Nvidia working on similar memory acess to AMD

[thechinchinsong]

51 minutes ago, RejZoR said:

AMD tends to do this a lot. Even feature sets on Radeon cards are mostly limited to RX 5000 series. If you have Vega, you just don't get it. Or if you have Polaris... Where NVIDIA tends to add features that go quite few series back. Currently I believe they go as far as Maxwell 2 (GTX 900 series) for the most part and the rest is hardware limited (like variable shaders that are only supported on Turing or newer because you just can't do it on older cards).

I think some of the limits that AMD puts on their graphics cards features aren't actually intentional limitations put in by the company like this one is. There is a reason that certain features are only enabled on RDNA, since it's their brand new architecture while everything before RX-5000 series is GCN based. Iterating on such an old architecture can only go so far with the features you are able to add/enable.

[StDragon]

2 hours ago, LAwLz said:

I mentioned this in the Ryzen 5000 series review threads:

 

Some things to keep in mind here:

1) Smart Access Memory, which seems to just be AMD branding for resizable PCIe BAR, is not something AMD came up with. It's a PCIe specification.

2) Resizable PCIe BAR is a standard in the PCIe specs. Any member of PCI-SIG can implement it if they want, including AMD, Intel, Nvidia, ARM, etc.

3) Resizable BAR has not been supported in Windows until fairly recently. That's one of the reasons why we haven't seen it before.

4) AMD seems to be artificially limiting it to their newest CPUs, motherboards and graphics card in order to sell more products. I can't think of any reason why they couldn't enable it on older products if they wanted to. It is fairly telling that AMD even has a requirement of a 500 series motherboard, despite the motherboard barely having anything to do with PCIe since that is controlled by the CPU.

 

 

Let's hope Nvidia enables it for all CPUs.

Would be a bit funny if Nvidia cards had better feature compatibility with AMD CPUs than AMD's own GPUs has.

3% to 8% performance gain with PCIe BAR is nothing to scoff at. That's HUGE! Why wouldn't this have been implemented years ago if all that was required is a driver update? It was already competitive between AMD and Nvidia. So the question is "why now?". Moreover, why not then?

 

This whole topic stinks. Something is amiss to have left that much performance on the table for so long. I really look forward to a GN in-depth analysis. 

 

 

[LAwLz]

6 hours ago, StDragon said:

3% to 8% performance gain with PCIe BAR is nothing to scoff at. That's HUGE! Why wouldn't this have been implemented years ago if all that was required is a driver update? It was already competitive between AMD and Nvidia. So the question is "why now?". Moreover, why not then?

 

This whole topic stinks. Something is amiss to have left that much performance on the table for so long. I really look forward to a GN in-depth analysis. 

Why not implement it years ago? Because windows only started supporting it like lady year, and it takes time to implement these things. 

We don't know how big of an impact this has either. Sure, it was like 3 to 8 percent in the games AMD tested but we don't know what it will be like for most games. AMD might have picked the best games for this and only got 3-8%. Chances are it turns out to be like 5% performance uplift but that's rather small. Nvidia might have been focusing on other stuff instead that could yield better returns. But now that AMD are pushing it in their marketing Nvidia we're forced to focus on it. 

 

I don't think it's strange that a new feature (in Windows) that only provides a few percent performance increase is low priority. 

 

Why now? Because AMD got handed a feature they could use to try and sell more CPUs and motherboards. They haven't really had the chance to market it this way before. 

 

Why is Nvidia focusing on it now? Because AMD marketed it and Nvidia wants to crash AMD's party a little bit. 

[Sauron]

On 11/13/2020 at 5:57 PM, GhostRoadieBL said:

Nvidia acting like school kids on the playground "well my toy can do everything yours can" 

Well, if it can, all the better

[xAcid9]

Guess what Nvidia going to name it.
I say, RTX Memory CUDA A.I.

[SFFDesigns]

53 minutes ago, xAcid9 said:

Guess what Nvidia going to name it.
I say, RTX Memory CUDA A.I.

RTX Memory doesn't actually sound that unlikely.

 

They may just call it resizable BAR though

[RejZoR]

They shall call it NFINITE MEMORY ACCESS !

[igormp]

That has been available on Linux since 2017, but depends on the motherboard to have this option or at least report it to the OS. Support is finicky so I guess that's why AMD is requiring both new CPUs and x570 motherboards, since they can easily validate and mandate how the bios works. Could be just a marketing scheme as well tho.

 

Fun fact, nvidia already has support for it on linux and that is well documented under their gpudirect-rdma docs (amd should have something similar, but I'm not so sure).

10 hours ago, LAwLz said:

despite the motherboard barely having anything to do with PCIe since that is controlled by the CPU.

The motherboard has to report that above 4g mmio support is enabled for the OS.

[LAwLz]

17 hours ago, igormp said:

The motherboard has to report that above 4g mmio support is enabled for the OS.

Yes but that's not a feature that's new to the 500 series boards. Pretty sure even my X370 board has it. But AMD says that their "smart access memory" will not work unless you have:

1) A 500 series chipset motherboard. 

2) A 5000 series CPU. 

3) An RX6000 series GPU. 

 

If you don't have all three, you won't get smart access memory. Seems to be like AMD are blocking it from working on older platforms, rather than there being some technical reason for it. 

[StDragon]

14 minutes ago, LAwLz said:

Yes but that's not a feature that's new to the 500 series boards. Pretty sure even my X370 board has it. But AMD says that their "smart access memory" will not work unless you have:

1) A 500 series chipset motherboard. 

2) A 5000 series CPU. 

3) An RX6000 series GPU. 

 

If you don't have all three, you won't get smart access memory. Seems to be like AMD are blocking it from working on older platforms, rather than there being some technical reason for it. 

The question I have is this: Is that a requirement for a guarantee of feature availability, or a hard requirement for the feature to be "unlocked" and available? There is a difference.

 

I could see 1) and 3) being a requirement due to both vendor BIOS and AMD GPU driver support. But there should (in theory) be no reason to exclude Zen 2 as @Hymenopus_Coronatus so correctly pointed out that both Zen2 and Zen3 share the same IO die.

[leadeater]

56 minutes ago, LAwLz said:

If you don't have all three, you won't get smart access memory. Seems to be like AMD are blocking it from working on older platforms, rather than there being some technical reason for it. 

It's an optional feature within the PCIe spec, like a lot of optional features it may not have been implemented or it's buggy so won't be enabled. When Nvidia comes out with their equivalent and it works on these other platforms it shouldn't take long for AMD to follow, competition wins out.

 

But I wouldn't just assume AMD are arbitrarily blocking support for it, getting a feature like that fully supported and consumer ready is a lot of work and we're talking about chipsets that are out of general support or about to be.

 

Just have a look at the issues Linus has had in the past with PCIe devices on his server videos and his stupid experiments he's done, motherboards have been the primary issue and has required to change them out on a number of occasions or get custom firmware fixes to resolve his issue.

[Guest]

According to GamersNexus Nvidia's implementation will only be for Ampere cards on Z490 and some AMD boards too. It won't be universal at launch either.  

[LAwLz]

1 hour ago, StDragon said:

The question I have is this: Is that a requirement for a guarantee of feature availability, or a hard requirement for the feature to be "unlocked" and available? There is a difference.

 

I could see 1) and 3) being a requirement due to both vendor BIOS and AMD GPU driver support. But there should (in theory) be no reason to exclude Zen 2 as @Hymenopus_Coronatus so correctly pointed out that both Zen2 and Zen3 share the same IO die.

AMD's own website lists all three as requirements for Smart Access Memory to work.

Since the RX 6000 series isn't out yet we don't have tests to check if it is required or not. The official word from AMD is that it you need a 500 series motherboard, 5000 series processor and 6000 series GPU.

 

1 hour ago, leadeater said:

It's an optional feature within the PCIe spec, like a lot of optional features it may not have been implemented or it's buggy so won't be enabled. When Nvidia comes out with their equivalent and it works on these other platforms it shouldn't take long for AMD to follow, competition wins out.

It's an optional feature, yes, but it is supported on far more than what AMD has said Smart Memory Access works on.

According to one of the AMD GPU driver developers for X.org, they have had it working on Linux for years already, and I assume they do not have time machines. So there is no hardware limitation for it in currently released hardware such as Zen2.

 

 

1 hour ago, leadeater said:

But I wouldn't just assume AMD are arbitrarily blocking support for it, getting a feature like that fully supported and consumer ready is a lot of work and we're talking about chipsets that are out of general support or about to be.

Well I assume they are. But I guess time will tell. If Nvidia can get it working on something AMD aren't supporting then we know for sure that it is a limitation imposed by AMD, not a hardware limitation.

[StDragon]

21 minutes ago, LAwLz said:

 

According to one of the AMD GPU driver developers for X.org, they have had it working on Linux for years already, and I assume they do not have time machines. So there is no hardware limitation for it in currently released hardware such as Zen2.

Would be real interesting to see these drivers compared with the ones without the feature enabled on Linux running Proton for Steam. Some comparative bench marking.

[igormp]

4 hours ago, LAwLz said:

Seems to be like AMD are blocking it from working on older platforms, rather than there being some technical reason for it. 

As I said before, I believe it's either a matter of validation or just a marketing scheme.

 

1 hour ago, LAwLz said:

So there is no hardware limitation for it in currently released hardware such as Zen2.

Some first gen AM4 boards (b350 and x370) had no option for above 4g decoding, boards that are still used by some people with Zen 2, so there's that. I'm not sure about b450/x470, but I wouldn't doubt that some manufacturers left it disabled by default without some option to turn it on.

[Loote]

Let's look at it from this perspective:
1. It's a fairly recent thing
2. Fast NVME storage is getting bigger and more popular every year

3. As was mentioned above, Windows only started supporting it a year ago
4. Nvidia would have to work with Intel to make it work

some of my thoughts:
-It may very well be a case of AMD not supporting the feature, but not blocking it either on platforms other than 570(+ RX 6xxx and Zen3), then you add marketing and say it like it's unique to 570, while only them supporting it is

-It made no sense to make it work on GCN/Vega when RDNA was coming, it isn't hard to think of some reason why it only appeared in RDNA2 too seeing how the firs iteration supposedly even had hardware bugs.

-Seeing how Nvidia had to work with Intel(and AMD, if they didn't do it for RDNA) to make it happen and some rumors about them getting lenient, I think that assumption that they put a low priority on it which caused the delay on their part isn't too far fetched.

 

Honestly, if it really is AMD pretending like it's impossible on older platforms to make the new ones sell more, I'd be quite annoyed, hopefully in a year both them and Nvidia will have a quite big base of supported hardware.

[igormp]

4 hours ago, igormp said:

I'm not sure about b450/x470

Actually, I found out that the asus prime x470 has no way to enable above 4g iommu, so yeah, AMD probably made it a requirement (and to be enabled by default) just recently.

 

On other news, seems like there's a beta bios for gigabyte boards that allow it (and even was renamed as "resizable bar support").

[igormp]

https://www.pcgamer.com/amd-smart-access-memory-not-proprietary-promise/

 

Quote

AMD has now released a statement explaining that yes, Smart Access Memory is based on the Resizable BAR PCIe, and yes, it can work with other hardware, but it's currently only been working on validating it with the Ryzen 5000 and Radeon RX 6000 pairings.

 So it seems that it was just a matter of validation anyway.

[LAwLz]

35 minutes ago, igormp said:

https://www.pcgamer.com/amd-smart-access-memory-not-proprietary-promise/

 

 So it seems that it was just a matter of validation anyway.

Lol, good one AMD.

 

AMD: Look at our amazing new technology that we have developed. AMD is the only company that can deliver processors and GPUs, and that has enabled us to make the Ryzen 5000 series, when you got a 500 series motherboard, to work better with our RX6000 series graphics cards. This is a platform performance innovation! Remember, it requires an AMD 5000 series CPU, 500 series motherboard and 6000 series GPU.

 

Nvidia: We got it working on other stuff too.

 

Customers: Wait, so AMD are blocking it from working on older platforms?

 

AMD: Ehm, ignore what we said before. Yeah it's just PCIe resizable BAR. We will make it work on our old products too.

 

 

I wonder if AMD would have just kept it locked to their latest platforms if Nvidia hadn't come out and said it would work on other platforms too.

If they had plans to support it on older platforms too then why didn't they mention that anywhere? Why did they go on to talk about how it was only possible because they make both CPUs and GPUs? Why did they explicitly put up marketing material where it clearly states that Smart Access Memory require a 5000 series CPU, a 500 series motherboard and a 6000 series GPU?

 

I think AMD tried to push sales of their newer products, get some hype for being a CPU and GPU company, and then were caught with their pants down and are now trying to save face.

[StDragon]

50 minutes ago, igormp said:

https://www.pcgamer.com/amd-smart-access-memory-not-proprietary-promise/

 

 So it seems that it was just a matter of validation anyway.

Well, that confirms it.

 

And all along I was thinking AMD might have implemented the Intel equivalent of DDIO (Data Direct I/O Technology) in Zen3. Nope!

[Fnige]

On 11/15/2020 at 5:26 AM, RejZoR said:

They shall call it NFINITE MEMORY ACCESS !

This is NVidia we are talking about. Not AMD. they didnt decide to go with calling something rage mode after intoxicating themselves at their local bat

[tim0901]

13 hours ago, LAwLz said:

I wonder if AMD would have just kept it locked to their latest platforms if Nvidia hadn't come out and said it would work on other platforms too.

If they had plans to support it on older platforms too then why didn't they mention that anywhere? Why did they go on to talk about how it was only possible because they make both CPUs and GPUs? Why did they explicitly put up marketing material where it clearly states that Smart Access Memory require a 5000 series CPU, a 500 series motherboard and a 6000 series GPU?

Honestly wouldn't surprise me if they had wanted to. They know they're in the lead now and are trying to take advantage of that position with anti-consumer tactics. Just look at the B450 kerfuffle - they only rolled back on their decision then because consumer response was "greater than expected", so in the future they just aren't going to make such promises to remove themselves from any obligation to be consumer-friendly in the first place.

 

The sneaky pricing of the Ryzen 5000 series is another example of this - the 5600X is "only" a $50 increase... compared to the 3600X that nobody bought anywhere near MSRP. Same with the 5800X vs the 3800X. In the real world, the 5600X and 5800X are a $100 markup over the comparable previous gen chip that people actually bought, AMD were just deceptive about how they showed that information.

 

The community seems to be a lot less critical of AMD when it tries things like this than they are of Nvidia or Intel, which imo is ridiculous. AMD are a company just like the others and they won't hesitate to screw you over in the name of profit. My prediction is that Ryzen 6000 will have even more of this - I certainly wouldn't expect to see a commitment to 4 generations of CPUs with AM5.

 

IMO Nvidia was probably preparing this in the background for their professional cards (like the upcoming A6000), but decided to roll it out to their consumer cards as well when AMD announced SMA.

[spartaman64]

7 minutes ago, tim0901 said:

Honestly wouldn't surprise me if they had wanted to. They know they're in the lead now and are trying to take advantage of that position with anti-consumer tactics. Just look at the B450 kerfuffle - they only rolled back on their decision then because consumer response was "greater than expected", so in the future they just aren't going to make such promises to remove themselves from any obligation to be consumer-friendly in the first place.

 

The sneaky pricing of the Ryzen 5000 series is another example of this - the 5600X is "only" a $50 increase... compared to the 3600X that nobody bought anywhere near MSRP. Same with the 5800X vs the 3800X. In the real world, the 5600X and 5800X are a $100 markup over the comparable previous gen chip that people actually bought, AMD were just deceptive about how they showed that information.

 

The community seems to be a lot less critical of AMD when it tries things like this than they are of Nvidia or Intel, which imo is ridiculous. AMD are a company just like the others and they won't hesitate to screw you over in the name of profit. My prediction is that Ryzen 6000 will have even more of this - I certainly wouldn't expect to see a commitment to 4 generations of CPUs with AM5.

 

IMO Nvidia was probably preparing this in the background for their professional cards (like the upcoming A6000), but decided to roll it out to their consumer cards as well when AMD announced SMA.

i mean doesnt this post sort of invalidate that point since people are complaining about this and lots of people complained about the 450 chips not supporting zen 3. which with the compromises they have to make to make that work they had a point. the market for b450 boards is going to be a complete mess when the bios for zen 3 launches and amd is going to take the blame for it even though its the board markers that decided to cheap out on flash chips

[tim0901]

4 minutes ago, spartaman64 said:

i mean doesnt this post sort of invalidate that point since people are complaining about this

This post that's a tiny thread compared to the huge rage thread that was going on here when the Nvidia 'paper launch' 3080 was released? Or compared to the rage thread that went on when Intel removed memory overclocking from non z-series boards? Yes, the community as a whole has an AMD bias rn, because they're happy that they've come in and created some competition. I also predict that the outrage towards the lack of RX 6000 cards will be smaller than vs the RTX cards, even accounting for the fact that people are used to this shit by now.

 

This forum is also far more balanced than the tech community as a whole. Reddit is far worse at this.

 

(Can I also point to them outright stealing Intel and Nvidia's naming schemes? People complained that the motherboard names were confusing, they didn't blame AMD for trying to 1-up Intel and causing that confusion in the first place. )

 

4 minutes ago, spartaman64 said:

and lots of people complained about the 450 chips not supporting zen 3. which with the compromises they have to make to make that work they had a point. the market for b450 boards is going to be a complete mess when the bios for zen 3 launches and amd is going to take the blame for it even though its the board markers that decided to cheap out on flash chips

Maybe AMD shouldn't have made promises that they couldn't keep. It was their fault that they ever were in that position so start with, because they promised such long-term support for motherboards that clearly weren't engineered to support it because of the small BIOS chips (and allowed their motherboard vendors to make similar promisies of their own.) 

 

And yes, I agree that the current solution is messy and horrible. Imo not going back on their decision would have been the better option for them as a company (although not for the consumer). At least that way they wouldn't have made the whole thing look even more anti-consumer than it already did by announcing they were bowing to the consumer rage and changing their mind.