Chips are in the pipeline for many years, so products being released today are probably on the tail end of being impacted by AMD cutting GPU budgets to the bone to focus on Zen. I wouldn't expect miracles.

Given AMD's market cap and financial success, I bet they're working on competitive designs now. There's no money in being an also ran in the space.

They would have been putting tons of development money into the graphics for the PS5 and the XBOX, and I think a lot of what went in to that is going to show up in their new GPUS. The timing of this stuff coming out right alongside the next generation consoles is confirmation of that, I think.

Agreed. Mark Cerny has said that some features in RDNA2 are a direct result of collaboration between Sony/AMD.

Intel is vulnerable. Now is the time for AMD to strike!

The information available around Big Navi (The GPUs expected to launch next month) put them at a +50% performance jump from the last generation and likely able to compete with the Nvidia flagship 3080: https://www.tomshardware.com/news/amd-big_navi-rdna2-all-we-...

I'd really urge anyone wanting to opine on these issues to watch the official analyst day videos and read through the releases AMD has put out so far in terms of architecture and other tech decisions they've made: https://www.youtube.com/watch?v=LMNwLVJQzGs&t=10064s

> The information available around Big Navi (The GPUs expected to launch next month) put them at a +50% performance jump from the last generation and likely able to compete with the Nvidia flagship 3080

I guess it depends on what benchmark you care about but for AMD to compete with a 3080, which is 70-100% faster than a 2080, which is 10-20% faster than AMD's current flagship, they need to do a lot better than +50%.

Some of the leaked benchmarks for the 3080 put it more in the range of 30-50% better than the 2080 series:

"30% Faster Than The RTX 2080 Ti & 50% Faster Than the RTX 2080 SUPER on Average"

https://wccftech.com/nvidia-geforce-rtx-3080-graphics-card-b...

AND those early benchmarks were from Nvidia on very specific DLSS and RTX workloads that may not have been representative to many games people play.

Does it really matter for gaming?

Disclaimer: I'm personally hoping to complete my current build as a "Red Box" (AMD CPU, AMD GPU). However, facts are facts.

DLSS 1.9 was impressive, but has its problems and would understandably be avoided by most gamers. DLSS 2 is a whole different story[1], and is highly likely to be something that most gamers enable. 4K/60 or 1440p/240 are the golden standards for PC gamers (depending on whether you prefer high resolution, or high framerates) and DLSS 2 would help you achieve either.

RTX is nice, Minecraft RTX has done a lot for marketing it. Probably still a coin-flip on whether a gamer would eat the performance hit for the quality improvement.

[1]: https://youtu.be/YWIKzRhYZm4

> 4K/60 or 1440p/240 are the golden standards

Those are $1200+ monitors at any decent resolution.

EDIT: by resolution I mean the horizontal dimension of the monitor, where pc enthusiasts typically go for 27+ inches in width which are more expensive as you scale in size, add a curve, etc.

And you're talking about $600-$1200 GPUs, plus similar or higher for CPU+MB+RAM. People who buy these things are willing to pay for maximum performance.

I bought this 4k/60 monitor for $500 3 years ago:

https://www.amazon.com/gp/product/B00SHZSXVI/ref=asin_title?...

No gsync/freesync, but I wasn't intending to game on it anyway.

You can get a real good quality 4k60 for $400, or one with shitty color for much less. The LG 27gn950 is 4k144 + full DCI-P3 for $800, but unfortunately has been sold out everywhere since it was released a week or two ago.

the resolution is specified, do you mean image quality? gamers tend to sacrifice that for refresh speed and low input latency.

There appears to be no FPS performance gain going from 144hz to 240hz though so you don't have to get crazy.

I meant the dimensions of the monitor i.e. 27in, 32in, etc. 27 inches seems to be one of the most popular dimensions.

DLSS and RTX are still game specific, with very few games that have developed for them. It has been a generation, if there was going to be a big flood of DLSS and RTX titles, it would have happened already.

> DLSS and RTX are still game specific

Note that DLSS 2.0's model is not game specific ( https://www.nvidia.com/en-us/geforce/news/nvidia-dlss-2-0-a-... )

DLSS support is still game specific as it needs additional data beyond just the color buffer to do the up-scaling, such as motion vectors, but it at least doesn't need the expensive part that put it out of reach of eg. indie games.

So it's now on the "if unity & unreal support it, you'll see everyone have it in a few years" trajectory. Probably, anyway, similar to things like TXAA & FXAA.

RT is here to stay, both next gen consoles have it.

Depends on price. Nvidia might have owned the tippity top with their 20xx series but the 5700xt was a very strong card for medium priced builds and as a pc building enthusiast would’ve been the card for my next build alongside an amdR5 3600 cpu. At launch it had driver issues but they were cleaned up and at $350 it was a total steal compared to a 2070 at $650 for similar performance.

I agree. They don't need to beat the 3080, they just need to match or beat the 3070 at a competitive price. I would say the majority of the market does not have $700 to blow on a GPU. Having compelling options under $500 would be enough of a win.

That's what AMD has been doing and it hasn't really worked out. While all the market sales are at the mid range price point, all the youtubers & lust worthy builds that drive mind-share are not. And marketing works, it's why you'll see basically nothing but Nvidia on the steam stats page despite how competitive the 5700 (xt) was.

https://store.steampowered.com/hwsurvey/Steam-Hardware-Softw...

The chart toppers are the mid range cards (1060, 1050 Ti, 1050), exactly as you'd expect. But what you then might not expect is that the 1070 and 1080 are both still higher than any AMD offering. There's almost as many 2070 SUPER's ($600) as there are RX 580s ($230 @ launch, currently sub-$200), AMD's most popular GPU among Steam users.

Ah, apologies the linked stories and videos go into more detail. The “50 percent jump” is from Instructions per cycle alone with process node improvements and better cache architecture also contributing gains.

Final perf numbers are being speculated as being “close to 3080” in the chip forums, but I’d take that with a grain of salt since those are rumors alone.

Isn't the Nvidia flagship the RTX 3090? The overclockers uk website has the most expensive 3090 (Asus one) at ~1600gbp. On paper this should destroy everything else available atm.

In the kitchen keynote, their CEO positioned the 3080 as the flagship card and the 3090 as the Titan replacement(higher vram for things like large model runs)

It's also well over $1000. I imagine that's not the competition that AMD is going after this iteration. They probably want to compete with the graphics card people will actually buy.

But nVidia was substantially ahead the last generation and are promising a greater performance delta than that.

AMD often has very high FLOPS but fails to translate that in to FPS in games, due to architecture inefficiency, drivers, or developers simply optimizing primarily for GeForce.

I would in turn urge you to take all of these figures with a large grain of salt until the third party benchmarks hit.

AMD’s one saving grace is that to optimize for consoles is to optimize for Radeon.

It is important to note that it is 50% performance improvement per watt, not absolute performance.

In at least datacenter settings, performance-per-watt is just as important as performance-per-dollar, since watts are continuous OpEx.

Even in the home, I like to consider performance-per-watt. More power draw means more heat and more heat means less comfort for much of the year.

I think you're overestimating the thermal pollution. A human is 100W-120W. The Zen 3 is about 180-220W. That pales to comparison to the cost of cooling a home where air conditioning systems have 36,000BTU or 10,000W.

Unless your AC is designed specifically with your PC in mind, and is adequately sized for your house (not a guarantee), a hot, noisy PC or laptop still makes its local area uncomfortable.

Most homes with central AC have adjustable vents, many have baffles or zones. It's trivial to tweak your airflow to boost cooling to the room with a gaming pc.

Ceiling fans are another great option for homogenizing the temperature of a room.

You don't need to build your climate control systems specifically with a PC in mind, they're designed to be adjusted.

Honestly I've never lived in a house or apartment that gave enough hot or cold air to the right rooms. Even with baffles open or closed, there are always super hot rooms and super cold rooms. This is exacerbated by having a high power PC (or, before 2020, a party with a lot of people).

Something running at 300W continuously (not that most home PC users would be in that position exactly) still uses about 220 kWh a month.

+50% performance per watt efficiency. Unfortunately that only puts them at the same efficiency ballpark as nvidia's past two generations.

> There's no money in being an also ran in the space.

AMD has spent nearly its entirely life in second place as the resident "also ran". But they're still around...

What's their CUDA competitor this year? Are they pushing out client libs for Vulkan or are we sticking with trying to make ROCm work?

ROCm and HIP - https://github.com/ROCm-Developer-Tools/HIP, as it has been for the last 3-4 years

unless you have navi, in which case rocm is not supported and doesn't work and I guess AMD expects you to use opencl.

Navi just began to compile a week or two ago. Still 'unsupported', but its slowly gaining features.

ROCm is a bit uneven with what cards it supports. Ex: Rx 550 never was supported (even though Polaris, the rest of the 5xx series worked). It seems like the only cards ROCm works for are the ones that share a chip and/or driver with AMD's "Machine Intelligence" line of cards. (MI50, MI60, etc. etc.). Which are Fiji (Rx Fury), Polaris (Rx 580), Vega, and Radeon VII.

CUDA is a "works on all NVidia devices" kind of thing. So I think people are surprised when they have to read the docs for AMD's ROCm.

> Navi just began to compile a week or two ago. Still 'unsupported', but its slowly gaining features.

14 months after release, still unsupported. AMD needs to step up to the plate if they want to be seriously considered in the AI world.

Its pretty clear what's going on: AMD doesn't have the resources to support all the GPUs they release on ROCm. They focus on the "MI" line of cards, which have similarities to the consumer GPUs. (Ex: Rx Fury is very similar to MI8).

Because MI8 and Rx Fury have similarities, AMD gives ROCm support to both. Ditto with MI25 and Vega, and MI6 and Rx 580.

EDIT: With "Navi" gaining features in ROCm repos, my bet is that a new MI-card based on RDNA2 might be coming out soon. Maybe its Arcturus (the Exascale GPU for the Frontier supercomputer).

No, Arcturus is CDNA (the Compute rather than Render architecture). Still with the Radeon VII EOL it would be good if they had a consumerish ROCm card, so RDNA2-support would be great.

Last time I tried, PCIe 3.0 was also a requirement, although that wasn't much of an issue unless it was some nonstandard configuration like an RX 580 in a junkyard find A4 APU machine.

The industry at large is moving to SYCL, which can be implemented on top of OpenCL / CPUs / FPGAs.

I'm not sure what part of 'the industry' you're referring to, but last time I looked (which was admittedly more than a year ago), there is only one serious SYCL implementation, it's proprietary and the only people pushing SYCL are... drum roll... the people selling that implementation. That's not even considering that OpenCL is in most cases (way) behind CUDA in performance, not to mention that the tooling for OpenCL is outright primitive. And all of that is not even mentioning that nvidia drags their feet massively in supporting newer OpenCL version on their GPU's.

Maybe things have changed massively very recently, but I came back from iwocl (the OpenCL and SYCL conference) two years ago massively disappointed. And I still don't see anyone in the slice of the HPC space I'm familiar with (environmental & GIS-related modeling) using it, for a reason I presume.

Intel are supporting SYCL in a big way. They are implementing SYCL in the open source LLVM project, the project is called DPC++. Argonne National Labs are also using SYCL to program the exascale supercomputer (Aurora) they are building. Whilst it's not the best solution for everyone there is certainly industry support for it.

https://software.intel.com/content/www/us/en/develop/tools/o... https://github.com/intel/llvm/blob/sycl/sycl/doc/GetStartedG... https://www.alcf.anl.gov/support-center/training-assets/road...

Oh, the third option, with absolutely no library support. Or, looking at Khronos's page, even device support. At least OpenCL has the benefit of wide compatibility, even if its performance meant it was unsuitable.

Nvidia has settled on CUDA, and has maintained 13 years of compatibility. If SYCL is going to ever compete with CUDA, it's going to need to stick around long enough for people to build on it.

My firm does GPGPU programming for audio processing. Nobody in this area even thinks of using anything other than CUDA. It has far more support, tools, compilers, etc.

What about the Mac side of things ? Mac's are still pretty big in audio production, but they don't support CUDA (and don't come with Nvidia GPUs any more anyway). How does Metal Compute stack up ?

We don't use Macs. While many uusicians use macs, the people on the engineering/scientific side of things don't. We're not doing music audio.

Well, Nvidia invented CUDA, it's no surprise that they won't give up their competitive advantage by properly supporting OpenCL or SYCL.

Edit: Also, SYCL should be supported by most things already supporting OpenCL, like Tensorflow for example.

I think OpenMP device offload has a better chance than SYCL or OpenCL. But I'm not in the industry.

CUDA's single-source environment is very good compared to SYCL or OpenCL. OpenMP still has single-source (just #pragma target), and is already adopted by the HPC crowd for multithreaded programming. NVidia and AMD support seems to be growing, especially because OpenMP is important to the supercomputer folks. (Summit and Frontier)

ROCm is CUDA-like and mostly works. The main issue is that CUDA will always remain a few steps ahead as AMD is forced to play "catchup". (ex: Cooperative groups are useful but unimplemented for now in ROCm). ROCm's main issue is being stuck in Linux, they need some Windows support if they want ROCm to really take off.

Didn't they plan to merge OpenCL and Vulkan?

this is nice, do you have any more details?

>gaining market dominance, both in the ML/GPGPU space

Is this really the case? Last I checked all the major Deep Learning toolsets ran off CUDA/cuDNN and there was nothing comparable for AMD hardware.

anyone have some experience on this? I'm in the market for an AMD laptop capable of running standard sci-kit/pytorch/etc but these seem optimized for NVIDIA cards. I'm curious about the outlook for these trending AMD cards.

Generally if you want to do ML tasks you want nVidia. They put the work in early to build the tooling so now the default assumption is that you're on nVidia hardware. It is possible to do some stuff on AMD cards, but you'll be on the cutting edge for that platform re-solving problems that were already solved on the nVidia side.

yes that was my conclusion after some research on this. there is an open issue on github for AMD support on pytorch, and looks like something works on arch linux, but really sounds like support is still in the hacking stage and far from production mode.

https://github.com/pytorch/pytorch/issues/10657

Support for PyTorch on ROCm is fairly good. I have built it from the dev branch without much trouble for all year. There has been ROCm CI even longer, my patch to print ROCm system information for bug reports was merged last week.

If you know where to look (i.e. it's public but unannounced) you can see that nightly wheels have been built for the last few days. So I would expect that some time between now and the Developer Day in November we'll see ROCm appear on PyTorch's "get started" page.

I'm almost positive that if you're talking about AMD GPU's your going to be out of luck. For Deep Learning especially NVIDIA is really the only serious option.

I agree. I think OP was referring to nvidia?

I recently ported a code to AVX512 (with ISPC) and was blown away by the performance. Sure, maybe fma can cause throttling but for me, I could replace a 28 core Haswell machine by a 10 core Skylake X (Xeon W specifically) with the new instruction set. A fatter one (16 core?) would compete with a high end Nvidia card and maybe a 32 core EPYC. Sure, it’s just my benchmark, but it is competitive.

Funny I just posted almost the exact same. Seems like we all have a soft spot for AMD!