207 comments

[ 10.8 ms ] story [ 258 ms ] thread
With Intel stuck in process hell for a while longer, I primarily hope AMD can gain ground on Nvidia. They are pushing a lot of innovation and gaining market dominance, both in the ML/GPGPU space and gaming (with ray tracing, upscaling, etc).

We need some competition to keep prices low and innovation high medium to long term.

Funny I just posted almost the exact same. Seems like we all have a soft spot for AMD!
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?
The industry at large is moving to SYCL, which can be implemented on top of OpenCL / CPUs / FPGAs.
this is nice, do you have any more details?
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.

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.

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.
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?
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...

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.

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.
> 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.
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.

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

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.
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.
> 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

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

> 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.
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.
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.

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.
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.
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)
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.

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

> 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...

>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.

I agree. I think OP was referring to nvidia?
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 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.
At the beginning of the year I thought AMD were going to have an amazing year. I think they’ll probably continue beating Intel but the new nVidia cards look really good, and very good value.

I hope AMD still have an ace up their sleeve.

> I think they’ll probably continue beating Intel but the new nVidia cards look really good, and very good value.

AMD have much lower power consumption, that was my main motivator for Ryzen and AMD Radeon as dev station.

Really? The general sense about AMD has almost always been the reverse, both as compared to Intel and Nvidia in those respective spaces.
5 years of process stagnation (Intel) will do that. Samsung's 8nm process used in Nvidia's cards is not as good as TSMC 7nm and that shows in their power consumption. Historically Nvidia's architectural efficiency has been much higher so it's not a given that RDNA2 will be more efficient even with the process advantage.
>I thought AMD were going to have an amazing year

its very amazing for AMD since Lisa Su became the CEO. we are talking about a stock not too long ago (2016) at $2 and in just 4 years it jump to $80.

Rumor has it AMD has a competitive GPU but will not pursue the high-end market mostly because of their limited TSMC slots, they can make much more money from a CPU wafer than a GPU one ( due to area size and yield )

But I really hope I'm wrong and they kick Nvidia ass since they become such intolerable arrogant anti-competitive pricks in the last +10 years.

TSMC slots are transferrable/sellable...

The fact Nvidia hasn't sold their slots to AMD for CPU production, yet AMD's GPU team has given up slots to their CPU team, tells you Nvidias GPU range is more profitable than AMD's one in this high-end market.

Doesn't that assume that the value of the slots is directly proportional to the price of the product you sell from those slots? That seems like quite a big assumption. For example, I would speculate that having a product shortfall (because you sold your manufacturing capacity) is going to significantly impact your market share and thus long term profitability.
For this generation, NV only makes the flagship datacenter/professional GPUs/A100 on TSMC, smaller GPUs are made on Samsung's 8nm process.
30xx parts are made at Samsung on a process inferior to TSMC's N7/N7P. This might indicate that Nvidia's datacenter/HPC parts are worth more than AMD CPU production, but that's not very shocking. Also it would be bad business to sell your slots to your competitor, so they could have ever greater production capacity on a leading-edge node.
The last time I owned an AMD video card was more than a decade ago. At the time their video card drivers needed work and could blue screen your box. There was a lot of back and forth on the internet about which driver version to load to get the most stable experience.

Fast forward to 2020 and the perception about AMD's video driver quality seems to be in a similar place: https://www.pcgamesn.com/amd/radeon-big-navi-driver-issues-f.... This may just be confirmation bias since there are also articles refuting that AMD still has driver issues.

On the gripping hand AMD themselves apparently acknowledge they have issues and are trying to address them . https://www.tomshardware.com/news/amd-drivers-update-downloa...

> Fast forward to 2020 and the perception about AMD's video driver quality seems to be in a similar place

Maybe for Windows, but the same cannot be said of Linux or macOS, where AMD's choice to open-source their drivers has led to these OSes having great support; while Nvidia, having to support their own closed-source drivers, seemingly hasn't had the bandwidth to get their drivers for these OSes up to par.

This. I've already decided my next GPU will be AMD regardless of if it's competitive at the top end in terms of performance. I am so sick of nVidia's Linux driver bullshit that I cannot wait to drop them. I don't game that much anymore anyway.
AMD is great for gaming on Linux, especially with projects like radv and ACO.
I'm not super interested in Linux gaming, I just want my desktop to be usable.
Honest question: If you're not interested in gaming, why even get a discrete GPU and not an integrated one? Are you interested in using it for other purposes?
I dual boot Windows for gaming. I work and code on Linux, except my monitor setup + nvidia graphics card results in a virtually unusable desktop. Integrated graphics can't run multiple 4k screens very well, if at all.
I do all my gaming on Linux and AMD works very well for it.
I'm not the author of the previous comments but I use Linux, not interested in gaming, and I have a dedicated GPU (2080ti). Mostly for working with Blender and Houdini, where the integrated one would work as well as using the CPU, at least for the workloads I'm dealing with, so having a dedicated GPU for both simulations and rendering helps a lot to speed up the workflows.
I can answer this question from my own perspective, at least. CPUs with integrated graphics are not as fast as standalone historically. Compare say, a ryzen 3600 to a 3400G.

Passmark below:

3400G: 9425

3600: 17850

for fun

1600: 12444

Their Windows drivers are a pain as well, to be honest. I've had more than a few blue screens from various versions of their drivers over the last few years.
Not really, drivers for Navi were equally bad on Linux. Second thing is that you need to wait several months before open source driver is stable, first months you have to use latest kernel + mesa-git. On the other hand, you can use Nvidia blob from day one.
And you can undervolt your AMD card with the open source drivers to improve efficiency and gain performance at the same time.
Linux support exists, but is a royal pain. With my 5500XT, even on Arch linux it wasn't properly supported unless I compiled a few things from Git.

I applaud the open source effort, but let's not pretend it's all sunshine and rainbows.

I happen to have a macbook pro now but 10 years ago that was the reason I avoided ATI cards. I had a Dell laptop with one and it always had driver's issues. On a windows machine I have nvidia card and I don't remember the last time I had an issue with it.
I've had multiple generations of Nvidia GPUs in my PCs, all the way back from a Riva TNT up to a GTX 460 as the last one.

I switched to an RX560 and I couldn't be happier. Linux drivers that actually work and aren't a mostly closed source garbage heap. The current Windows drivers are fine too.

Am I the only person on this site who has had no issues with the blob (aside from Wayland), especially on desktop? The last actual thing that I remember being a bit of a pain was an Optimus laptop years ago but from what I understand that’s finally been officially fixed; the desktop cards worked quite well, including the 2060 in my current desktop.
No, I've been using NVIDIA Linux drivers on RedHat/CentOS for the last 14 years and seldom had anything like the problems everyone complains of here.

I've not had much luck on laptops with NVIDIA Optimus though, maybe that's improved.

The current situation with Wayland and EGLStreams is regrettable though.

(comment deleted)
Yup, I've got about a dozen motherboard/video card combinations that I'm adminning. All of them require some sort of hack to get the binary driver to work with our code. Most require a kernel flag, some require sub-optimal bios settngs, some require pinning to an old version, and one requires a monitor attached to boot.
I've had 2 Nvidia cards (GTX760 and GTX 1070) and between those 1 AMD card (Sapphire Radeon R9 390) and it was the AMD that constantly BSOD on me. It would run well most of the time but crash my PC often. I appreciate the competition but not willing to risk spending so much if AMD still can't manage.
Sounds like a hardware defect to me. I've had the 380 and never had a BSOD even once.
Sapphire 580 here too zero issues
I was on the market for a GPU and I would agree with you and wouldn't say it's confirmation bias. On Amazon, most Radeons have 3.5 stars vs 4.5 for Geforces; many reviewers say they purchased a Radeon only to return it later because of driver issues.

Also, as the owner of Macbook 16" with a Radeon 5500M Pro, I'll add that I find the bootcamp drivers much subpar. They're also rarely updated although the issue seems to be Apple there.

Back in 2018 I was upgrading my computer, and I wanted an all-AMD system, so I picked up a 2700X and a Vega 64. I’m really happy with the 2700X, but I returned the Vega because the Linux drivers (open source and proprietary) were shoddy for separate reasons. I picked up a 1080ti instead and never had a problem.
Not had a single crash on the Linux open source drivers though, here my RX 550 is rock solid with excellent performance for the given hardware. We have two different Vulkan drivers and two different shader compilers to choose from, all open source.
Yeah I got a good deal on a RX590 last year and I still regret it. Tons of crashes and outright hard lockups on my desktop. Looking forward to the RTX3070 so I can toss the 590 in the bin
I don’t think AMD need to kick Nvidia’s ass to be successful or even competetive.

If they can land a decent GPU that’s between the 3080 and 3090 for an affordable price people will buy it.

They don’t need to beat the 3090 to win, I think that ship has sailed, let Nvidia have the high end market, there’s still plenty of space in the midrange, low end to succeed.

> but will not pursue the high-end market mostly because of their limited TSMC slots

Your second assertion that they can make more money from a CPU wafer than a GPU is probably correct, but assuming they don't abandon the GPU market entirely, they will still be using fab capacity to produce GPUs.

If they're using limited slots for GPUs anyway, doesn't it make sense to focus on the high-end high-margin parts first?

It makes sense to focus on the GPUs with the highest price per area of die. And remember that the larger the individual chip, the more of the wafer is lost to part failures. So for this rumour to be plausible, AMD’s better GPU tech would probably require larger chips.
If you don't mind, I have what is probably a dumb question.

Why does a GPU use PCI-Express still? I understand why they used to use it, but now they are not only standard, but in some cases the most important part of a PC.

I feel like if we could break the basic components out of the card, and bring it closer the system with a more modular design a bunch of things would be better.

Wouldn't it be nice if I could choose more CUDA cores if that's what I needed? Wouldn't it be nice if I could add more dedicated video RAM? Wouldn't it be nice if the the PCI slot wasn't a bottleneck? They're not graphics cards any more... often enough they're used for AI and scientific computing. It feels to me like moving these pieces to the motherboard would be super beneficial. Plus eaiser for other companies to get their fingers in things.

What makes you think that PCIe is a bottleneck? You can just populate your machine with another GPU if you need more memory or cores (thanks to the PCIe bus system).
While reading about OpenGL programming, one of the first things nearly all books mention, is the cost of moving data from your system to the GPU's memory.... So I assumed that's still true, is it not? Adding more GPU's does not change things as much as reduce the problem by providing more space.
That is true in a sense of: Of all the things in the graphics pipeline moving data between CPU and GPU is the slowest thing you can do. But that doesn't really mean the PCIe bus is slow. The truth is that most machines have a single GPU, connected to the CPU with all 16 PCIe lanes but they usually hit their performance cap utilizing much less then that.
This is coming out to PCs nowish: https://devblogs.microsoft.com/directx/directstorage-is-comi...

Nvidia's 3000 series will support it, the new consoles have it one form or another, and we can safely presume AMD's 6000 series will support it as well.

Under Windows, obviously. Linux will be a completely different bucket of chum.

> Wouldn't it be nice if I could choose more CUDA cores if that's what I needed? Wouldn't it be nice if I could add more dedicated video RAM?

How would you achieve any of that? The single meaningful difference between data rates in normal DDR memory and GDDR is that the latter is designed for signaling where the chips are soldered directly on a pcb with very short traces, whereas the former is designed to accommodate the longer distances and narrower buses necessitated by having socket connections.

Any hypothetical new GPU connector would still have to have the memory soldered on the same board as the GPU. (Or, if you are going for the more advanced packaging methods, connected through an interposer or EMIB or something.)

> Why does a GPU use PCI-Express still?

Because PCI-Express is basically the best anyone can do. It has been rapidly evolved to the point where in the past there was a lot of win by using custom busses, but now everyone just uses PCIe because it's the best available. (And I mean everyone, almost all of the remaining custom buses used in high-performance workloads are basically PCIe with a few restrictions loosened.)

Basically the only way you can go up from PCIe is to go wider. And if you do that, what you will likely to do is to make a LGA or PGA style socket format, that talks PCIe.

> Why does a GPU use PCI-Express still? I understand why they used to use it, but now they are not only standard, but in some cases the most important part of a PC.

PCIe is the fastest I/O mechanism in computers these days.

> I feel like if we could break the basic components out of the card, and bring it closer the system with a more modular design a bunch of things would be better.

GDDR6 and HBM2 RAM cannot be broken out. GDDR6 requires you to solder the RAM directly onto the board, while HBM2 RAM requires an interposer with 4096 individual wires connecting the HBM2 stacks to the GPU.

> more modular design a bunch of things would be better.

Its very difficult to make a "modular" set of 4096 wires (HBM2). Or a "modular" set of GDDR6x (which is now using 4-bits per clock cycle per pin, thanks to PAM4: https://www.anandtech.com/show/15978/micron-spills-on-gddr6x...)

The quality of the connections requires direct soldering in the case of GDDR6x, or incredibly specialized packaging in the case of HBM2. These RAM architectures, while incredibly fast, are almost impossible to extend.

-------

> It feels to me like moving these pieces to the motherboard would be super beneficial.

Xeon Phi tried to do that, and it failed. I'm not entirely sure what you expect to gain from moving the GPU closer to the center of the motherboard.

Again: high-speed RAM trades flexibility for high-performance. So any GPU using the high-speed RAM like HBM2 or GDDR6 will be stuck at very limited configurations.

-------

DDR4 and DDR5 are an order of magnitude less bandwidth, but their ability for multiple configurations is what makes the motherboard even possible.

Remove the ability to buy more or less RAM per computer, and what do you have? You just have a graphics card. Its not like you'll be able to pipe information from your hard drives or SSDs faster than a PCIe pipe anyway.

If the rumors are true about increased frequency and IPC for Zen3, we should be in for quite a ride. Zen4 and a shrink to 5nm next year ought to produce some further benefits.

I’m too old to “hate” Intel but the improvements we are seeing from both sides these last couple of years make clear how bad the stagnation had gotten. Gotta have competition!

I’m worried that the choice of not going first with the RDNA2 presentation is because they know it’s going to be disappointing and not competitive with RTX 3000.
It's not out of the question to save the best part of a presentation for last.
Sure - but waiting to announce (and hopefully release) for over a month after your competitor is SHIPPING is a sure way to lose sales. At the very least they need to leak potential pricing and performance in the mean time.
Also, many are probably looking into a new pc upgrade with the new RTX gpu's. I know I am. But I will most likely go for Intel 10900 CPU instead of the 3900X, the 10900 is also cheaper than the 3900X and XT in my country.
Is October 8th really a month after Nvidia is shipping?

I'm not sure how many units Nvidia is really going to move before then...

Per the article you're commenting on... RDNA2 is being announced October 28th. NVidia is SHIPPING RTX 3k on the 17th.

https://www.techradar.com/news/rtx-3080

So yes, October 28th is more than a month after September 17th. Maybe read the article next time before you downvote.

I didn't down vote.

Sorry I misread the article.

Anyway I am still highly, highly skeptical Nvidia is going to move a lot of units so close to launch. The first batch is going to sell out in moments and from then on they will only be available way above retail price. Happens every time...

Unless they are going to announce that they will stop letting motherboards lock your cpus, I don't see why I should be excited.
Why wouldn’t you buy a Ryzen because some epyc chips get locked?
I can't support AMD as a company as long as they are doing that on any of their products.
So what kind of CPUs do you buy, with both Intel and AMD out of the game?
ARM based platforms such as Allwinner and Rockchip CPUs which are free of these security measures. Though the performance is very poor in comparison, however over the next few years the situation might improve.
Not my niche, but Talos II probably serves that niche well.

Its workstation class, aimed at $5000 computers on a completely open framework.

I'm not convinced that the AMD EPYC feature is malicious however. So I personally would still go with Threadripper or whatever.

I'm mad at intel for what they did in the past, im mad at amd for what they are currently doing, which do you think will win out?
As Intel's Boot Guard is essentially the same feature, and not a thing of the past, my guess was neither of the two. I didn't try to be snarky, I'm really interested in alternatives.
I can't find anything on google about Intel boot guard meaning that plugging a cpu into a motherboard can bind that cpu to that motherboard such that putting it into another motherboard no longer works.
Currently using an Ryzen at the moment, my next CPU is likely to be a used Xeon because of AMD's PSP nonsense, which has reached a new low with the EPYC CPU 'bricking'.

On the Intel chips the ME security has been cracked completely so in theory someone could write an open source ME firmware replacement, although this might require a hardware mod to bypass the signature checks.

This seems more like a reason to avoid Dell than AMD.
I've always avoided dell, AMD should be avoided for giving Dell a chip API that let dell do what they did.
It was already a dire situation running a Mac with eGPU or a Hackintosh even before the RTX 30-series announcement, and much more dire after that.

The likeliness of Apple allowing NVIDIA to produce Mac drivers again is close to zero. It’d be nice for AMD to be able to compete with GPUs again.

Apple Silicon will contain an Apple-designed GPU, so it might soon be irrelevant for Mac users what AMD has to offer.
Apple's GPUs achieve "integrated" level performance and it would be a large step for them to achieve AMD or nvidia performance levels. Maybe they can scale it up sufficiently, or put it on a separate die with its own heat dissipation, but that seems like biting off too mcuh.

My MBP has the intel integrated graphics, and AMD discrete graphics. I imagine as Apple moves to Apple silicon it will be the same arrangement at tiers that currently have discrete graphics, with the intel integrated just being swapped out for Apple integrated, leveraging the discrete graphics when appropriate.

Given the performance of iPad Pro graphics, I could easily see the Apple GPU being competitive with AMD’s graphics in thermally constrained scenarios.
I highly suspect Apple will still ship (and support) AMD GPUs - maybe for the iMac Pro and Mac Pro. I doubt that Apple Silicon GPUs will be close to the performance to AMD/NVidias flagships. And I assume the market share for iMac/Mac Pro is just not that big for Apple to invest in competing with AMD/Nvidia on the Highend Segment.
OTOH Apple also needs dedicated GPUs for some of their MBPs and the bigger iMacs.

I think they will end up ditching AMD to have 100% E2E control on all their products.

Apple's MBP drivers have been broken since 2016 and getting worse, high-frequency locked and thereby thermally throttling the CPU at idle. I don't know if that means Apple Silicon will be better or if they are just incompetent at it.
I can also easily imagine AMD in the Mac Pro, but only as an accelerator rather than to drive displays.

(And I’m also predicting Apple will have more FPGA programs ready for the Afterburner card. This accelerator may even be the distinguishing feature of a future iMac Pro.)

>And I assume the market share for iMac/Mac Pro is just not that big for Apple to invest in competing with AMD/Nvidia on the Highend Segment.

Well it is the same argument for CPU as well. Shipment volume dont make much sense for a Custom CPU in those TDP.

But it really seems they are going All in. Custom CPU, GPU and likely NPU on all of their platform.

What do you even do with a GPU on a Mac these days? I guess there are a decent amount of games now, but it seems like it'd be mostly limited to "creative" stuff since CUDA is out of the question.

Imagine how much more dire it's going to get once Apple has ditched Intel. Are they going to continue using third-party, discrete GPUs? You definitely won't be able to use one off the shelf (at least without flashing the card's BIOS)

You need to watch youtube these days.

Intel integrated GPU has a huge downside in laptops - offloading to it causes it to heat up and since, so close to the CPU...CPU gets heavily thermal throttled. That's true for any laptop that isn't inch or more thiccc.

Also, Apple will be switching to their own GPU in a few years.

As a linux-only user (and gamer), nvidia is a plain no-go for me so i just hope that rdna2 is not entirely disappointing like RDNA1 was. For us Linux gamers a working VFIO on a consumer card would be a dream of course since nvidia is locking that to their workstations cards, maybe AMD just throws it in as a goody.

Zen3 is propably gonna kick-butt, especially in any non-gaming task perhaps AMD will inch closer to Intel in gaming loads aswell, but i don't care too much.

Wouldn't SR-IOV support (which is rumored in nvidia's chips) make vfio on this kind of setup possible?
VFIO doesn't work with Nvidia devices because of a design decision by Nvidia. Their drivers refuse to load in a virtualized setting unless you buy their enterprise GPUs. You can patch out a few checks and everything works again. Another option is to reload the GPU firmware before the virtualization host boots so the GPU doesn't know it's the second time it boots.

Even if it's possible to implement proper VFIO for Nvidia GPUs, I expect them to fight such an implementation with tooth and nail in their drivers. This is the same company that artificially limits the amount of transcoding streams in their drivers, mind.

VFIO is already possible with amd if you have supporting hardware i.e. i was running single vega guest windows guest and linux furyx host on tr4 board but with sr-iov you could run many guests using same gpu, i think stadia uses this now. There were also rumor that vega has sr iov but disabled on hardware
Uh, yeah you are right i meant SR-IOV
I have had perfect working VFIO setups with both the 1070 and 1080ti from nVidia, there are many comprehensive guides and working around the driver error code 43 is quite trivial.
That's nice but an error thats designed to fuck me as a customer is a good reason not to buy from that company, and the general state of nvidia drivers on linux makes me not want to chose them for a new build. Nonetheless, i'm glad to hear you got it working, it sure is a fun thing to do.
> the general state of nvidia drivers on linux

... is great? They have been well supported for decades, has working drivers for every GPU on release day and even 10+ year old cards get updated drivers for new kernels / X servers whereas AMD takes months/years to have usable drivers in a stable kernel.

Is that so? I always here: stay away from nvidia on Linux, especially but not limited to using Wayland. Is that truly not the case?
Well, my experience [1] has been nearly perfect, certainly much better than what I've had with AMD. Nvidia has been pretty much plug&play for me for a long time, with very few issues.

For Wayland yes, Nvidia is using a different approach than AMD/Intel so most desktop environments don't support the Nvidia way, resulting in a poor experience. That said, Wayland desktops as a whole are still quite some time away from reaching feature parity with X, so I'm not too worried just yet.

[1] For reference, I've been using Linux (and only Linux) on the desktop for ~18 years.

Until NVIDIA decides to "fix" the workarounds, again.
I will get a 3000 series. While I've never actually owned a AMD graphics card.... (25 year Linux guy) For a really long time AMD on Linux was absolute garbage, like lower end NVIDIA beating the this crap out of high end AMD. And I really appreciate the open source stuff, it seems to still be problematic. The other issue is, I use DaVinci Resolve which works a million times better with NVIDIA than AMD, even if it works at all.
I'm pretty excited to see what they'll offer in the 10- to 16-core range for Zen 3 and if they'll hit the magical 5GHz mark.

Anyone else remember the promised 10GHz of the NetBurst? It's nice to see the industry slowly creeping up to that number.

Besides processors, AMD knows how to design a good hype train.
The major difference is that they have a product to back it up. Also, a large part of the hype comes from the "market" and you can blame Intel and Nvidia for that, since they have been screwing and exploiting said "market" for more than a decade now.
They are masters of the hype cycle. The hype bicycle, nowadays.
Including the most important part of the AMD hype bicycle, backpedaling.
I wonder if Zen 3 will finally let AMD pass Intel in single threaded performance. I thought they had already, but just a few days ago I checked benchmarks and saw that my 8086K still beats a 3800XT.

Though maybe I'm stuck in the past a bit thinking that single threaded performance is even still important for games.

> single threaded performance is even still important for games.

I'm looking at you, TF2 & PCSX2.

>TF2

Is Team Fortress 2 really that single-thread bound?

Source 1 does basically everything in one thread and is a ~15 year old D3D9 renderer.
Maybe they meant Titanfall 2.
It only is if you're running low resolution and/or really high refresh rate. Otherwise a top tier GPU is still your limiter.
Its not just games, firefox with ublock+ and a dark mode plugin are noticeably faster on fast single threaded machines.
Agreed, it's funny how nearly all discussions about reasons for single threaded perf just focus on games.
I was going to add "and typical PC tasks" but ended up taking it out because everyone's version of typical on a PC is different.
Yes single threaded perf is still important, but really only if you have top-end GPUs. 8 - 12 threads is where multi-threaded gaming gains seem to wane at bit, which might make sense given last console cycle's architectural optimizations. Once you're at 8 - 12 threads frequency starts becoming more important. Most recent Intel K CPUs can hit 5GHz+. The 10900K is still the best performing gaming CPU out there, and 8086K is still very competitive. Still seems many game engines are optimized for Intel and even with Zen2 having an IPC advantage, it cannot muster high enough frequencies to beat Intel's top gaming CPUs.

Maybe Zen3 will change that though, more IPC gains and possibly higher frequencies. If AMD can put out at 5GHz chip, it would probably demolish Intel's current offerings.

Where do you consider the top-end to start in Nvidia's lineup? I know that gets a bit awkward to answer since we're not far from being able to buy 30xx cards, so probably best to just pretend we don't know about those for this question.
2080 Super / 2080 Ti. Really seems like you need to be pushing the cutting edge to justify a top-end consumer CPU. If you're spending the money on a 2080 Ti class of card, then you should also be spending money on the best CPU to drive it, though when you're at the top-end the money/performance ratio starts getting significantly more biased towards the money side.
Very excited about this. I wasn't really a fan of AMD, but I built a machine with a Ryzen 3700X last year and I've been super impressed.

Personally I don't really care if RDNA2 is not as powerful as the Ampere GPUs as long as the price is competitive. I want to update my 1070 just for playing Cyberpunk 2077 when it comes out, but since I don't play that much anymore it doesn't make sense to go all-in on the GPU.

The elephant in the room is (non-free) drivers/compilers

You can lose 90% of theoretical performance due to code quality and (especially in a laptop) thermals.

AFAIK AMD is more open than the alternatives
I recently bought a Ryzen 3900X. I hope Linux eventually supports these processors better. There has been some serious stabilization issues I eventually tracked down to being related to my processor. These issues required me to spend a significant amount of time researching. Things seem to just work with Intel processors. I'm hoping as popularity increases support for these processors will improve.
>serious stabilization issues

Care to go into more detail? I've just installed a server with a 3700X and the only issues I've experienced have been that I need a custom kernel module in order to get CPU temperature and power data (zenpower)

>I need a custom kernel module in order to get CPU temperature and power data (zenpower)

Is this on an old Ubuntu version or something? This was only true in the early days of these CPUs, which was last year. Temps have been working for a long time in the Arch world, but maybe it's different in point release hell.

>Is this on an old Ubuntu version or something?

I don't know, although I'm running the latest Ubuntu (20.04.1), with kernel 5.4.0-45-generic

It's been a while, so I have since forgotten the details... but I believe I needed to make a few changes to the BIOS to turn off idling or something like that.
I've been on Linux (mostly Manjaro, some Arch) with a 3900X for over a year, and I haven't had problems.
FC31 and FC32 on 3900x for half a year and zero issues either.

I suspect old kernels can be problematic. I had to upgrade past LTS to get my Ryzen 4700U laptop to even get video output.

I had 3900x on Ubuntu 18.10, 19.04 and 20.04 no problems whatsoever (Except with the nVidia GPU).
Linus has a Threadripper now - so it will be stabilized. :-)
I've been running a 3900x under Linux since release (got it 14.07) without any issues. Support was great day one, but of course you needed a new kernel, not something that was running on your 2010 web server.
I'm in the market for an apu ? should I wait for these new processors or just bite the bullet on a 3400G. not a serious gamer, or hardly play games at all. but an alternative to the xbox one s, I have would be nice. games I would like to play e.g strategy | tactics are not available on console.
The AMD naming conventions are a little confusing, so it's not obvious, but the 3400G is based on Zen+. They have announced Zen 2 APUs [0], but they are not available directly - just in pre-built systems. That may change at some point. These are a huge improvement over Zen+, but I wouldn't wait around for retail availability of Zen 3 APUs!

[0] https://www.amd.com/en/products/apu/amd-ryzen-7-4700g

I feel RDNA 2, at the very best would prove to be something in between of what Zen+ and Zen 2 was against Intel, so while it may not outperform Nvidia on highest end, it ends up providing an overall better value, leading to people buying over Nvidia.

It is worth noting that unlike, Intel, Nvidia isn't sitting ducks when it comes to GPU innovation so that's a harder hill to climb for AMD. And fortunately AMD now has their cash cow running with Ryzen, EPYC and Consoles so they can allocate more R&D budget for GPU.

I commented earlier, but figured to add some more:

I'm rooting for AMD and also looking to replace a 2012 MBP with an AMD laptop. The defacto cards for ML work seem to be NVDIA (eg their RTX line, https://timdettmers.com/2020/09/07/which-gpu-for-deep-learni...)

Does anyone know if pytorch/sci-kit might offer more support to AMD cards?

PyTorch is increasing support. It had CI for a long time, has been buildable for all this year without too many caveats, and I would expect nightlies to be advertised sometime soon, maybe the next stable release will even also have it.
Would an AMD event like this usually be an announcement about future products, a paper launch, or could one expect to buy one of these processors soon like? I'm looking to build a new system around an RTX 30-something. I would even consider a pre-built system if I can't get my hands on retail parts.
Same here; I decided about a month ago that I wanted to upgrade, and that I wanted to go AMD this time, but then I found out Zen 3 was right around the corner so I've been anxiously awaiting its release so I can move forward
AMD has said in the past that the GPUs will be available before the new consoles. I'd expect very limited availability sometime early November. CPUs maybe a bit earlier and a bit better availability.
This is an announcement of an announcement. Those are off topic here. There's no harm in waiting.

https://hn.algolia.com/?dateRange=all&page=0&prefix=true&sor...

https://hn.algolia.com/?dateRange=all&page=0&prefix=false&so...

This announcement of an announcement is very important for people looking to buy a GPU now.
Fair point, but we have to go for the global optimum, which is having the most interesting front page. That means sacrificing local optima, so you may be right that there isn't zero harm in waiting. Most of the people you mention (at least the ones who read HN) are probably googling for GPU information, though, so I doubt there will be too many GPU purchase disasters.

The trouble with "announcement of an announcement" posts isn't just that they're unsubstantive and lead to generic discussions (https://hn.algolia.com/?dateRange=all&page=0&prefix=true&sor...). It's also that when the real announcement is made, we end up with a duplicate thread.

I suppose this follows from the competing interests of the publishers and this forum. Their interest is to copy the same information to as many places as possible; our interest is to deduplicate it.

(comment deleted)
(comment deleted)