Current Generation Hardware Speculation with a Technical Spin [post GDC 2020] [XBSX, PS5]

[psorcerer]

Ok ... so we agree, it is a hardware feature? You need the TSS hardware to do SFS ... so it's a hardware feature.

It was in a response to some tweet that implied the feature is MSFT-specific.

 

[Scott_Arm]

It was in a response to some tweet that implied the feature is MSFT-specific.

Yah, I think it's very likely PS5 has the same hardware, because their gpu is listed as RDNA2.

 

[chris1515]

https://twitter.com/x/status/1241454613993152512

Maybe the PS5 OS RAM reservation is lower than the Xbox reservation. This guy which worked on Ps5 before leaving Sony praise the NX Gamer video.

 

[psorcerer]

This guy which work on Ps5 before praise the NX Gamer video.

People told me that the vast majority of the reserved RAM comes from the "game VCR" features.
I.e. some shitheads that want to record how they are playing waste 2GB of precious GDDR6...

 

[3dilettante]

AMD has claimed a 50% uplift in power efficiency for RDNA1 over RDNA2, to which they specifically mention higher clocks:


It remains to be seen how much of these 50% come from better IPC and how much it comes from higher clocks.

The 5700 had a typical clock (game clock) at around 1700MHz. If you consider 35% of the total 50% improvements are coming from increased clocks, then we should expect a 5700 XT equivalent card at the same power budget to be clocked at 1700*1.35 = 2.3GHz like you suggested.

One interpretation of Cerny's statement about ~2% change in clocks producing ~10% in power is that the range the PS5 is operating in is past the knee of the curve in terms of power efficiency.
AMD may not be choosing its reference points in an entirely transparent manner, just as its reference point for the GCN to RDNA transition was the 16nm Vega 64 rather than Vega VII.
In prior generations, I think there were examples of AMD using mobile SKUs or lower-power implementations like the Fury Nano to set up their PR slides, giving an impression for the overall family that wasn't applicable to most of the products in it.

If there is a "big Navi2" or whatever it turns out to be, it could use its larger size to step clocks back from whatever region the PS5 is operating in and get better results--unless a competing GPU forces AMD to upclock and overvolt again.

1. I think NV has coherency hashing in their hw. Doesn't seem to help much though.
2. I think most of the problems with RT come from: "what to do after" rather then "how to find the intersection".
I.e. I suspect that main problems come from pretty random memory access patterns when doing any real sampling for materials on the objects.
And BVH lookup is not that sequential either.

I'm not quite sure how to determine what form of coherence ordering Nvidia has for its rays, but how would we know it was helpful without knowing how it performs with it on and off?

The custom BVH formats the vendors use should be trying to pack nodes with an eye on efficiently utilizing cache lines and increasing reuse, if possible. If the RT method matches AMD's patents, can we rule out at least some cache line reuse by the hardware in the CUs? In AMD's case, if the execution loop depends on the L1 (edit: L0 for RDNA) cache, those latencies are particularly poor and might be noticeable as part of its hybrid execution loop between the intersection hardware and SIMDs.

 

[Kaotik]

AMD may not be choosing its reference points in an entirely transparent manner, just as its reference point for the GCN to RDNA transition was the 16nm Vega 64 rather than Vega VII.

The reference points in the +50% energy efficiency slide are the same, Vega 64 & 5700 XT

 

[3dilettante]

The reference points in the +50% energy efficiency slide are the same, Vega 64 & 5700 XT

And at the time that claim was made, a lot of people didn't read the fine print or had trouble finding the footnote that it wasn't the 7nm Vega VII. The choice of the start and endpoint can change how we should regard the percentage.
When AMD gave the power improvement for the Fury generation, it was even less transparent and apparently used a Hawaii or Tonga start point and apparently a mobile-clocked or Fury Nano endpoint, but implied the improvement was for the Fiji family.

The claim for RDNA2 has an unknown endpoint that could be one of the SKUs it doesn't push too far, and we have ample evidence that AMD didn't try that hard to optimize the 5700XT's voltage or clocks for power-efficiency.

 

[Adonisds]

Do you think it's possible now to build a PC that will play all multi platform games decently, for the next 9 years, after we received the most important console specs? Or is there still a lot of uncertainty regarding the SSD requirements of PC versions?

 

[Cyan]

https://twitter.com/x/status/1241454613993152512

Maybe the PS5 OS RAM reservation is lower than the Xbox reservation. This guy which worked on Ps5 before leaving Sony praise the NX Gamer video.

dunno who has the best explanation of the specs, but it doesn't matter. Both the PS5 and XSX were shown at a bad date with all the things going on in the world right now, a lot of technical mumbo jumbo but boring presentations. I think with a console you have to sell the hardware with a tech explanation of course but also with something more, better marketing, fun ads, etc.

Also, how to forget the PNG image of 5 people looking at Mark Cerny doing the presentation, and the guy might be a genius, but man, is he boring.

 

[Silent_Buddha]

Do you think it's possible now to build a PC that will play all multi platform games decently, for the next 9 years, after we received the most important console specs? Or is there still a lot of uncertainty regarding the SSD requirements of PC versions?

If you plan to build a PC similar to next gen consoles, I would wait until MS has officially launched DirectStorage on PC and then see if it requires specific DirectStorage compatible SSDs. And even more than that see benchmarks of said SSDs in games optimized for DirectStorage.

That said, likely all modern PCs will be able to play multiplatform next gen games. It's just at what level of performance and with which IQ features enabled or disabled?

Regards,
SB

 

[anexanhume]

And at the time that claim was made, a lot of people didn't read the fine print or had trouble finding the footnote that it wasn't the 7nm Vega VII. The choice of the start and endpoint can change how we should regard the percentage.
When AMD gave the power improvement for the Fury generation, it was even less transparent and apparently used a Hawaii or Tonga start point and apparently a mobile-clocked or Fury Nano endpoint, but implied the improvement was for the Fiji family.

The claim for RDNA2 has an unknown endpoint that could be one of the SKUs it doesn't push too far, and we have ample evidence that AMD didn't try that hard to optimize the 5700XT's voltage or clocks for power-efficiency.

There probably aren’t too many unique “Big Navi” SKUs to generate this comparison. We can also assume 5700XT is the reference as it would generate the most favorable numbers. The biggest concern is we don’t know what the comparisons are iso of. I’m sure big Navi does quite well at the same frequency because the voltage is lowered. We also have no idea how broad the comparison is since it’s “internal estimates.” One game that leverages VRS could have a huge delta.

One interpretation of Cerny's statement about ~2% change in clocks producing ~10% in power is that the range the PS5 is operating in is past the knee of the curve in terms of power efficiency.

I think that’s pretty much a given since Cerny intimated they’re operating on the edge of static timing closure.

 

[Pinstripe]

Is there a reason why Cerny mentioned only Primitive Shaders for the PS5, but nothing of Mesh Shaders or VRS support in his presentation? I thought Primitive Shaders were a RDNA1 feature, while RDNA2 uses Mesh Shaders.

 

[anexanhume]

Is there a reason why Cerny mentioned only Primitive Shaders for the PS5, but nothing of Mesh Shaders or VRS support in his presentation? I thought Primitive Shaders were a RDNA1 feature, while RDNA2 uses Mesh Shaders.

He did talk about mesh shaders in all but name when he referenced the Geometry Engine and its features. We’re all wondering why he didn’t mention VRS. My guess is the interest of time. His talk wasn’t an exhaustive coverage of the GPU capabilities.

There does seem to be a restriction if the OS uses an IOMMU and enables virtualization-based security. It's possible the PS5 will, given the collection of other IO processors and DMA engines running freely in the SOC.
Whether Sony implements some other form of security measures around the PCIe subsystem, or the system software has other safeguards is probably something Sony wouldn't discuss. Sony has displayed surprising amounts of neglect in its security schemes before, however.

The m2 slot's PCIe link goes to the IO block on the main SOC, same as the custom SSD does.
We know that decompression and priority arbitration would be handled in that complex, and while I don't recall a mention of decryption, I think the most likely place for encryption and tamper checking is on the main SOC. That should give some defense of the data in persistent storage, barring security flaws or Sony's already mentioned neglect.

I'm not clear where the Tempest block is, or where its DMA might be.

Left unsaid is where the OS and firmware are being loaded from or into. I'd like to think Sony learned something from how stringing the boot process over the link from a separate southbridge compromised security.
We'd need to find out whether there's still a southbridge chip, since at least one of its functions was serving as a controller for the HDD, which the PS5 would have on-die.

The anti-tamper measures in this generation are fairly extensive:

https://www.crn.com/news/components...on-work-led-to-a-big-security-feature-in-epyc

 

[Barrabas]

He did talk about mesh shaders in all but name when he referenced the Geometry Engine and its features. We’re all wondering why he didn’t mention VRS.]

Is there a context between VRS and Geometry Engine?

https://twitter.com/x/status/1240748262736269313

 

[iroboto]

Is there a context between VRS and Geometry Engine?

https://twitter.com/x/status/1240748262736269313

I’m not sure. It should have both. Geometry engine I assume is another label for Mesh shader. And VRS is VRS. You can do both to save cycles.

 

[Betanumerical]

I’m not sure. It should have both. Geometry engine I assume is another label for Mesh shader. And VRS is VRS. You can do both to save cycles.

Its a bit weird to say it in the way he said it then, because im not a aware of a GPU that has VRS but no Geometry engine in that case.

 

[Mskx]

Is there a reason why Cerny mentioned only Primitive Shaders for the PS5, but nothing of Mesh Shaders or VRS support in his presentation? I thought Primitive Shaders were a RDNA1 feature, while RDNA2 uses Mesh Shaders.

"He didn't mention 'x feature' so that it means they don't have it" is obviously a flawed way of thinking and not really proof of anything.
But Xbox has harped quite hard on VRS in particular, and even called it "Our patented form of VRS", that it makes me wonder if they know something.

 

[function]

.... But Xbox has harped quite hard on VRS in particular, and even called it "Our patented form of VRS", that it makes me wonder if they know something.

Yeah, it's got me wondering too. If it was merely an oversight or not appropriate for the talk, I would have expected Sony to clarify afterwards, as they did for the backwards compatibility issue.

 

[iroboto]

Yeah, it's got me wondering too. If it was merely an oversight or not appropriate for the talk, I would have expected Sony to clarify afterwards, as they did for the backwards compatibility issue.

I think they want the discussion to be dominated by their SSD tech right now. Things like VRS, SFS etc, are all things that fall dull on ears.
As we speak, the narrative trying to put PS5 as the front runner over XSX as being the better machine. Multiple journalist, devs, etc are coming out to re-write the narrative of the internet.

So they may just sit on that strategy going forward. And keep pushing that narrative that SSD is the only thing that matters and they have 2x of it over XSX. Which is honestly, the narrative is ridiculous because the internet only sees the 2x factor and sort of just ignores the fact that XSX is running a very high speed SSD as well. Apparently 50x faster hard drive speeds is not enough for a generational leap, we need 100x faster speeds.

I sort of get it.

XSX is safe. They're going to beat PS5 in every single 3P game, there's no doubt in my mind on this. MS's strategy extends from X1X: they know well in advance what their performance numbers will be like with live code; they know it likely for RT code as well. When they set out a target to hit 4K with X1X, they did a good job of hitting that target. The target for XSX is 4K@60 now. I suspect we'll see most of their titles at that range with only a few titles in the 30fps range.

The 3P games need a baseline for performance and PC is going to be well below XSX which is 2x below PS5. And that just moves things in favour of MS in this case. XSX will have better frame rates, better resolution, and higher graphical settings. Even if they are minor and most people won't notice it without the help of DF.

But if you can sell the public that this SSD tech will be the shining light for this generation, then that's all you need to do is convince them of it. SSD isn't as safe as what MS did, so I'm sort of drawn to it.

I don't know if there will be better design paradigms as a result of SSD. But the narrative that CPU was supposed to bring those design paradigms seems to have been tossed out in favour of SSD already.

Things move fast

 

[zupallinere]

don't know if all is true, but interesting read

It sure looks like an optimistic take and I would think the OS memory footprint might be conservative as it was for the PS4 at the beginning. I sure did enjoy seeing things laid out that way however and hope for the best as well.
A couple of things of note that might backup some of the claims from the DF article:

The controller itself hooks up to the main processor via a four-lane PCI Express 4.0 interconnect, and contains a number of bespoke hardware blocks designed to eliminate SSD bottlenecks. The system has six priority levels, meaning that developers can literally prioritise the delivery of data according to the game's needs.

Might this prioritization system help if some of the OS is placed into the SSD ?

it could really hurt the GPU performance - so we've implemented a gentler way of doing things, where the coherency engines inform the GPU of the overwritten address ranges and custom scrubbers in several dozen GPU caches do pinpoint evictions of just those address ranges."

This reminds me of a discussion about garbage collection for VMs (recent vintage Java in that case ) where well managed clean up can really keep something moving reducing stalls (race conditions as well but that might not be an issue with a GPU per se)