Could someone please update the guide on Building Blender/Windows - Blender Developer Wiki - or provide a link to the latest relevant documentation - on how to compile HIP (on Windows)? I’m not really sure which packages to install, and if they mess with my nVidia based system.
I’d like to compile with both HIP and OPTIX because we use a slightly modified version of Blender internally in the company, and some computers here are AMD based systems.
We are waiting on AMD to make the HIP SDK more widely available before we can have instructions.
Regarding benchmark results, I think lack of hardware ray-tracing support is what makes the biggest difference. If you look at only CUDA and HIP results are much closer, though there are not that many CUDA results at the moment. In the 3.0 we found speedups on AMD GPUs compared to 2.93, but in some ways the new Cycles architecture can benefit more from hardware ray-tracing than the old one, so because of that the gap became bigger.
Regarding HIP vs. OpenCL, I don’t think the language or API makes a big difference in performance. Optimizations in the compiler or driver are not really specific to either.
For Linux, as far as I know the plan is still to have that this month in Blender 3.2 builds, we’re waiting for a new ROCm driver release.
Thanks for taking the time to respond. Just in case you know the answer, is there any plan for hardware ray tracing?
EDIT
Apologizes to confuse, there was a rearrangement of words from here. The intention was to show there is a long complex journey to hardware ray tracing support as was previously completed for Optix.
What is planned is listed in the task here, and that includes hardware ray-tracing:
https://developer.blender.org/T91571
We don’t have answers to questions about when things will be done until we are close to having something working, just what is planned or not.
(Also for anyone reading @L_S’s comment and confused like me, that’s not a real quote)
I will jump in here. (And yes I am back from parental leave today, I tried to avoid work stuff for a few months). Something like an Optix ray tracing library for HIP is what is needed here to use in Cycles. These types of things don’t happen overnight and it’s not something we are just doing for Blender of course. So there’s no definite date, I would like to give something more concrete, but have been told by our management not to. It is being worked on!
More generally I do think there is some conversations on here that are non-helpful to development or testing, let’s please try to keep it on topic, and understand we won’t reply to these posts.
Linux support - ROCM 5.1 driver needed coming this month. Will then be enabled in Blender 3.2
Hardware RT support - In progress internally at AMD, no date. You’ll probably see the first news about a HIP ray tracing library on https://gpuopen.com
Older hardware support - Still under investigation and lower priority from the AMD side than the above two.
Being realistic, and as a Vega 64 (GCN5) owner, should I even bother continuing to check if it will ever work again with Cycles? “Low priority” is normally just corp speech for “not gonna happen”. I know it’s 5 years old, and I can deal with being slower or missing some features. But being tossed straight out the window? Damn, man. That’s just rough.
As a Linux user, you know that it physically hurts me to say this, but next time I’m going Nvidia. You know, whenever mid-range GPU aren’t going for two grand.
If there will be no HIP support for Vega II I’m also going team green.
Even if AMD pull a performance winner out of the sleeve with the next gen GPUs, how should I trust them if they effectively deprecate their flagship, not-even-2-year-old professional GPU?
Radeon VII PRO was released may 2020.
I have Quadro P1000 from around 2017 and it still works with CUDA today. Hell I wouldn’t be surprised if GTX 980 would work if I bother to dust it off.
Same here, I’m using Vega 64, even though I could barely afford it because I thought it was in the same class as 1080 which was 15% expensive back then. Now I’m unable to render in 3.0, 3.1 or 3.2, because there is no OpenCL or HIP that works with my card. It has reduced my faith in AMD GPU team to nearly nothing. And I have to buy a Nvidia or even a Intel/Arm card next time, instead of AMD, if ever it becomes economically viable for me.
Appreciate your time to respond and how difficult it can be when faced with criticism, however as per the first post “This is a topic for feedback”.
Unfortunately some feedback here is about as negative as it comes but it is still feedback and would be foolish to ignore or suppress. I suggest AMD and Blender take this feedback to heart and hopefully GPU usability can be improved for all.
I also think the negetive feedback here is representative of the broader community, at least what I have seen:
I think AMD just doesn’t care and doesn’t want us to hurt their brand at this point. I’ve come to terms with that I was sold Vega 64 based on lies, and now I am being swiped under the rug because I am too unimportant to deal with anymore, and I think we never were in AMD’s eyes realistically. So yeah, I should’ve listened to all the experts online and bought NVidia instead. Fool me once - shame on you.
Let’s please come back to the topic of giving feedback on the HIP device in Cycles.
I understand that people like support for older cards, but you made your point. @bsavery confirmed in the above post, that support is “Still under investigation”. Once there is news on this, it will be shared here.
Also a Vega 64 + Vega 56 owner here, both of which I bought specifically to use with Blender back in 2017-2018. You removed support for my GPU from your software. I really don’t see how feedback about that would be off-topic in any way.
When only a fraction of existing cards are supported and linux support is nonexistent, you’re not going to get a lot of people posting feedback, let alone positive… The feedback is that nothing works. I feel bad for bsavery for having to deal with this.
I fail to see how responding to
Hey yeah, sorry your cards don’t work anymore (and may have possibly never worked with cycles to begin with), we’re taking a look at it but it’s not a high priority so no promises.
with
Ah, that sucks. As a consumer of AMD’s products, I am dismayed with that statement and will not consider their products in the future, as I do not see GTX 1080 owners lacking support in such a way.
doesn’t qualify as feedback on HIP support in cycles (or lack thereof). People feel as though they have been left high and dry by AMD, and posting in this thread is unfortunately the closest thing many see as lodging a formal complaint with those who are directly connected to the issue. Sending an email to AMD will get an automated response, and posting on any other forum (especially AMD’s own forum) would be equally as effective as shouting about it in the middle of an empty field.
I do understand your perspective, that you’d rather see meaningful discussions from people with newer cards utilizing HIP. But please understand that the perspective of people who are dismayed is not that they “like support for older cards”. It’s that they can’t use your software, and have been told the same message on the matter for the past four months. Or in other cases, have not been able to use their card with cycles for almost 3 years. “Development doesn’t happen overnight” is not something that needs to be stated when it is already very much understood.
With all due respect, asking people to sit around silently and wait for a completely indeterminate amount of time for the possibility of news on the matter is, in my view, unreasonable.
No one is saying you shouldn’t post feedback here. But how constructive is a post saying “Just don’t buy AMD” towards improving the device support? There seems to be a lot of those posts. If the message is dissatisfaction, message received.
Just to make it perfectly clear. If your decision to buy AMD was based on OpenCL support in Blender at the time, that support was removed for Blender 3.0 (not by AMD’s decision). Our solution was to add HIP support to cover as much support for AMD GPUs as possible. No one did anything “wrong” here but assigning that blame for drop of OpenCL squarely on AMD is not fair. If you want to blame AMD for targeting the latest GPUs first, that IS fair.
And trust me, I do feel your pain here.
Not constructive, however most in this thread appear to own AMD GPU’s and want a competitive solution from AMD, which would benefit all Blender users.
There are no bad GPU’s, just bad prices, one should compare the cost of a prospective GPU against the performance in the tasks they wish to use it for.
Old cards aside, as a Blender user who is currently in the market for a new GPU, would you say that purchasing something like an RX6900 today would be a worthwhile investement for the next few years? From what i’ve gathered by reading this thread and other discussions online, HIP is really intended for scientific computations and AI while Blender/Cycles rendering for desktop cards was more of a last minute afterthought in response to OpenCL’s removal. AFAIK even Radeon ProRender doesn’t currently use HIP, and it has HW raytracing already. I’m just having a hard time seeing a long-term strategy here and feel like i’d be betting on the wrong horse again like i did with OpenCL which was “coming eventually” for the past decade but never really went anywhere. Can i expect a real commitment from AMD to make HIP a proper alternative to CUDA/Optix for Blender in the future?
In a word, yes. This is AMD’s compute language alternative to cuda going forward. And not just for HPC applications. This is why it moved from Linux to windows and added support for rdna cards. Also Prorender just added HIP. Primitive variables, flexible ramp node, beta support for HIP in Radeon™ ProRender SDK 2.02.11 - GPUOpen
Expect to see more applications moving from OpenCL to hip this year.
Can someone confirm, Windows 10, driver 22.1.2, 6900xt - Threadripper 1950x
BlenderMark during render consuming only 90% of resources?
edit>> tested on my own scenes and also not going full 100 <<edit
Gpu-Z and Task manager both show arpoximately 90% of usage during render.
And related question, does HIP render also require a powerful CPu to get 100% out of the GPU?