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

[function]

I wonder how much of that could be shifted to some DDR4 if they take the ARM+DDR approach again. Or maybe just the DDR approach.

There'd certainly have to be some amount of the GDDR6 reserved for the OS, but screenshots, DVR, downloads, notifications, Spotify, party chat, streaming, the web browser, and the store could probably be served just fine by a couple of GB's of DDR4.

I think there was something like 512MB of memory on the PS4 that had to be available to the OS when requested. So it's possible they've iterated on that with the PS5.

My guess is 1GB of GDDR6 reserved by the OS at all times, and 1GB that can be used by either the OS or the game. It only takes a fifth of a second to fill that from the SSD. I can wait a fifth of a second to see my dynamic Subnautica theme.

Then 2GB's of DDR4 for the other stuff I mentioned. Maybe 4GB's of DDR4, and the 1GB of OS-available GDDR6 gets swapped in here. The extra 1GB available to developers if they want to venture out of the GDDR6.

I'm assuming there isn't any additional DDR4 simply because they haven't mentioned it (yet! ). Even 2GB of DDR4 that was directly addressable by the main APU would be a game changer in this regard, and allow for a significant reduction in GDDR6 reserved.

Come to think of it, didn't an early PS5 main board "leak" describe DDR4 mounts near the SSD....?

 

[PSman1700]

Almost certain there wont be any ddr4.

 

[disco_]

Could you put up a link please, interested to read a bit more.

It's just those tweets people have posted.

https://twitter.com/x/status/1240825295994793984

https://twitter.com/x/status/1240841832503934976

https://twitter.com/x/status/1240842648224698368

https://twitter.com/x/status/1240823350399811586

His tweets sound like someone curious about BCPack rather than someone who owns the tech.
Bcpack allows Ms a compression of 2:1 while Sony has ~1.6:1. Is that of any significance if the raw speed is faster than compressed?

 

[Barrabas]

Come to think of it, didn't an early PS5 main board "leak" describe DDR4 mounts near the SSD....?

Yes there was a rumor, but I am certain that would have been brought up in Cerny's presentation if that was the case. Now I am interested to find out if there is anything in the talk about maybe the OS can be handled in the SSD?

 

[RobertR1]

Sony yields might be interesting to learn about and keep an eye on. 2.2ghz+ is a lot to validate en masse.

 

[Jay]

His tweets sound like someone curious about BCPack rather than someone who owns the tech.
Bcpack allows Ms a compression of 2:1 while Sony has ~1.6:1. Is that of any significance if the raw speed is faster than compressed?

Thanks.
I thought people meant someone else as they said the creator of it.
He's obviously very knowledgeable, but he's just making educated guesses until we're told otherwise.

Sony has gone into more depth now, so will be good to hear ms about audio, bcpack, VRS, any other customizations.

 

[PSman1700]

Very intresting the tweets on bcpack.

 

[Deleted member 13524]

Off the top of my head from the games I have, BFV, Assassins Creed, Shadow of the Tomb Raider and a few others are known to be "CPU heavy" on the PC due to their use of AVX. I'm sure that number will continue to grow.

If AVX is the culprit on those games, then it's most probably AVX1, which AFAIK is mostly 128bit operations.
Cerny seemed to claim that 256bit operations (AVX2) were the biggest culprits for scaling up power consumption.
Has AVX2 been in any PC game? That would make it incompatible with every pre-Haswell and pre-Excavator CPU out core out there.

I'm yet to find a game that is incompatible with my Ivy Bridge E system, though I think that's bound to happen this gen with Zen2 being used as baseline.

Bcpack allows Ms a compression of 2:1 while Sony has ~1.6:1. Is that of any significance if the raw speed is faster than compressed?

Cerny said Kraken with 5.5GB/s raw bandwidth is typically 8-9GB/s effective bandwith, but it can go as high as over 22GB/s. That's 4:1.
At 17m23s:

https://youtu.be/ph8LyNIT9sg?t=1043

He claims the SoC-embedded Kraken decompressor is doing the same job as 9 dedicated Zen2 cores. It seems to be a pretty hefty compression tech, and a hefty decompressor to make this work.

Come to think of it, didn't an early PS5 main board "leak" describe DDR4 mounts near the SSD....?

Yes. 2* 4GB DDR4 next to the SSD complex and one 4GB DDR4 chip closer to the SoC.

Yes there was a rumor, but I am certain that would have been brought up in Cerny's presentation if that was the case.

This was a presentation for developers. Whatever they put there for the OS to use internally is of little concern to developers..
It's not that he couldn't talk about it (if it's there), but the presentation was already almost a full hour long and he didn't talk about a lot of things like the cooling system or HDMI 2.1 features or any peripheral.

 

[VitaminB6]

Sony yields might be interesting to learn about and keep an eye on. 2.2ghz+ is a lot to validate en masse.

It'll be interesting to see what the typical clock frequency range will be on AMD's RDNA 2 GPU's. Maybe 2.3-2.5Ghz will be the norm for cards released to consumers?

 

[PSman1700]

Maybe 2.3-2.5Ghz will be the norm for cards released to consumers?

Doubt it, not gameclock atleast. Higher end gpus most likely will be wider then 36CU count. Mid range perhaps, when boosted, to those speeds. Even then i doubt it, you usually dont design a gpu with rediculous speeds like that, there would be no room for any non-reference factory oc cards. And since were on pc, where the user might run whatever fps/settings where other problems could occur hardware related.
But who knows, maybe we will see 2.5ghz midrange gpus running at their toes.

There were rumors of monstrous die sized 80+ rdna2 gpus close to 18TF on hbm. One thing is sure, a gpu lile that wont run at those speeds.

 

[Barrabas]

This was a presentation for developers. Whatever they put there for the OS to use internally is of little concern to developers.

Certainly the amount of memory available for the developers is interesting.

 

[Deleted member 13524]

It'll be interesting to see what the typical clock frequency range will be on AMD's RDNA 2 GPU's. Maybe 2.3-2.5Ghz will be the norm for cards released to consumers?

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.

Certainly the amount of memory available for the developers is interesting.

Did he say how much memory is available to developers?
Even if he did, it wouldn't say how much DDR4 the OS is taking for its features, e.g. for video recording.
If the DDR4 is there.

 

[RobertR1]

If AVX is the culprit on those games, then it's most probably AVX1, which AFAIK is mostly 128bit operations.
Cerny seemed to claim that 256bit operations (AVX2) were the biggest culprits for scaling up power consumption.
Has AVX2 been in any PC game? That would make it incompatible with every pre-Haswell and pre-Excavator CPU out core out there.
.

Yes, AVX1. I don't think I've come across any AVX2 games. AVX1 is 256 but AVX2 expands the instructuion sets. AVX512 is 512 bit.

On my 9900k @ 5ghz running P95 small the load is as follows:

No AVX = 145amps
AVX = 187amp
AV2 = 198amps

 

[Scott_Arm]

Oh, OK. Dictator didn't say it, Dark1x did. That must be one of the threads I'm not following as well. That said, I wouldn't take their word as sacrosanct. DF has been wrong in the past and do not always accurately interpret the information shared with them.

And I will reiterate this behavior they are describing is the opposite of what Cerny talked about in the presentation: that certain GPU loads cause the GPU to throttle, and certain CPU loads will cause the CPU to throttle, not that CPU loads will throttle the GPU, or back the other way. Of course, even if that's true it's not clear how much that matters if we assume the GPU load that causes the CPU the throttle means you're GPU limited anyway, and vice-versa.

dictator does say it.

 

[chris1515]

Reads like NXgamer

No this is true streaming work like this GPU scrubbers too and Xbox memory work like this. This is the same addressable memory the more you use the 6 GB/s more your bandwidth goes near 336 GB/s you need to let CPU and low frequency system inside this memory of no it will decrease seriously the bandwidth addressable by the GPU.

 

[Barrabas]

Did he say how much memory is available to developers?

From about 12:40 in the presentation he says "most of the ram is working on the games behalf" not much to conclude from that, but anyhow the developers need to know the amount available. Most likely they are given this in the devkit documentation and he didn't feel to say it there. I still think it's a bit strange not to be clear about it in a GDC presentation.

 

[psorcerer]

There's some theorizing that it's not strictly bandwidth, but also the memory transaction rate that the RT hardware can achieve. Since traversing the BVH means analyzing nodes and then following pointers to successive entries, there's more round-trip traffic rather than raw read bandwidth. Some of the research for ray tracing involves finding ways of gathering rays that hit the same nodes of the tree, which along with saving bandwidth saves on the number of round trips the hardware may need to make.
Depending on how well the BVH hardware caches data, or how often BVH data can be accessed again while still in a cache, there can be a benefit to higher clock speeds. Cerny did note that the primary downside is that memory is effectively further away in terms of latency, which means more divergent rays or less effective cache situations would make trips to memory or contention for memory more expensive for higher-clocked units.

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.

 

[disco_]

Thanks.
I thought people meant someone else as they said the creator of it.
He's obviously very knowledgeable, but he's just making educated guesses until we're told otherwise.

Sony has gone into more depth now, so will be good to hear ms about audio, bcpack, VRS, any other customizations.

He is the founder of Binomial and BCPack is a Binomial product, unless I'm mistaken.

 

[Jay]

He is the founder of Binomial and BCPack is a Binomial product, unless I'm mistaken.

Maybe it is a binomial product, but where did you hear that from?
I didn't see anything in tweets that says that.
Looks like he may have worked at MS at some point though.

 

[iroboto]

Yes, AVX1. I don't think I've come across any AVX2 games. AVX1 is 256 but AVX2 expands the instructuion sets. AVX512 is 512 bit.

On my 9900k @ 5ghz running P95 small the load is as follows:

No AVX = 145amps
AVX = 187amp
AV2 = 198amps

you mean milli amps right

hahah 198 amps could charge 5 teslas simultaneously