AMD refuses to set handicap on any workload

[Kisai]

On 3/18/2021 at 12:17 PM, StDragon said:

When it comes certified graphic hardware AMD isn't much better with their Pro and FirePro lineup.

 

https://knowledge.autodesk.com/certified-graphics-hardware

 

On 3/18/2021 at 12:05 PM, Grabhanem said:

For me, the difference is that AMD's architecture isn't maliciously bad at anything. Take CAD for example. Geforce cards have been maliciously limited in CAD via disabled features like antialiased lines for years because Nvidia knows they can make more off people who do real work by selling them Quadros. I'm sure AMD's been doing something similar with Radeon Pro, although much less aggressively.

Interestingly, Nvidia had to "unlock" Quadro features on Titans to compete with the Vega FE's less restrictive drivers, and I suspect the reason they rebranded the Titan as the 3090 was to allow them to go back to more restrictive drivers on their "prosumer" cards without breaking that promise.

You want to know something interesting?

 

Did you know that you can just install the Radeon Pro/FirePro drivers on each other, no messing around at all? I'm not sure how true this remains between generations, but several HP laptops (that have since been disposed of) had the FirePro chip, but every second one had the Radeon driver on them and worked without issue (that I'm aware of.)

 

The same can not be said for nVidia, where it will usually just straight up block you from installing "the wrong driver", despite the exact same chip being used. This also breaks installing Quadro and Geforce on the same computer.

 

At any rate you're likely to hit some bugs if you use the wrong driver because the driver wasn't tested in that configuration. I'm also sure AIB's would love to just sell overpriced Geforce parts as Quadro's during shortages. Imagine if all the 3060ti's were being sold as RTX A1000's

 

[Uttamattamakin]

On 3/19/2021 at 7:50 PM, igormp said:

Ok, looks like you saw way too much marketing stuff.

 

What you're talking about applies to ANY GPU,

 

As I said in the post you quoted  "In fact I am pretty sure if you have a dGPU and use productivity software that can benefit from it it will use that. "

On 3/19/2021 at 7:50 PM, igormp said:

 

They haven't updated their HSA SDK in years and instead are relying on public consortiums to maintain it.

Which is known as being open source you know like Linux, or for that matter Chromium is to Chrome etc. 

 

On 3/19/2021 at 7:50 PM, igormp said:

 

AMD's integrated GPU has better performance than Intel's integrated one, yes, but that's solely due to having more hardware that's also better, and has nothing to do with the whole HSA stuff.

Ahh but it does.   AMD's APU's are built with this in mind.  It does require a measure of hardware support that intel's CPU's with IGP's don't have.  

 

Even if that was somehow not the case a more powerful GPU means better GPU accelerated task.  

 

On 3/19/2021 at 7:50 PM, igormp said:

 

Intel had GVT-g first. AMD's implementation is called MxGPU, and is only available in Radeon FirePro S GPUs (so not even all of their enterprise GPUs support it). So Intel is the only one that offers that feature in consumer hardware without requiring a license.

 

Keep in mind that hardware passthrough has nothing to do with what you're talking about. SR-IOV allows you to "share" a PCIe device between multiple guests. 

 

Feel free to ask any questions about GPUs or AMD's APUs. I actually had a laptop with their first APU arch (llano), which was kinda okaish at the time but nothing really impressive, and I also work with CUDA on a daily basis (mostly due to ML stuff).

Yeah you know I do that too.   I've seen the difference between an AMD APU and a intel CPU with it's integrated graphics all the time.  AMD's offering is vastly superior in any type of GPU accelerated workload.  They are made with this use case in mind.  

AMD makes their APU's with the assumption it will be all that a given comptuer will have so it has to be able to do everything one might want/need a GPU to do.  Not as well as a dGPU.  That's no my claim.    All I've ever said on this, is that at least under current conditions, such products are far more obtainable than the new Nvidia Unobtanium Pro GPU's we see displayed on youtube.  

[igormp]

23 minutes ago, Uttamattamakin said:

As I said in the post you quoted  "In fact I am pretty sure if you have a dGPU and use productivity software that can benefit from it it will use that. "

Yeah, and most can't, specially due to the lack of proper support support from AMD (sadly).

 

23 minutes ago, Uttamattamakin said:

Which is known as being open source you know like Linux, or for that matter Chromium is to Chrome etc. 

Chromium is fully maintained by google though, while AMD's HSA SDK has been abandoned since 2015 (where the last significant update happened): https://github.com/HSAFoundation/HSA-Runtime-AMD

 

26 minutes ago, Uttamattamakin said:

Ahh but it does.   AMD's APU's are built with this in mind.  It does require a measure of hardware support that intel's CPU's with IGP's don't have.  

No, it doesn't, AMD isn't even using its own HSA stack anymore in the desktop space, and the iGPUs in their APUs are regarded as a regular GPU with shared memory by all means.

 

Intel has way better software support than AMD in all fronts, too bad their current graphics offering sucks.

 

28 minutes ago, Uttamattamakin said:

Yeah you know I do that too.   I've seen the difference between an AMD APU and a intel CPU with it's integrated graphics all the time.  AMD's offering is vastly superior in any type of GPU accelerated workload.  They are made with this use case in mind.

So you should know that both are pretty useless for productivity tasks, Intel's QuickSync is way better than AMD's VCE, and your only actual options for any serious GPGPU development is an Nvidia GPU, or cry with a high end AMD GPU + opencl/sycl.

 

30 minutes ago, Uttamattamakin said:

AMD makes their APU's with the assumption it will be all that a given comptuer will have so it has to be able to do everything one might want/need a GPU to do.  Not as well as a dGPU.  That's no my claim.    All I've ever said on this, is that at least under current conditions, such products are far more obtainable than the new Nvidia Unobtanium Pro GPU's we see displayed on youtube.  

And I agree with that, but your first claim is that they had some "special integration" due to the whole HSA stuff, which is simply not true. It's just a really good integrated graphic option with nothing else to add.

[cj09beira]

with cxl and the frontier supercomputer using a single virtual address space, HSA should start to bring actual benefits and be available for consumers around when pcie 5.0 arrives