AMD Ryzen 3000 and Older Zen Chips Don't Support SAM Due to Hardware Limitation, Intel Chips Since Haswell Support it

[Outback Bronze]

Lucky I've still got a few Haswell chips lying around...

 

[Halo3Addict]

Sooo what if I own a 5600x paired with an RTX 3080? Are you going to play nice then, AMD?

 

[TechLurker]

This doesn't make sense; Zen 2 shares the exact same I/O die with Zen 3. The only thing that changed is the CCX/CCD, but the I/O die handles the communication with the PCIe lanes anyway. This seems like either some half-arsed Intel hit or AMD just not wanting to officially bring SAM to Zen 2.

That said, Intel isn't likely to support Resizable BAR on older CPUs. Like AMD, they want to only sell their newest. For that matter, Re-BAR support seems to be locked to their Z400-series Chipset too, so they don't even target the mid-range like AMD does with their B400-series.

 

[z1n0x]

I'm sure support is coming to Skylake and Turing, momentarily. /s
Do you honestly believe support on Intel/Nvidia will come to anything other than the latest parts?

 

[Colddecked]

I know what I have to do. Replace my ryzen 3600 with an i5 10600k and z490 motherboard.
I will finally enjoy SAM with my rx6800XT. Thank you AMD for taking your customers for idiots.

Bro why wouldn't you just buy a 5600x and not have to switch motherboards? Look I wish my 3800x was able to SAM, but apparently its a new HW feature of Zen3 so just let it go. You still have a pretty damn good CPU even if Zen3 has a NEW feature thats not possible on it.

 

[Makaveli]

This doesn't make sense; Zen 2 shares the exact same I/O die with Zen 3. The only thing that changed is the CCX/CCD, but the I/O die handles the communication with the PCIe lanes anyway. This seems like either some half-arsed Intel hit or AMD just not wanting to officially bring SAM to Zen 2.

This has been answered already in this thread.

 

[mechtech]

What is really interesting here if it’s part of the pcie spec and Intel has had it so long did they implement on their integrated graphics??

Kind of reminds me of vesa adaptive sync being around and no one taking advantage or implementing it for years.

Still almost impossible to find a monitor with official DP1.4 certification and look how old that spec is now.

Also strange that AMD has not implemented until now, but I guess back in 2014 they probably had bigger things to worry about?

 

[iuliug]

Let's be honest here. The fact that is it theoretically supported does not mean that owners of 9k series and below will see that feature enabled.

 

[Makaveli]

Let's be honest here. The fact that is it theoretically supported does not mean that owners of 9k series and below will see that feature enabled.

Yup it will be up to the motherboard makers if they want to implement it in older boards, and they most likely will want you to buy newer products for the functionality. So even though Haswell cpu's support the instruction set I wouldn't expect bios updates for those older motherboards.

 

[InVasMani]

I think I might've gotten this right in another thread on the matter. Seems like it's pretty plausible you can mod the bios to support the feature. I think the naming is a little difference in some cases, but it represents the same thing aperture size. There are some other things as well like "payload size" as well that could have minor impacts as well potentially if you dig into the bios settings on a MB chip. I'm not too shocked that Intel's chipsets were more advanced than AMD until very recently. I don't see that as a surprise to anyone given their R&D budget and complications over the last decade.

https://www.techpowerup.com/forums/...-intel-z490-h470-b460-platforms.275415/page-2

That's a bit embarrassing for AMD - introduce a new performance improving feature; limit it to new GPUs *and* new CPUs; watch the motherboard manufacturers introduce it for existing competitor CPUs....

Hopefully this will be another kick for AMD to "find a way" to introduce SAM on older AMD CPUs that are compatible with the same motherboards that support the latest AMD CPUs.

I'm not certain it can be implemented on a MB chipset older than x570/B550 and I think the decoding on the CPU side itself might not even be present on ZEN2 to properly accomplish it for that matter. It compresses data sent across the bus which I believe the chipset controls then decompresses it on the CPU. It requires both far as I can tell it's much much like a GPU and display with different HDMI/Display port specs need to be met.

Haha, thats funny However, the performance difference in games (outside of Forza Horizon 4 for some reason) is minimal, I doubt it will matter much..

If old(er) Intel boards requiree UEFI update to support it anyway, how many manufacturers are going to release an UEFI update!? Not many I expect..

QUESTION; If the motherboard does not have SAM enabled, will Nvidia be able to "enable" it thru drivers anyway!?

It's actually dead simple to implement if it's what I believe it to be I can probably do it myself. It's pretty much already implemented within the bios spec, but was set conservatively for the older 256MB standard and option hidden to the end user. As far as MB makers actually enabling the option that's hard to say and how far back they'd go in doing so is another matter though it's certainly quite plausible. It's interesting it goes back to Haswell on Intel side of things that means a Broadwell chip could potentially enable it which is really quite a interesting thing to contemplate with that EDRAM chip makes me wonder how nicely they play together.

 

[Bytales]

Its interesting how intels CPUs supported it since 2014, yet nobody seemed to bother to give it importance and bring it to light and supported. This feature was basically unknown in the tech world before AMD reminded everyone, theres this thing here, that can bring more performance. And intel was like, oh, we didnt bother about it because we was like already on the top and milking customers year to year, no point in giving them more performance.

Its disgusting what intel is, is all.

 

[InVasMani]

Its interesting how intels CPUs supported it since 2014, yet nobody seemed to bother to give it importance and bring it to light and supported. This feature was basically unknown in the tech world before AMD reminded everyone, theres this thing here, that can bring more performance. And intel was like, oh, we didnt bother about it because we was like already on the top and milking customers year to year, no point in giving them more performance.

Its disgusting what intel is, is all.

GPU VRAM along with storage have evolved a lot since 2014 keep that in mind. On a typical system in 2014 would they have noticed or saw a difference? Could be a simple as proof of concept and not really having ever been put to the test and kind of then subsequently forgotten about entirely. It was probably a bit forward thinking about future tech initially and people then somewhat had forgotten about it by the time it became more relevant.

What is really interesting here if it’s part of the pcie spec and Intel has had it so long did they implement on their integrated graphics??

Kind of reminds me of vesa adaptive sync being around and no one taking advantage or implementing it for years.

Still almost impossible to find a monitor with official DP1.4 certification and look how old that spec is now.

Also strange that AMD has not implemented until now, but I guess back in 2014 they probably had bigger things to worry about?

I think the integrated graphics is a good point for Intel and AMD.

 

[Toothless]

Its interesting how intels CPUs supported it since 2014, yet nobody seemed to bother to give it importance and bring it to light and supported. This feature was basically unknown in the tech world before AMD reminded everyone, theres this thing here, that can bring more performance. And intel was like, oh, we didnt bother about it because we was like already on the top and milking customers year to year, no point in giving them more performance.

Its disgusting what intel is, is all.

That's like saying we need to have five wheels on a car when four wheels has been fine for years. Intel had a bad business practice sure but no one is forcing you to use it.

 

[Steevo]

What is really interesting here if it’s part of the pcie spec and Intel has had it so long did they implement on their integrated graphics??

Kind of reminds me of vesa adaptive sync being around and no one taking advantage or implementing it for years.

Still almost impossible to find a monitor with official DP1.4 certification and look how old that spec is now.

Also strange that AMD has not implemented until now, but I guess back in 2014 they probably had bigger things to worry about?

It wouldn't make any difference on integrated graphics as this is for the VMEM access on a discreet GPU.

 

[warrior420]

Shocking!!! Company releases new technology for only new products, news at 11.

 

[InVasMani]

I'm not sure how it would would work integrated graphics because the data is encoded and decoded between the GPU and CPU and it's really just the address size to communicate burst transfers of the encoded/decoded data itself. It's really just adjusting the size of data sent. I think that 256MB was just a conservative value for the era of tech and hadn't been readjusted until now as a proof of concept until AMD came along and opened a can of worms. I'm not sure if it's applicable to IGP/APU or not in practice there could be technical reasons where it is or isn't just because there isn't a PCIE slot doesn't mean there isn't PCIE wiring somewhere with shorter traces involved. I'm not sure how it's electrically engineered on the CPU itself with integrated graphics and if there is some sort of a PCIE communication or equivalent that would fall under the same umbrella for the aperture size to matter.

 

[Tom Yum]

Gee, it would be nice if the author could provide a source to support this article, because many elements of the story don't make sense at face value, nor has it been reported elsewhere.

 

[Colddecked]

Sources are right under the picture of the pcie slots.

 

[B-Real]

I know what I have to do. Replace my ryzen 3600 with an i5 10600k and z490 motherboard.
I will finally enjoy SAM with my rx6800XT. Thank you AMD for taking your customers for idiots.

Did they promise you it will be supported on older CPUs or what? Don't be so angry.

 

[Zmon]

As interesting as this is, there's a slim to 0 chance that motherboard manufacturers put out an update for Z370 to Z87 boards (and their H/B variants). This will likely only be for the LGA1200 boards and newer.

 

[Tom Yum]

Sources are right under the picture of the pcie slots.

Apologies, didn't see the tiny light grey on white font. Looking through them, no conclusive evidence has been provided as to why hardware accelerated PDEP instructions are required to support SAM. The whole case seems to hinge on: Zen3 implements single cycle PDEP commands (instead of slower microcode) => PDEP supports 32/64 bit manipulation => Zen2 doesn't support SAM because PDEP too slow. The sources don't demonstrate why the PDEP commands are required to support SAM (or the underlying PC-SIG spec), or why different commands couldn't be used that do perform better on AMD, and therefore why AMD's implementation pre-Zen2 prevents SAM support.

Secondly, many elements of the article are incorrect. PDEP commands were introduced in BMI2 which is supported by AMD since Excavator. They aren't a PCIE physical layer 'feature'. Whilst Zen3 accelerated these instructions to a level in line with Intel (with lower latency), it hasn't been shown that the Zen2 performance was too low to support SAM/Resizable BAR. The conclusion is therefore not supported by the evidence presented.

 

[MikeSnow]

The conclusion therefore is therefore not supported by the evidence presented.

But it's not contradicted either. And it seems plausible.

 

[InVasMani]

The latency relative to the gains might've been too high. It could also a be a 32-bit vs 64-bit memory addressable size limitation too perhaps think system memory on a 32-bit OS. A aperture size of 4GB or more I think requires 64-bits. That said it's hard to say if that aperture size still could be increased up to a certain point on Zen/Zen 2 like up to 2GB size possibly 3GB similar to PAE I'm not certain that's some speculation though on my part. I'd be a bit surprised of 256MB (aperture size/BAR size) is the actual hard limitation wall on Zen/Zen2.

 

[Tom Yum]

But it's not contradicted either. And it seems plausible.

That is not how proof or evidence works though. If you make a definitive statement, then the evident needs to support it definitely. If the article was 'SAM may be supportable back to Haswell, not Zen 2' then I'd agree with you.

 

[MikeSnow]

That is not how proof or evidence works though. If you make a definitive statement, then the evident needs to support it definitely. If the article was 'SAM may be supportable back to Haswell, not Zen 2' then I'd agree with you.

Do you have any evidence that SAM is using just the resizable address bar feature, not a combination of features? AMD didn't say exactly what it is, just that it requires re-size BAR support, the rest are just assumptions.

It seems plausible that after AMD optimized those CPU instructions on Zen 3 they started using them in their video card drivers, and that's probably where the main benefit comes from. Then they probably found that things will work a bit better if you give those instructions access to the video card memory as one continuous address space as well, so they used the two features together under the SAM name.

Would have made any sense to implement the two optimizations as separate features, assuming it could be done? From our perspective as users, probably. From their perspective, if the actual benefits from the resizable BAR were minimal, no. Especially since it would have made their video cards run faster on old Intel CPUs which already include optimized versions of those instructions, compared to older AMD CPUs that do not. By withholding these details, they would give you one more reason to buy a Zen 3 CPU instead of staying on Intel, at least until the Intel BIOSes received the required updates.

Does it suck for us the consumers if that's what they actually did? Yes, it does. But is it actually a lie? No, it's just withholding the details. Basically, doing this gave them a temporary advantage over Intel and gave users more reasons to switch to AMD, at least initially. I don't like it, and I'll remember it, but it won't stop me from buying AMD components if they are worth it. It just proves you need to wait for the dust to settle for a bit and not buy something from the first day, or even worse, preorder, unless you really need it. That stands for both software and hardware. It's not worth the risk.

Anyway, these are just my assumptions. The only way to know for sure is to wait for NVidia to release their implementation. Then we'll see what happens when you try to use it with old Zen 2 CPUs. My guess will be that on Zen 2 the performance increase from the resizable BAR will be minimal, and any significant benefits will be from the use of those CPU instructions on the CPUs that have the proper optimizations.

One question, does anyone an AMD motherboard with a new BIOS, which has re-size BAR support, but a Zen 2 CPU installed? If so, can you check if the required options "Above 4G Decoding" and "Re-size BAR support" are available, and can be enabled? I'm not asking if they enable SAM or not for the video card, just if they are available and can be set.

Because it's not clear to me if the check for Zen 2 vs Zen 3 is done in the BIOS of the motherboard, disabling those two options in the BIOS completely, or the video card driver, which would disable just SAM, not the required features.

Depending on this, NVidia might or might not be able to enable something similar to SAM on Zen 2. It would also depend where the performance benefit comes mostly from, the optimized CPU instructions or the resizable BAR support, so I wouldn't bet much on NVidia being able to bring any performance improvements on Zen 2.

And if it's just the motherboard BIOS preventing the activation of those features, it would be very interesting what would happen if somebody could patch the BIOS to bypass this limitation.