crazycrave
2[H]4U
- Joined
- Mar 31, 2016
- Messages
- 2,243
Just a joke like the RTX 5090 turned out to be so far missing ROP's and going Pop with the power stage.
![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
-
Some users have recently had their accounts hijacked. It seems that the now defunct EVGA forums might have compromised your password there and seems many are using the same PW here. We would suggest you UPDATE YOUR PASSWORD and TURN ON 2FA for your account here to further secure it. None of the compromised accounts had 2FA turned on.
Once you have enabled 2FA, your account will be updated soon to show a badge, letting other members know that you use 2FA to protect your account. This should be beneficial for everyone that uses FSFT.
RTX 50 ends 32 bit phyx support
- Thread starter ChadD
- Start date
- Joined
- May 7, 2007
- Messages
- 13,438
The original Ageia PhysX processors were the equivalent to an NVIDIA Series 8 8500GT from 2007.
lol, I think that 960 would be more than enough.
![Zarathustra[H]](/data/avatars/m/9/9298.jpg?1723407683){kind=avatar}
- Joined
- Oct 29, 2000
- Messages
- 42,357
I don't have a good understanding of how PhysX calculations have scaled over time and with resolution. Maybe we need more power today than back then?
Also, I wonder how much PCIe bandwidth a dedicated PhysX GPU might use. Back in the day, a 960 would have benefitted from 16x Gen3, but with modern motherboards, you either have to use slot 2, and drop your main GPU down to 8x, or stick it in one of those tertiary slots off of the chipset where you may only have 1x or 4x Gen4
I do agree that a 3050 seems like overkill though.
The last time I did this (again I think it was with a 768MB GTX 460 running PhysX in support of my Kepler Titan on my x79 3930k system) when I monitored GPU load, the dedicated PhysX card only saw like 1-2% load, so I can't imagine anything newer or more capable is really required, unless PCIe bandwidth is an issue.
As has been mentioned, the oldest Consumer Nvidia GPU that is still supported by current drivers is Maxwell, 900 series I think.
If one is a huge fan of older titles and really wants to have accelerated PhysX turned on, one could probably get away with a cheap GTX950 off of eBay. There are a few listed with "Buy it Now" prices of $30. That is, if 1x or 4x Gen3 bandwdth does the trick.
If you have only a Gen4 1x or 4x chipset slot available, and PCIe bandwidth is a concern, it might make sense to try to maximize that bandwidth by getting the cheapest Nvidia Gen4 card one can find, which might have been why he opted for a 30-series, as that was the first Gen4 Nvidia series.
Maybe an alternative would also be a low end workstation model, like the RTX A400. Though searching for that, I see they seem to be going for the same amount of money as a 3050, and the 3050 definitely has more computing power (like 3x more) so if someone was uncertain how much computing power they needed, and wanted to maximize PCIe bandwidth in a low lane count Gen4 slot , I can definitely see how they might land on the 3050 as the best option.
Though I have to admit, I like the fact that the A400 is powered entirely off of the motherboard, so that might be my choice if I ever did anything like this (and a 950 was not an option due to PCIe bandwidth, because otherwise I'd of course opt for the older cheaper option)
Last edited:
- Joined
- Jan 1, 2014
- Messages
- 2,798
Yes. NV Control Panel shows CPU or NV card. You may be able to use AMD on some games/software if the developer added the ability. I don't know if there are any settings in AMD drivers for physx.
AMD CPU:
- Joined
- Jun 16, 2008
- Messages
- 7,284
So, run amd cpu or apu for physx support on 5xxx with old games?
jobert
[H]ard|Gawd
- Joined
- Dec 13, 2020
- Messages
- 1,790
What? Are you trying to say to use the CPU for hardware level physx? Have you not bothered to pay any attention at all? In the vast majority of those old games, the CPU cannot run those effects at a proper frame rate and that's clearly shown in all the benchmarks that have been posted about this. Plus there's some games where those effects don't even run on the CPU at all and you have to have an Nvidia GPU.
- Joined
- Jun 16, 2008
- Messages
- 7,284
Isn't AMDs physx support through the cpu and isn't hardware accelerated? While it won't be to nvidias level, the 5xxx series doesn't support it hardware side anymore anyways. Just wondering if it would be practical, considering itx doesn't have an extra slot for older cards to run in conjunction with the 5xxx just for physx.
Ur_Mom
Fully [H]
- Joined
- May 15, 2006
- Messages
- 20,903
I remember thinking the Ageia PhysX cards were going to usher in a whole new era of gaming along side GPU graphical updates. Graphics were excellent but adding in realistic physics and particles were going to add in a huge layer of realism. Some games took advantage of it, but once NVIDIA bought it and held onto it, it just kind of sat there without being the huge deal I thought it would be. I know modern GPU's can handle Havok and other physics models, but those old PhysX games seem to do better than many modern games. I really hope that it's a focus with newer technology as soon as the AI stuff runs it's course. There's so many times I've seen a game and it's super nice looking, ray tracing, everything on ultra. But, nothing really interacts with the environment. You can have fog, water, dirt, trees, etc., but it's just bleh when interacting, walking through it, stepping on something, etc.. It's almost like sprites on their own layer above the background. :/
Nothing to argue or add to it, just a bit of a sidetrack thought.... Which having this to be said is a killer testament to how amazing those "old" specific purpose PhysX and GPU's really were in comparison to a "new" general purpose CPU. With all the newer technology, higher speeds, higher core count, more memory, higher bus speeds, etc., they still can't hold a candle to those older cards. Really shows how amazing a GPU really is and how the CPU is just doing the basic things in a computer (although, working with old 8/16 bit homebrew stuff and older consoles, it's always been the case...). The CPU itself (not considering the integrated GPU in those that have them) couldn't handle modern graphics processing, physics like those, etc.. And they can really do a hell of a lot of things! Makes you realize why those data centers are using massive racks of GPU's rather than multiprocessor servers like in the past. Just thought it was cool as fuck these "old, deprecated, 32-bit" PhysX instructions and support still hold their own against modern CPU's.
Nothing to argue or add to it, just a bit of a sidetrack thought.... Which having this to be said is a killer testament to how amazing those "old" specific purpose PhysX and GPU's really were in comparison to a "new" general purpose CPU. With all the newer technology, higher speeds, higher core count, more memory, higher bus speeds, etc., they still can't hold a candle to those older cards. Really shows how amazing a GPU really is and how the CPU is just doing the basic things in a computer (although, working with old 8/16 bit homebrew stuff and older consoles, it's always been the case...). The CPU itself (not considering the integrated GPU in those that have them) couldn't handle modern graphics processing, physics like those, etc.. And they can really do a hell of a lot of things! Makes you realize why those data centers are using massive racks of GPU's rather than multiprocessor servers like in the past. Just thought it was cool as fuck these "old, deprecated, 32-bit" PhysX instructions and support still hold their own against modern CPU's.
jobert
[H]ard|Gawd
- Joined
- Dec 13, 2020
- Messages
- 1,790
Again, look at the benchmarks that are being shown on this topic. You cannot properly run the GPU level hardware physx in these old 32-bit games on the CPU and have it work right in every game. Performance is going to be all over the place and drop to unplayable levels in most titles and again, some titles are simply going to disable those effects if you try to run them on the CPU. Those effects in the older games like that were never set up to take advantage of a modern CPU like later versions of physx.
Not sure if there is really an opposition here, ai inference for realtime video game physic, that seem a possible use case for it:
https://developer.nvidia.com/blog/r...gital-twins-with-nvidia-omniverse-blueprints/
https://www.run.ai/guides/ai-open-source-projects/nvidia-modulus
https://github.com/NVIDIA/modulus
Kind of things that benefit to run on the edge instead of the cloud for latency and game are something being non-exact/that precise can be a plus or at least not an issue and error/hallucination as no cost. Maybe some dlss where your train on your game engine running super high quality physic models that would require a supercomputer vs what a very simple-fast that the game will ship with and get ai enhanced result would be possible (that something I am trying to do right now)
If they run single thread and no AVX2 type of usage on their CPU fallback version, it is a bit running those new CPU with a hand in their back too.
Last edited:
- Joined
- Jun 16, 2008
- Messages
- 7,284
I don't think you get what I am saying. I am not talking about running physx at all. I am talking about the alternative support that amd runs. I'm just asking. Physx is something I never thought much about. Between Nvidia and AMD cards, I never bothered to try and spot the difference.
3050 might be overkill but they can be found cheap.
The RTX A1000 would probably be the perfect card if it cost 1/4 of what the MSRP actually is.
https://www.pny.com/rtx-a1000?iscommercial=true
Don’t think of PhysX as its own thing, instead consider it a curated CUDA library for specific tasks.
https://developer.nvidia.com/physx-sdk
It just so happens that may aspects have some CPU/GPU cross compatibility and exclusivity’s. But the GPU components are CUDA based and as such Nvidia exclusive.
- Joined
- Oct 29, 2000
- Messages
- 42,357
You don't think a lower end and cheaper an RTX A400 will do the trick?
- Joined
- Oct 29, 2000
- Messages
- 42,357
Yeah, that's the part that doesn't make sense to me.
PhysX is just a custom API that runs on top of CUDA compute.
It doesn't make any sense to me why they would pull the plug on it, when 50-series GPU's certainly can still run CUDA.
There is absolutely nothing about the 50-series hardware that should prevent running PhysX on it. Heck - if you knew what you were doing - you could probably hack the drivers to allow it.
So why exclude a software feature that would be nearly free to include? Just a desire to support less stuff?
That why PhysX still run perfectly fine on a 50-series GPU, just need to be compiled with the 64 bits version.
Why they decided around ~2012 to kill cuda 32 bits and took around 10 years to fully get rid of it (it went, around 2013 telling people to stop to use it, around 2014 to stop making compiler and including it in the sdk, etc...)
Is there nothing in the 50-series hardware that prevent running 32-bits CUDA (like 8 bits code do not run on 16 bits cpu quite often, 32 bits code do not run on apple cpu anymore, etc... they need to use virtual mode-emulation), it is possible to make hardware gain to be made in MMU-ALU to support a single set of instruction rather than 2.
The question can become intel-apple make effort to support and still be able to run in non-native mode older bits size pointer and instruction, slow cuda seem useless, could run on the cpu instead if the emulation cost is big, that tech could take the size they saved and more importantly about nothing would use Cuda 32 of any importance, what are the cuda application not make to take advantage of more than 4gb of ram ? There could be virtually nothing of importance still used that still running on 32 bits after a full decade of clear instruction of stopping using it and really hard to use it.
That why when they said in 2022 that Lovelace was the last to support 32 bits cuda, we did not hear about it at all. Game made ~2013 are a strange case of something that use CUDA but would have never been updated since and people that can't for some reason do it (publisher not existing, caring, and so on)
I read that as A4000, yeah the A400 would probably cut it just fine.
Yeah there is no physical limitation, it's purely software.
But in all fairness, Nvidia announced the start of their 32-bit CUDA depreciations back in November of 2013.
https://www.linuxcompatible.org/sto...-32bit-linuxx86-cuda-toolkit-and-cuda-driver/
It's just taken a while for them to finally end it all.
I suspect that what ever hardware version comes next, will be different enough of an architecture that support for it would be impossible, so this is the Band-Aid moment if you will, so when the next gen gets here and physically can't nobodies shocked.
How do we know such things ? Pretty much all cpu around have physical limitation to run natively some older-lower bits program of the past.
- Joined
- Oct 29, 2000
- Messages
- 42,357
Huh. I tried to play with CUDA under Linux back in 2013 when I first got my Titan, but I never got very far.
I wonder how hard it would be to port it to 64 bit instructions. I imagine it might be a little bit more complicated than just recompiling it for the different instruction set, but I can't imagine it would be that big of a deal.
If they are going to abandon it anyway, they should quit being dickheads and just open source it, and let the community take it from there, sort of like what happened with EAX/Alchemy.
This is the problem with proprietary API's on the PC platform, and why they should all die.
Issue could be more the rest of the game than the CUDA part.
You would want to port the game to 64 bits, using a 64bits->32 bridge to talk to the CUDA could destroy the purpose (which is purely speed here), how hard it is to port the game to 64 bits, depends how much pointers arithmetic-optimisation that assumed 32 bits, did you use third part library you do not own the code (non-closed source for what you buy got more and more popular, early 2010 was not like now) that did not had a 32 bits version, does your game publisher still exist and if he do, any contract that make it illegal for you to update your game you made.
When I had to do old 32->64 bits giant project transition, the issue was always old libraries, more than the code base you know a bit and fully control, that said games can be more intense than usual in memory-pointer arithmetic that assume 32 bits than the average program.
Video game turning open source to let people compile it in 64 would work obviously, but not sure they are being dickheads about it.
But we also see how much benefit there can be in the incentive with proprietary APIs (the speed it can go), if Cuda endup curing a bunch of cancer in the next 20 years, playing batman without physx set to on on a new gpu because the game dev do not want to update it will look silly, that why we always need to balance the pro and cons.
- Joined
- Oct 29, 2000
- Messages
- 42,357
Is there really that much CUDA has done for us over the years that OpenCL couldn't though? Sure, maybe it wouldn't have been quite as convenient, and maybe it would have been slightly slower, but it also would ahve been more flexible from a hardware perspective and probably allowed for that to make up for the efficiency losses. I feel like these compute problems would have been solved one way or another with open platforms if the proprietary shit hadn't sucked all of the oxygen out of the room.
..but because of the proprietary nature of the likes of CUDA and PhysX they just wind up overcharging for them once they lock their customers in, and turning them into dead ends for users once they no longer feel it is worth supporting it.
...and thats not to mention the bullshit they pulled with driver lockouts preventing AMD users from using a secondary Nvidia card for PhysX. For that - IMHO - they ought to have been broken apart by the DOJ. That is classic illegal market manipulation.
Heck, Ageia's PhysX existed before Nvidia turned it into shit. Was it open source? No, but it was nowhere near the proprietary mess it became under Nvidias stewardship, where - once CPU's got faster and got more cores and were able to do the work on the CPU, they just refused to support and update the CPU code because it didn't benefit their owjn prporietary strangelehold on the technology.
Nvidia has a long history of acquiring or creating propiertary standards and using them toeither lock users in to their own compatible hardware, or lock users out of the competition, and that is straigh tup illegal by the laws that area lready on the books, but for some reason we as a country choose to turn a blind eye.
Nvida should have been Ma Bell:ed already back in the mid to late 2000's.
Last edited:
Will never know if the openCL world and the destruction to the giant incentive created by CUDA we will be closer or not to self-drivings cars, but I would not talk as if one was obviously better than the other for humanity. Everything pretty much as cost and we only have compromise, no solutions.
Diminishing the expected financial reward associate with really had work, always can have consequence, we only not see the counterfactual when we do.
Diminishing the expected financial reward associate with really had work, always can have consequence, we only not see the counterfactual when we do.
The level of cuda work on everything as been quite good too, people make way too much of the 32 bits things, nothing of any matter run on this that we know of (if they do, not heard of the issues with recompile 64 bits)
- Joined
- Oct 29, 2000
- Messages
- 42,357
Maybe I don't fully understand the architecture (actually, I'm pretty sure I don't) but I thought the API was just a abstracted pipe that received computation requests in numbers, and spit back the results. A 64bit system should be able to compute 32bit numbers just fine, the biggest difference should be instruction set and memory addressing, which ought to be completely abstracted by the API.
Seems like it ought to be faily trivial to modularly just maintain the same application interface API, while reworking the back end completely. Heck, if they could do it while moving from whatever MIPS64+custom architecture Ageia used to x87 code for CPU use and later and 32bit CUDA, it seems like 64bit CUDA should be well within the realm of possibility.
Many of those old games really didn't expect to run at more than 60fps either, so they ahve 16ms to get the job done befor ehte next frame. Thats a bloody eternity to do the calculations needed on modern hardware. I can';t imagine that 32bit to 64bit translation would slow it down much, Maybe on early to mid 2000's hardware, but certainly not today.
This is in all likelyhood not worth it for Nvidia, but I'm pretty sure that if they released some combination of the x87 CPU codebase and the 32bit CUDA code some data scientist who understands CUDA coding, who is also a gamer would take this on as a community passion project, and we'd get something like Alchemy for EAX.
It's not like the code has any real value for them anymore. They could release it like ID Software released the Quake code, and allow for exciting commuity projects.
Number is different than instruction.
When we talk about 32 bits vs 64 bits cpu and compilation, it is not the support for say int64 that change, 64-128 bits operation on 16 bits cpu can work, it is usually the instruction and the memory adress size that change like you say. You can have your ALU an a MMU that detect which is which and support both, but that not absent of a transistor cost necessarily.
As API had any difference between 32 and 64 ? I did not find one, there is some extra instruction only to the 64 bits (like always) for example the ability to use the memory of multiple gpu at the same time, but the call should stay the same as I can see.
Lot of CPU that only run 64 bits (take apple since M1 and their OS since 2019) retro support 32 bits application but that involve a lot of works, virtualization mode and performance impact, one big difference there is a lot of valuable code and critical code to be run, we do not have a single example of cuda 32 bits program that matter even just a little bit that cannot run on a 5000 series GPU computer today, to be worth the trouble if there is one.
You mean the ptx to machine language compiler ? (cuda to ptx is already open source), if in any part you endup needing access to game source and to recompile it ... , may has well just compile it in 64 bits no ? if it does just just in time and working with the original PTX that nice, but that seem more complicated than just recompiling the 3-4 game that would matter themselve.
Anything that you can do in CUDA, you can do in some combination of OpenML, OpenCL, OpenMP, OpenACC, and the tinkering of your development environment to work with those in some combination of C++ and Python; it would just be a matter of putting in the work.
That's where CUDA's big selling point comes in, you can have your team spend months building that environment, and the libraries to do the work, then pour over those libraries they built to painstakingly optimize them so they can finally get to doing the task you wanted them to do that they had to build the system for, or they can use CUDA.
Chances are that whatever you want to be doing, there is already a well-maintained, well-documented, previously vetted, and verified CUDA library to do it, and it may even be Open Source.
When buying into the Nvidia Hardware/Software ecosystem at that level, it's because they have already put in the work, and that work has been verified, qualified, and optimized to a very high degree.
It is primarily a Time/Money/Resources issue. Do you have the Money to afford the Time and Resources needed to do it without Nvidia?
Flogger23m
Fully [H]
- Joined
- Jun 19, 2009
- Messages
- 16,865
I assume a software fix would not be possible?
Translation layers are always an option, somebody just needs to build it.
Though I suspect that anybody who is truly bothered about losing 32-bit CUDA and PhysX will either just keep an old machine or eventually break down and get a sub-70w, non-5000 series workstation card to slot into their gaming PC.
Nvidia keeps support for the workstation cards going for a LONG time so the chances of loosing drivers for them any time soon is unlikely.
- Joined
- Dec 19, 2005
- Messages
- 17,807
- Joined
- Dec 19, 2005
- Messages
- 17,807
- Joined
- Oct 29, 2000
- Messages
- 42,357
I did some more thinking on this. I think the oldest and cheapest card that ought to work with current unified Nvidia drivers (572.60) is the old Maxwell based Quadro K620.
There is one on eBay now for $12.
Back when I ran dedicated PhysX on my old GTX460 (768MB version) I only ever saw 1-2% GPU loads on that card. Now the K620 and the GTX460 are very different architectures from different times, but I think the K620 is only 10% or so behid the GTX460 from a raw number crunching ability, so, no guarantees, but it will probably work. (provided Nvidia enabled PhysX in those old Quadros. I think they did, but I am not 100% sure. For $12, it's not too much to just test it
Who knows how much longer it will be supported in the latest unified driver, but if you need PhysX acceleration right now, that seems like an affordable option.
Last edited:
- Joined
- Oct 29, 2000
- Messages
- 42,357
Yeah. 32 bit OpenCL was dropped as well as 32bit CUDA.
For most compute workloads, I'd argue, it is time to move beyond 32bit. I'd suggest that PassMark is no longer measuring significant metrics with this compute benchmark.
PhysX is the lone feature I am sad to see go.
I am really disappointed about this because my kids and I have a tradition of playing a Borderlands co-op game every Christmas after dinner and the subsequent food coma. We have done this for probably the last 8 years and we just loved the PhysX effects in the games. I couldn't tell you how many times I had to call it quits because the sun was coming up.I did some more thinking on this. I think the oldest and cheapest card that ought to work with current unified Nvidia drivers (572.60) is the old Maxwell based Quadro K620.
There is one on eBay now for $12.
Back when I ran dedicated PhysX on my old GTX460 (768MB version) I only ever saw 1-2% GPU loads on that card. Now the K620 and the GTX460 are very different architectures from different times, but I think the K620 is only 10% or so behid the GTX460 from a raw number crunching ability, so, no guarantees, but it will probably work. (provided Nvidia enabled PhysX in those old Quadros. I think they did, but I am not 100% sure. For $12, it's not too much to just test it
Who knows how much longer it will be supported in the latest unified driver, but if you need PhysX acceleration right now, that seems like an affordable option.
I really don't think it is too bad for me and my kids because I don't think we are going to be upgrading soon from my RTX3070 and their RTX360ti's.
- Joined
- Oct 29, 2000
- Messages
- 42,357
Well, luckily getting an older, still supported dedicated PhysX card isn't too expensive!
Just have to have a motherboard that has a slot you can stick it into. Preferably one that does not degrade the primary slot to 8x.
StoleMyOwnCar
2[H]4U
- Joined
- Sep 30, 2013
- Messages
- 3,526
Sorry but... I think that's some copium. At that point you're pretty much close to running SLI, unless you specifically look for a very small form factor and TDP GPU to run in that bottom slot. But I reckon most of us are just going to repurpose an old card if needed. Which is probably not going to be small.
This gen is just such a shitshow. Every single post, I feel like we keep finding more and more ways in which they decided to skimp out.
- ROPs? "Yeah it's just 0.5% of cards affected okay?" *Under breath: damn, hoped they wouldn't notice*
- Connector and PCB? "Smallest ever, look our cooling is so good" *Turns into fire hazard when exposed to 600W, surprised Pikachu face* Almost like it was just made to save money
- PhysX? "Yeah we're just ending support for that." You know, despite this gen having no large improvements in any way shape or form that make it worth it actually worth cutting that off. Pretty much did it just to save costs I guess... goodbye to some of those older titles that you might have wanted to run.
- Supply? What supply?
- Performance uplift? You love framegen right?
At this point hilariously buying the 4090 might have been the best investment I've made. Because it's pretty clear this isn't going anywhere good, and the 4090 is still better than the 5080 (which is an unusually bad gen to gen improvement). The 5090 is about a 41% improvement (and some more VRAM) in Stable Diffusion, which might be the only reason I even consider getting it at this point, but that was only at flat MSRP. At current price and availability (and hell, reliability), no fucking way in hell. For 2k they definitely also should have included more VRAM but they clearly don't want to encroach upon their precious AI space... you know, the one where we're lifting all of our "gen gains" from.
Also as far as frame gen, I have attempted turning it on briefly in a few games and it was pretty much unbearable. I didn't expect to notice a little extra input delay, but it was massive and noticeable to me (maybe because my base frame rates weren't great). The game also looked like shit because I wasn't actually reaching great FPS without frame gen. Wtf is the point, if I need it then I can't use it.... if I don't need it then I'm not going to use it anyway.
At this point I'm almost ready to tell AMD to just take my fucking money if they just make anything even half decent with the 9070/XT, just to shove it to Nvidia. This gen pisses me off that much. I've rarely been actually angry about PC part releases. I've ranted and rambled. I thought the 4000 series was also a bit of a shitshow. But this is something else.
Last edited:
- Joined
- Oct 29, 2000
- Messages
- 42,357
Don't get me wrong. I'm disappointed about this as well. This is the problem with closed source bullshit. It was never going to make sense for Nvidia to keep maintaining 32bit CUDA forever, and it is technically EOL by now.
There are some 40 titles that are affected in all, and the newest one is (Borderlands: The Pre-Sequel) is over a decade old at this point.
It was going to come to an end at some point. I had hoped that when that time came, CPU's would be able to handle it. And CPU's - especially modern ones with gobs of mostly idle cores - can handle those calculations. That's what the physics engines built into most modern engines do, but because of the shitty state of CPU PhysX and it's terrible legacy x87 code, it causes games to choke on it.
It sucks. But at least there is a cheap workaround. Get a $12 Quadro K620 off of eBay and stick it in that bottom x2 to x4 slot off the chipset. If you don't have an older GPU kicking around you can use - that is - (but m,any of our old GPU's are going to be higher power models, and then you might run into PSU issues and needing more power connectors than you have handy)
Totally agree. I'm not defending this launch. It is awful. The worst GPU launch in recent memory. Probably even worse than the awful GeForce FX launch in 2003.
The only worse Nvidia incident I can think of was the solder substrate issue on the 7000-9000 series cards back in like 2008 or so. when close to 40% of everything Nvidia sold in that era was going to die from this issue at one point or another, and they did nothing to cover the losses of AIB's or end users. AIB's ate warranty claims from users who were still under warranty, and all the users who were out of their warranty were just SOL.
I'm not going to lie though, despite all of the issues I still want a GPU faster than the 4090 as a 4090 continues to be inadequate for my performance requirements. I wish there were other options out there, but the 5090 is the only option. I hope I can find one at some point in the next few months that does not suffer any of the ROP issues. I have some concerns about the connector, but I can be very careful with it...
Ideally Nvidia would recall anything with that connector and replace it with something else, but you know they are never going to do that. Company policy seems to be to just ignore any and all problems until people stop talking about them, and stiff AIB's and end users. And as long as there is no alternative in the high end, they will likely constitute to get away with it. The only way they will ever be held to account is by there being competition, and people ditching them.
Ideally someone other than Nvidia would provide a product that gives me a notable upgrade over my 4090, but that likely won't happen for a long time.
OpenCL is fully open source and has the exact same issue it seems... CPU supporting 32 bits native or not is not something open source fix by itself. Or maybe I have not followed the conversation and it shifted about something else ?
If you mean the application that use openCL or Cuda being close source, yes.... but at some point, I am not sure how they would finance software development if everything all the way up the stack is always open, for a lot of stuff you cannot just sell service over the app (by making them hard to use).
Been a long time when you compiled you had a warning in your compiler telling you unsupported, deprecated, etc... Support for it in term of creating a new program became not recommanded and less and less supported around a decade ago, the ability to run old binary is something different but was also I can imagine not expected to be forever, that why they told people to not make anything expected to have a life with it.
Last edited:
People that bought RTX 40xx series must feel rather cozy
View: https://www.youtube.com/watch?v=h4w_aObRzCc
View: https://www.youtube.com/watch?v=h4w_aObRzCc
- Joined
- Mar 11, 2007
- Messages
- 1,006
https://www.phoronix.com/news/ZLUDA-Q1-2025
"The ZLUDA project hopes to have PyTorch up and running on it this year along with eyeing 32-bit PhysX support since NVIDIA has dropped support upstream for the 32-bit PhysX libraries with the recent RTX 50 Blackwell launch."
"PhysX on the modern ZLUDA hasn't been a big focus but since NVIDIA dropping support for 32-bit PhysX with the GeForce RTX 50 / Blackwell GPUs, that thinking changed slightly. There's a belief that ZLUDA can 'rescue' the 32-bit PhysX support to get it working for both AMD and NVIDIA GPUs."
"The ZLUDA project hopes to have PyTorch up and running on it this year along with eyeing 32-bit PhysX support since NVIDIA has dropped support upstream for the 32-bit PhysX libraries with the recent RTX 50 Blackwell launch."
"PhysX on the modern ZLUDA hasn't been a big focus but since NVIDIA dropping support for 32-bit PhysX with the GeForce RTX 50 / Blackwell GPUs, that thinking changed slightly. There's a belief that ZLUDA can 'rescue' the 32-bit PhysX support to get it working for both AMD and NVIDIA GPUs."