Wish they would release\support gpufort or whatever amd calls their platform, so stable diffusion works, and we can get benchmarks. Nvidia cuda is really a market monopoly, and AI is everywhere now. Voice changers for discord/streaming, Stable Diffusion, etc.
All in all HIP is pretty decent, from what I understand. There's just a lot a lot a lot of algorithms CUDA & associate also-on-metal add-on libraries have & it's hard to have great coverage. AMD keeps adding more. With Intel also trying to get into GPUs I feel like there's more chance something longer term viable & open emerges, whereas it's been AMD with their compatibility tooling fighting this war alone for a long time.
Sounds like GPUFORT is out but I dont think it's very relavant...
> Most of those efforts to date have been focused on C/C++ code while GPUFORT is about taking CUDA-focused Fortran code and adapting it for Radeon GPU execution. GPUFORT supports the source-to-source translation of CUDA Fortran and OpenACC-based Fortran code over to OpenMP 4.5+ for GPU execution or Fortran + HIP C++ code.
Longer term, it should matter less. More and more machine learning platforms will target Vulkan or other generic acceleration frameworks. Developers wont target CUDA directly forever- they increasingly do & increasingly will use good ML frameworks. Everyone but bad guy Nvidia at least will harken increasingly in that direction.
Honestly, the bit I’m happy to see is DisplayPort 2.1, as it might finally open up a lot more options in the display market.
I’ve been waiting a decade for 5K120 to be ubiquitous. That it might start existing with real selection (and competitive pricing) in the next year is fantastic. For the record, I’m talking about for productivity, not gaming. I’m aware that 4K is everywhere, but I’ll take the extra 77.8% logical work area every time.
Absolutely. Personally, I'll be excited to see higher colour depth for my next HTPC GPU and monitor upgrade. I can really notice colour banding, and having 12 bit colour will be a game changer for content consumption.
I'm disappointed to see a max 3 DP + 1 HDMI output though. Was hoping I'd be able to attach a VR set without unplugging my 4th monitor some time in the future.
Wait, how big does your screen physically need to be to even call 5K a win over 4K. Surely under normal monitor sizes (~32 inch) the dots are equally invisible?
4K at 32-inch isn't enough, which is why Apple uses 5K for 27-inch screens. It's not just about the pixel size, but using integer scaling and getting reasonably-sized display elements.
Personally, I find the 8K ultra-wide announced via the Samsung partnership more exciting; that's exactly what I've been waiting for.
What content are we viewing? If a movie, I can't tell the difference between 4k and a higher resolution. If I'm looking at a desktop UI, I can easily tell the difference.
I think around 27"/8k (326 ppi) is where I really couldn't tell. 4k laptops are already at this density.
>Wait, how big does your screen physically need to be to even call 5K a win over 4K
As I mentioned with "productivity, not gaming", preferring 5K over 4K isn't about making picture prettier.
This is about having a 5K 27" screen with 177.8% of the work area of a 4K 27" screen. A 5K display (5120 x 2880) is 14,745,600 pixels, whereas a 4K display (3840 x 2160) is 8,294,400 pixels, yet a 27-inch panel is a 27-inch panel. It's about more stuff in the same area. It's why I've been using 1440p panels since 2010 instead of 1080p panels (and a 1920x1200 panel back in 2004). Just now everything on the screen can be crisper, since we use four pixels (2 x 2) to define the same logical area we used to define with a single pixel.
It's like having a 77.8% larger desk, except the desk is the same physical size as the "smaller one." This is about more of my work being visible at once.
Sure but why would you want that extra pixel realestate on a 27 inch or even 32 inch monitor? It's like zooming out and scaling everything down so you can fit more on your screen with the same sharpness, the thing is, you wouldn't want things to be that small on a 27 inch screen
Great to see AMD focusing on performance per watt. I might very well jump ship now; will be interesting to see the reviews - I'm into the kind of "poorly optimised" simulators that saw a massive impact from the 5800X3D. The NVIDIA 4090, whilst having amazing performance is just too much in both price and wattage.
NVIDIA's questionable business practices have also left a lot to be desired; they easily forgot who got them into the #1 position when miners were lined up with their money during the lock-downs.
If you are referring to DCS, I think that's absolutely down to the cache on the CPU and the memory subsystem as a whole.
I ran a Threadripper Pro system and it had the most stable VR FPS I have seen anyone put up (if not the fastest; clocks matter). I attributed it to cache and RAM bandwidth (8 channels... is a lot) and the 5800X3D numbers seem to support that.
That's awesome re the Threadripper Pro! Yes, DCS World in particular, but also Star Citizen (where nothing else comes close at the moment) and to some degree iRacing.
It went from stuttering all the time (bad 99% lows) on a well overclocked (all core 4.4ghz) 3900XT to butter smooth frame rates in DCS - and that's including custom scripts I had to exclude before (Streamdeck integration, etc).
I can now push DCS World to three screens; even on my lowly (for the task) 3070, the lowish 30-40-50 depending on weather FPS numbers combined with Free(G)sync monitors are never stuttering so it still feels smooth. I attribute it to the CPU never being the weak link; also, have reasonably good memory (4x 16GB C16-3600 Neos at XMP/3600mhz).
For gaming, there's no better value for money right now than the 5800X3D - if you already have an AMD system. And even then, if not, and you're not made of money.
I'm likely going to do another CPU upgrade as well, probably end of next year once/ if DDR5+AM5 mobo pricing has settled down a bit - but mostly because I want my son to have a better PC for his gaming as well. (His 6850k system is getting a bit old.)
Considering how Nvidia uses GPU RAM to segment their products between high end consumer GPUs (where the 4090 hasn't progressed at all from the 3090 in still having 24 GB RAM) versus the crazy high end ML accelerators, I've recently wondered how it would work out for AMD if they slapped 48+ GB RAM (sufficient to run GPT-NeoX-20B..) on one of their higher end parts (ideally without a large price premium) in order to excite the people trying to run ML on the cheap that are currently bottlenecked by RAM on consumer cards and spur the software development required to get more things running on AMD.
Like an AMD version of an RTX A6000? Granted $4500 still isn’t “cheap” but for a dual slot blower card with great performance and 48GB they do really well in the mid-range of ML.
I don’t really pay attention to AMD cards anymore. I’m not a gamer and every time over the past five years I’ve put a toe in the water of the disaster that is ROCm, their drivers, “support” in frameworks, etc I run back to CUDA.
It really seems like AMD has ceded everything other than supercomputer cluster ML to Nvidia and CUDA. They do well in gaming and they can sell and support the extremely high end of ML with mostly bespoke applications that have teams of professional support and institutional backing. When you’re spending $100 million or whatever on hardware flop/$ really counts and you can throw people at making it work.
In the low to mid ranges of ML the 25% or even 50% savings on hardware is quickly lost when compared to the human time sink that is the entire AMD ML “ecosystem”. If you’re buying 10 GPUs you’re talking about saving what, $20K? That “savings” gets wrecked really quickly when you look at salaries and time spent chasing yet another bug/quirk/showstopper with ROCm.
The problem is “on a budget” applies if your time is free. Most of the ML ecosystem is still gold plated - talent, hardware, etc with talent being the biggest cost for most applications.
Frankly, Nvidia doesn’t exactly foster goodwill (just look at any HN thread when they come up). I suspect that a decent number of people are willing to throw time and pull requests into the ROCm ecosystem just out of spite for Nvidia. Good for them because Nvidia needs a serious competitor!
I’m rooting for AMD but I’ve always been in the position of “it needs to work now in the shortest and cheapest path of least resistance” which has always been, and will be for the foreseeable future, CUDA.
I've been telling myself I'd buy the next flagship AMD since before Vega. I think I might actually do it this time!
AV1 encode is exciting. The rumormill has been going back & forth between monolithic vs chiplet. Cool to see they seemingly did it. It definitely comes at some kind of power cost, some struggle, but given the expense of modern processes it just seems infeasible to build cutting edge chips at the dame huge sizes. This had to happen, so faults in fabbing affect smaller pieces of the chip.
My only ask here is, I wish AMD would add some USB4 outputs. It's not normal- monitors arent built for it but they could be- but ideally with 60w+ power delivery too. Add pcie/thunderbolt. Make this the io card of the computer.
Passing even more power through the videocard sounds like an expensive challenge. Presumably you'd also prefer the card to be able to drive 2-3 monitors, which is already 1-2 additional PCIe 12v 8 pin connectors (old tech). That also assumes monitors which can be driven by 60W, which seems unlikely where these cards are used.
USB cables are also only rated for such high data transfer rates over very short distances, or active cables; which are inherently more expensive. DisplayPort cables have claims, by well known brands near the top of google results, of supporting passive cables of at least 15 meters. A single HDMI is included to accommodate devices like TVs, presentation systems, etc.
I would be unshocked to see a usb-c to displayport pigtail adapter & long DisplayPort cable have degredation. Given that cards are often space limited & sometimes use things like mini displayport to cram more ports in (since the gpu can often drive 5 or 6 displays), having a small usb-c port or 2 seems like a really good alternative, with litte sacrifice, and great potential usability boon.
The elrlectrical doesnt seem daunting to me. People hem and haw over how a $5 BOM cost can escallate in real world impact, but it feels like significant over-concern, amd when you are talking about a GPU costing up to $1000, it feels like an obvious way to differentiate & make users happy.
The jack wouldnt have to be used for video. Being able to plug a laptop in or a powered hard drive or a battery to charge would be a gratis service. I certainly hope over time we see more usb pd powered monitors come out, but right now I just feel like this ultra high end io monster should also support modern io standards & connections, and adding a stupidly efficient (98%) $2 buck regulator, a $2 control chip to support thr obviously better nice version of the io standard that also does power is not just obvious, but painfully obvious, and it stinks & is a glaring smell that there's literally no options here, for these expensive adapters.
25 comments
[ 4.0 ms ] story [ 59.6 ms ] threadAll in all HIP is pretty decent, from what I understand. There's just a lot a lot a lot of algorithms CUDA & associate also-on-metal add-on libraries have & it's hard to have great coverage. AMD keeps adding more. With Intel also trying to get into GPUs I feel like there's more chance something longer term viable & open emerges, whereas it's been AMD with their compatibility tooling fighting this war alone for a long time.
Sounds like GPUFORT is out but I dont think it's very relavant...
> Most of those efforts to date have been focused on C/C++ code while GPUFORT is about taking CUDA-focused Fortran code and adapting it for Radeon GPU execution. GPUFORT supports the source-to-source translation of CUDA Fortran and OpenACC-based Fortran code over to OpenMP 4.5+ for GPU execution or Fortran + HIP C++ code.
Longer term, it should matter less. More and more machine learning platforms will target Vulkan or other generic acceleration frameworks. Developers wont target CUDA directly forever- they increasingly do & increasingly will use good ML frameworks. Everyone but bad guy Nvidia at least will harken increasingly in that direction.
I’ve been waiting a decade for 5K120 to be ubiquitous. That it might start existing with real selection (and competitive pricing) in the next year is fantastic. For the record, I’m talking about for productivity, not gaming. I’m aware that 4K is everywhere, but I’ll take the extra 77.8% logical work area every time.
I'm disappointed to see a max 3 DP + 1 HDMI output though. Was hoping I'd be able to attach a VR set without unplugging my 4th monitor some time in the future.
* 2x DisplayPort 2.1
* 1x HDMI 2.1
* 1x USB-C, likely supporting DisplayPort as said above
Some add-in board (AIB) partner cards may replace this USB-C port with another DisplayPort port.
Now if you have an APU or CPU with iGPU, then it may be possible to enable both discrete and onboard graphics so you can fully utilize them.
Personally, I find the 8K ultra-wide announced via the Samsung partnership more exciting; that's exactly what I've been waiting for.
I think around 27"/8k (326 ppi) is where I really couldn't tell. 4k laptops are already at this density.
As I mentioned with "productivity, not gaming", preferring 5K over 4K isn't about making picture prettier.
This is about having a 5K 27" screen with 177.8% of the work area of a 4K 27" screen. A 5K display (5120 x 2880) is 14,745,600 pixels, whereas a 4K display (3840 x 2160) is 8,294,400 pixels, yet a 27-inch panel is a 27-inch panel. It's about more stuff in the same area. It's why I've been using 1440p panels since 2010 instead of 1080p panels (and a 1920x1200 panel back in 2004). Just now everything on the screen can be crisper, since we use four pixels (2 x 2) to define the same logical area we used to define with a single pixel.
It's like having a 77.8% larger desk, except the desk is the same physical size as the "smaller one." This is about more of my work being visible at once.
NVIDIA's questionable business practices have also left a lot to be desired; they easily forgot who got them into the #1 position when miners were lined up with their money during the lock-downs.
I ran a Threadripper Pro system and it had the most stable VR FPS I have seen anyone put up (if not the fastest; clocks matter). I attributed it to cache and RAM bandwidth (8 channels... is a lot) and the 5800X3D numbers seem to support that.
It went from stuttering all the time (bad 99% lows) on a well overclocked (all core 4.4ghz) 3900XT to butter smooth frame rates in DCS - and that's including custom scripts I had to exclude before (Streamdeck integration, etc).
I can now push DCS World to three screens; even on my lowly (for the task) 3070, the lowish 30-40-50 depending on weather FPS numbers combined with Free(G)sync monitors are never stuttering so it still feels smooth. I attribute it to the CPU never being the weak link; also, have reasonably good memory (4x 16GB C16-3600 Neos at XMP/3600mhz).
For gaming, there's no better value for money right now than the 5800X3D - if you already have an AMD system. And even then, if not, and you're not made of money.
Currently waiting for a Ryzen 7000 series X3D CPU. Should be available early next year I think. I'll upgrade to the 7900XTX then too.
I'm likely going to do another CPU upgrade as well, probably end of next year once/ if DDR5+AM5 mobo pricing has settled down a bit - but mostly because I want my son to have a better PC for his gaming as well. (His 6850k system is getting a bit old.)
I don’t really pay attention to AMD cards anymore. I’m not a gamer and every time over the past five years I’ve put a toe in the water of the disaster that is ROCm, their drivers, “support” in frameworks, etc I run back to CUDA.
It really seems like AMD has ceded everything other than supercomputer cluster ML to Nvidia and CUDA. They do well in gaming and they can sell and support the extremely high end of ML with mostly bespoke applications that have teams of professional support and institutional backing. When you’re spending $100 million or whatever on hardware flop/$ really counts and you can throw people at making it work.
In the low to mid ranges of ML the 25% or even 50% savings on hardware is quickly lost when compared to the human time sink that is the entire AMD ML “ecosystem”. If you’re buying 10 GPUs you’re talking about saving what, $20K? That “savings” gets wrecked really quickly when you look at salaries and time spent chasing yet another bug/quirk/showstopper with ROCm.
Frankly, Nvidia doesn’t exactly foster goodwill (just look at any HN thread when they come up). I suspect that a decent number of people are willing to throw time and pull requests into the ROCm ecosystem just out of spite for Nvidia. Good for them because Nvidia needs a serious competitor!
I’m rooting for AMD but I’ve always been in the position of “it needs to work now in the shortest and cheapest path of least resistance” which has always been, and will be for the foreseeable future, CUDA.
AV1 encode is exciting. The rumormill has been going back & forth between monolithic vs chiplet. Cool to see they seemingly did it. It definitely comes at some kind of power cost, some struggle, but given the expense of modern processes it just seems infeasible to build cutting edge chips at the dame huge sizes. This had to happen, so faults in fabbing affect smaller pieces of the chip.
My only ask here is, I wish AMD would add some USB4 outputs. It's not normal- monitors arent built for it but they could be- but ideally with 60w+ power delivery too. Add pcie/thunderbolt. Make this the io card of the computer.
USB cables are also only rated for such high data transfer rates over very short distances, or active cables; which are inherently more expensive. DisplayPort cables have claims, by well known brands near the top of google results, of supporting passive cables of at least 15 meters. A single HDMI is included to accommodate devices like TVs, presentation systems, etc.
The elrlectrical doesnt seem daunting to me. People hem and haw over how a $5 BOM cost can escallate in real world impact, but it feels like significant over-concern, amd when you are talking about a GPU costing up to $1000, it feels like an obvious way to differentiate & make users happy.
The jack wouldnt have to be used for video. Being able to plug a laptop in or a powered hard drive or a battery to charge would be a gratis service. I certainly hope over time we see more usb pd powered monitors come out, but right now I just feel like this ultra high end io monster should also support modern io standards & connections, and adding a stupidly efficient (98%) $2 buck regulator, a $2 control chip to support thr obviously better nice version of the io standard that also does power is not just obvious, but painfully obvious, and it stinks & is a glaring smell that there's literally no options here, for these expensive adapters.