but the OS I love itself saps another huge chunk of the performance on top of that. Metal is one thing, but Apple really needs to update its OpenGL and OpenCL versions and drivers, and add Vulkan for good measure.