I don't know where they make their money, but if the hope is to compete in compute, I'd suggest that they do not "play the game" at all in GPUs and instead put together something 7nm and TPU-like. If I were them, I'd be working on that right now. 1TFLOPs/W would be a good goal, 100TFLOPs in 100W, using bfloat 16. Sidestep the GPU bullshit entirely, print money in a niche instead.
AMD reckons it can win the high-end GPU battle
71–80 of 160 posts
Re: AMD reckons it can win the high-end GPU battle
#72Earlier quoted context omitted.
Any recent hackintosh build guide recommendations?
Not used it, but this was posted on HN a few days ago https://infinitediaries.net/my-2020-hackintosh-hardware-spec...
I recently built an AMD hackintosh and although I love it — this build guide is exactly what I'd do if I did it again.
Re: AMD reckons it can win the high-end GPU battle
#73AMD cards don't yet have any raytracing support, right? I know it's a young technology, but they shouldn't be dragging their feet on it if they want to make a serious push for the top.
Real time ray tracing hardware is in a funny place right now. Firstly, it's not really as impressive as it sounds. You aren't going to be playing a game where you explore a house of mirrors any time soon. Control is a really good implementation, and it's so good it makes it really clear how quickly it falls over. Better than not ray tracing, but nowhere near non-real time traditional ray tracing. Secondly, at least w…
>Secondly, at least with the current generation of RTX, the performance hit you take is criminal. We're talking 40%.
I never really understood this argument. If you really need tons of frames for a competitive game, just turn it off. Otherwise, the games that implement it correctly show a marked visual improvement, that can't be computed any faster as far as we know. It's like saying 4k renders 4x slower than 1080p, so it's not worth it unless you can get the same frames. You can, it's just a premium technology and you'll need to budget accordingly.
(this ignores the games that half-ass their raytracing and make almost no perceptible improvement, which to be fair is most of them)
I do agree though, that Control (and Metro Exodus) are the best examples. Every other RTX title just uses it for reflections, which are neat but not as much of a categorical upgrade as RTGI, imo.
People who spend serious chunks of time gaming (but who also remember what reality looks like...) will increasingly notice the inaccuracy or lack of ambient occlusion and shading overall.
This time last year, RTX, in popular discussion, was a joke. I'm seeing more and more people who've looked into it, can see the objective improvement, and are ready to adopt it once as it enters their budget range. It would definitely be a sore sport if Navi had even worse raytracing performance than the RTX series, or none at all. As the hardware proliferates, there will be more and more reason to developers to look into implementing it at its best.
Crazy idea: Maybe we could train a neural network on before/after images of raytraced scenes and implement it with a constant, minimal penalty, like DLSS, or maybe that requires too large of a network or maybe it results in a trippy mess.
Re: AMD reckons it can win the high-end GPU battle
#74Earlier quoted context omitted.
Same with a 580. Linux, many monitors, work, light gaming. Drivers are the first time I've ever not had issues with Linux graphics. Three years in I'm still waiting for the other shoe to drop. Where are the driver issues? I'm certain it was like a law of physics in Linux, but somehow, not this time, not for me?
I have 2x 24" monitors on Ubuntu 18.04.03 with a Radeon RX 580 I can't watch a video on Youtube without substantial tearing on Firefox or Chrome. I've followed close to 20 guides and tried a number of drivers which all fail provide any benefit. Running glxgears will run at 7400FPS, with minor tearing. My monitors don't have FreeSync or whatever it's called since they're pretty old (over 10 years). Since most guides a…
Look at your compositor, I've fixed this in KDE by enabling Full Screen Repaints under vsync options.
Re: AMD reckons it can win the high-end GPU battle
#75Earlier quoted context omitted.
Agree, I dropped Linux due bad graphics drivers (AMD)
Are you really saying that AMD is worse than Nvidia on Linux ??? I have a Nvidia GPU (on my laptop), and I've had a truly horrible experience on Linux with Nvidia proprietary drivers. I was actually considering ditching my laptop, and buying a new one with AMD chipsets, due to the issues I've had on Linux due to the presence of the Nvidia graphics card. I occasionally play games on my Linux laptop, using Proton (a St…
He might not be saying that; but I can.
I have a fair number of machines and I /typically/ buy laptop/desktops with intels' graphics which works wonderfully under linux.
On the occasion I use a descreet graphics card (Quatro M1200, GTX 670) it' usually nvidia, and it works pretty good if you have the proprietary driver installed. (even on FreeBSD).
I did, however, recently build a full AMD machine that I intended to use for linux gaming, I did the most "happy path" thing and installed Ubuntu.
It was a Ryzen 1700x with an AMD R9 390 (old at this point, but it was built when Ryzen just came out) and I genuinely couldn't get 3d Accel working at all. Games ran like absolute garbage. I ended up installing all kinds of "amd" "amd-pro" and "radeon" drivers which came with various warnings about being out of date or proprietary.. or "this should be in the kernel as FOSS".. but it never worked.
I ended up selling the machine 2 years later.
Contrastly my laptop has hybrid graphics with Nvidia and is still working fine. :(
Re: AMD reckons it can win the high-end GPU battle
#76Earlier quoted context omitted.
Cuda is a proprietary Nvidia framework. You don't need cuda support to train deep learning models on TensorFlow. ROCm also works - and is built by AMD and is open-source.
Tensorflow is only a small subset of the massive engineering investment that has been done into cuda by various orgs and open source projects. We have GPU databases, GPU accelerated analytics packages, GPU accelerated DSP, etc. All with hand-written cuda kernels. Unless AMD can provide a toolchain which takes cuda code and generates whatever it takes to run it with performance parity to Nvidia cards, it'll never take…
> Unless AMD can provide a toolchain which takes cuda code and generates whatever it takes to run it with performance parity to Nvidia cards, it'll never take off.
They have this. It's called HIP.
Re: AMD reckons it can win the high-end GPU battle
#77Earlier quoted context omitted.
I have 2x 24" monitors on Ubuntu 18.04.03 with a Radeon RX 580 I can't watch a video on Youtube without substantial tearing on Firefox or Chrome. I've followed close to 20 guides and tried a number of drivers which all fail provide any benefit. Running glxgears will run at 7400FPS, with minor tearing. My monitors don't have FreeSync or whatever it's called since they're pretty old (over 10 years). Since most guides a…
My condolences. On Nvidia there is a Full Composition Pipeline option in Nvidia driver settings which gets rid of vsync issues. It's odd it's not on by default.
Wow, that's good to know! How do you turn it on?
(I don't see an option for it in the Nvidia X Server Settings app.)
Re: AMD reckons it can win the high-end GPU battle
#78Earlier quoted context omitted.
Tensorflow is only a small subset of the massive engineering investment that has been done into cuda by various orgs and open source projects. We have GPU databases, GPU accelerated analytics packages, GPU accelerated DSP, etc. All with hand-written cuda kernels. Unless AMD can provide a toolchain which takes cuda code and generates whatever it takes to run it with performance parity to Nvidia cards, it'll never take…
> Unless AMD can provide a toolchain which takes cuda code and generates whatever it takes to run it with performance parity to Nvidia cards, it'll never take off. They have this. It's called HIP.
https://rocm-documentation.readthedocs.io/en/latest/Programm...
"HIP is designed to ease the porting of existing CUDA code into the HIP environment."
Not even close an drop-in swappable alternative.
Re: AMD reckons it can win the high-end GPU battle
#79Earlier quoted context omitted.
Same with a 580. Linux, many monitors, work, light gaming. Drivers are the first time I've ever not had issues with Linux graphics. Three years in I'm still waiting for the other shoe to drop. Where are the driver issues? I'm certain it was like a law of physics in Linux, but somehow, not this time, not for me?
I have 2x 24" monitors on Ubuntu 18.04.03 with a Radeon RX 580 I can't watch a video on Youtube without substantial tearing on Firefox or Chrome. I've followed close to 20 guides and tried a number of drivers which all fail provide any benefit. Running glxgears will run at 7400FPS, with minor tearing. My monitors don't have FreeSync or whatever it's called since they're pretty old (over 10 years). Since most guides a…
Switch to Wayland or just enable TearFree on Xorg, something like:
`xrandr --output DisplayPort-0 --set TearFree on`
of course you have to adjust to your output (`xrandr --props`).
Re: AMD reckons it can win the high-end GPU battle
#80AMD's rather old Radeon VII can really smash a 2080 Ti if you're doing relatively simple streaming computation, like say video processing. That 1TB/s HBM is no joke, and there's 16GB of it, at ~half the price of the 2080 Ti with 11GB. Nvidia's memory compression is very impressive and allows them to do better with a narrower memory bus in many situations (such as games), however if your data is not so compressible (o…
AMD's rather old Radeon VII can really smash a 2080 Ti if you're doing relatively simple streaming computation, like say video processing. That 1TB/s HBM is no joke, and there's 16GB of it, at ~half the price of the 2080 Ti with 11GB.
The Radeon VII is also pretty fast in training transformer networks. In some models that I trained, it was only ~25% slower than our Quadro RTX 5000s.