Baseless Next Generation Rumors with no Technical Merits [post E3 2019, pre GDC 2020] [XBSX, PS5]

Status
Not open for further replies.
The more you rely on streaming from the SSD, the more you are limited by the SSD performance.
I don't think that's the right way to look at it. RAM is just cache between the data we want (on storage) and where it needs to be to be processed (register files in the processors). Presently we have maybe 8 GBs of data in RAM caching what's on the drive, and something like 8 - 32 MBs of L3 cache on the CPUs. That L3 cache is tiny but effective because at any given moment, the CPU doesn't need access to all 8 GBs of data but only a few kbs to feed the L1 and register files, so the caches are just prefetching, just enough to keep those L1/registers populated and CPUs (thanks to multithreading) are kept busy with high utilisation.

Ideally, we'd have all 80 GBs in L1 cache at 4 cycles access, but that's impossible. So we use a small cache to keep it fast and feed that from a bigger bucket.
Ideally, we'd have all 80 GBs in L2 cache at 10 cycles access, but that's impossible. So we use a small cache to keep it fast and feed that from a bigger bucket.
Ideally, we'd have all 80 GBs in L3 cache at 100 cycles access, but that's impossible. So we use a small cache to keep it fast and feed that from a bigger bucket.
Ideally, we'd have all 80 GBs in DDR at 15 ns cycles access, but that's expensive. So we use a smaller pool keep it affordable and feed that from a bigger bucket.

Each pool is only as large as it needs to be to account for the slowness of the next pool. The only reason for large RAM is because the final step, storage, is so incredibly slow it needs an epic pool to basically cache all the data. Once that's fast enough, pressure on the RAM cache is greatly reduced and we can think of it more in terms of the other cache topology, where 1/10th the data size is a huge amount of caching, and quite frankly wasteful but needed because at the moment software design is around using the RAM as the storage instead of a cache.
 
I don't think that's the right way to look at it. RAM is just cache between the data we want (on storage) and where it needs to be to be processed (register files in the processors). Presently we have maybe 8 GBs of data in RAM caching what's on the drive, and something like 8 - 32 MBs of L3 cache on the CPUs. That L3 cache is tiny but effective because at any given moment, the CPU doesn't need access to all 8 GBs of data but only a few kbs to feed the L1 and register files, so the caches are just prefetching, just enough to keep those L1/registers populated and CPUs (thanks to multithreading) are kept busy with high utilisation.

Ideally, we'd have all 80 GBs in L1 cache at 4 cycles access, but that's impossible. So we use a small cache to keep it fast and feed that from a bigger bucket.
Ideally, we'd have all 80 GBs in L2 cache at 10 cycles access, but that's impossible. So we use a small cache to keep it fast and feed that from a bigger bucket.
Ideally, we'd have all 80 GBs in L3 cache at 100 cycles access, but that's impossible. So we use a small cache to keep it fast and feed that from a bigger bucket.
Ideally, we'd have all 80 GBs in DDR at 15 ns cycles access, but that's expensive. So we use a smaller pool keep it affordable and feed that from a bigger bucket.

Each pool is only as large as it needs to be to account for the slowness of the next pool. The only reason for large RAM is because the final step, storage, is so incredibly slow it needs an epic pool to basically cache all the data. Once that's fast enough, pressure on the RAM cache is greatly reduced and we can think of it more in terms of the other cache topology, where 1/10th the data size is a huge amount of caching, and quite frankly wasteful but needed because at the moment software design is around using the RAM as the storage instead of a cache.

A bit of an unrelated question, but what do you think is happening on mobile phones? RAM keeps increasing while storage is getting faster as well. Although I would think that storage on a mobile phone was always faster than an HDD since it's solid? I don't know if Android and iOS are already optimised to use storage in that way, but then again it's only now that mobile storage is getting large enough. Earlier mobile phones probably couldn't use it like this because there was not enough storage? But I guess there might need to be something on the SDKs for software developers to use storage in this way rather than letting the OS decide?
 
And none have the scale or density as the open worlds of GTA or WACTH DOGS. Its' easier to manage persistence when you're juggling 100,000 fewer things in realtime.

Of course, you can create a lot more assets with a team of 200-300+ and outsourcing on top of that which indies will never be able to match.

The point is that any AAA developer could have been making these types of games in massive scale on current gen consoles if they didn't have to target a certain level of graphics.

Or to put it another way graphics and presentation are holding back what AAA devs can do, not the hardware. Hardware will allow them to do more with those graphics, or they may choose to significantly increase graphics and presentation with little to no improvements to game play and then talk about how they had to cut X feature in order to have Y level of graphics (which are required to be better than the previous generation for AAA devs). They don't have the option to go game play first, graphics second.

Regards,
SB
 
Of course, you can create a lot more assets with a team of 200-300+ and outsourcing on top of that which indies will never be able to match.
I'm not referencing the number of unique assets, just the issue of scaling up persistence when the world is more dense.

The point is that any AAA developer could have been making these types of games in massive scale on current gen consoles if they didn't have to target a certain level of graphics.

The nearest anybody has come is Bethesda and even then the persistence is heavily limited to small in-game objects in immutable environments. Even on meaty PCs, tracking the changes requires an aggresive approach by the engine to cull 'orphaned' objects older than an in game week.

Or to put it another way graphics and presentation are holding back what AAA devs can do, not the hardware.

It's more the I/O and keeping track of all those changes. Even the PS3, which easily had the most compromised version of Skyrim, could manage to render and calculate physics for hundreds and hundreds of cheesewheels But PS3's split RAM made maintaining all those changes and objects in RAM an impossible challenge. And tracking game objects is orders of magnitudes easier than tracking changes made to the environment across a world; chunk of wood taken out fo this beam because of sword swing, this hut half burned down by a dragon, that window smashed because of a poorly aimed arrow, these new NPCs serving in this shop due to the previous shopkeepers being slain, this bridge being rebuilt because of a battle.
 
Last edited by a moderator:
There was tes on shaderbench (?) in December 2018. It contained Zen CPU at 3.2GHz and GPU 13E9 Ariel (clocked at 1+GHz).

I have to go through this data but seems to me Ariel was discrete GPU (it was actually referenced as GPU, and was running on actual motherboard) and it maybe was test workbench for Sony while Oberon is actual SOC with Zen2 and iGPU. This is why Oberon iGPU is under Ariel folder tree and Oberon_Native regression test runs Ariel specs.
 
Last edited:
There was tes on shaderbench (?) in December 2018. It contained Zen CPU at 3.2GHz and GPU 13E9 Ariel (clocked at 1+GHz).

I have to go through this data but seems to me Ariel was discrete GPU (it was actually referenced as GPU, and was running on actual motherboard) and it maybe was test workbench for Sony while Oberon is actual SOC with Zen2 and iGPU. This is why Oberon iGPU is under Ariel folder tree and Oberon_Native regression test runs Ariel specs.
Yes, that reminds me the 1ghz ariel gpu models. Now I realise those have being most probably created specifically to start testing PS4 and Pro BC. Because 1ghz never made any sense in the first place to test regular GPUs IMO, but perfect to test 0.8ghz and 0.911ghz BC mode. The later higher clocked being probably used as PS5 devkits.

This goes towards my theory: Ariel has being designed for both reasons: Test BC well in advance for security (2 years before launch date is a lot for finalized silicon) and use them as early devkits in order that dev get used to their SSD tech (API, tools) and start developing their games on PS5 as their main target...and port their games on XBox later.
 
Yes, that reminds me the 1ghz ariel gpu models. Now I realise those have being most probably created specifically to start testing PS4 and Pro BC. Because 1ghz never made any sense in the first place to test regular GPUs IMO, but perfect to test 0.8ghz and 0.911ghz BC mode. The later higher clocked being probably used as PS5 devkits.

This goes towards my theory: Ariel has being designed for both reasons: Test BC well in advance for security (2 years before launch date is a lot for finalized silicon) and use them as early devkits in order that dev get used to their SSD tech (API, tools) and start developing their games on PS5 as their main target...and port their games on XBox later.
Or it was early Navi GPU specced for PS5 before RT/VRS was ready for actual SOC.

Would explain why Oberon uses Ariel GPU tests in Oberon_Native spreadsheet.
 
Ariel was tested on shaderbench (?) in December 2018. It contained Zen CPU at 3.2GHz and GPU 13E9 Ariel (clocked at 1+GHz).

I have to go through this data but seems to me Ariel was discrete GPU (?) and it maybe was test workbench for Sony...
What sort of test workbench? That sounds a lot of expense to me. I'd expect any company to use some already-available substitute part until the test silicon for the target platform was ready. eg. XB360 had Macs for PowerPC coding and 9800/X800 GPUs. MS didn't commission some Unified Shader design a year ahead of the final Xenos design. Why wouldn't Sony just go with a Ryzen CPU and Navi GPU?
 
Yes, that reminds me the 1ghz ariel gpu models. Now I realise those have being most probably created specifically to start testing PS4 and Pro BC. Because 1ghz never made any sense in the first place to test regular GPUs IMO, but perfect to test 0.8ghz and 0.911ghz BC mode. The later higher clocked being probably used as PS5 devkits.

This goes towards my theory: Ariel has being designed for both reasons: Test BC well in advance for security (2 years before launch date is a lot for finalized silicon) and use them as early devkits in order that dev get used to their SSD tech (API, tools) and start developing their games on PS5 as their main target...and port their games on XBox later.
Reminds me of the time early 2019 when the first rumours of AMD co-designing Navi with Sony. Raja made a claim that Vega was impacted because all these engineers were reallocated over to help build the PS5 chip.

I do wonder if somehow these are related items in some fashion (perhaps not the way we envision it, but still related).

Also other rumours swirling around a 2019 release date.
 
What confuses me about this 'Sony contributed to Navi' theory is: How can it be the main competitor gets the same chip (probably even a more powerful version)?
Would this mean Sony gets it cheaper?
 
What sort of test workbench? That sounds a lot of expense to me.

There's been word of different devkits with very different specs being distributed. Smaller devs getting lower performing 9TF kits while 1st parties and AAA devs getting 12TF ones with documentation saying the target will be 13TF.


I don't think that makes much sense, especially if there are 2 distinct SoCs involved, so don't hurt the messenger.
Only way I'd see that as possible is if the final console has a SoC + dGPU.


IMO there wouldn't be documentation claiming 13TF on the grounds of RDNA showing higher performance per theoretical throughput than GCN.
But that might be a good discussion for another thread.



What confuses me about this 'Sony contributed to Navi' theory is: How can it be the main competitor gets the same chip (probably even a more powerful version)?
Would this mean Sony gets it cheaper?

Sony partnered with IBM and Toshiba to create a CPU that consisted of a PPE and 8 SPEs.
At the end of the development cycle, Microsoft came in and ordered a CPU consisted of 3 PPEs for their new console.

So Sony contributing to a tech and then Microsoft adopting it wouldn't be a first time in the console market.
 
Last edited by a moderator:
Yes, that reminds me the 1ghz ariel gpu models. Now I realise those have being most probably created specifically to start testing PS4 and Pro BC. Because 1ghz never made any sense in the first place to test regular GPUs IMO, but perfect to test 0.8ghz and 0.911ghz BC mode. The later higher clocked being probably used as PS5 devkits.

This goes towards my theory: Ariel has being designed for both reasons: Test BC well in advance for security (2 years before launch date is a lot for finalized silicon) and use them as early devkits in order that dev get used to their SSD tech (API, tools) and start developing their games on PS5 as their main target...and port their games on XBox later.

To me if they went through all this process.....then that's the chip...

It's also possible 2GHz is only for the final dev kits and retail will be ~1.8GHz. 2GHz is probably doable if it's for several thousand dev kit units.

Edit: Actually wait. It's probably unlikely that they different clocks between dev kit and retail. The dev kit is probably at the same clock but just using the full 40CU's.
 
Last edited:
What confuses me about this 'Sony contributed to Navi' theory is: How can it be the main competitor gets the same chip (probably even a more powerful version)?
Would this mean Sony gets it cheaper?
I think Brit hit here:
Simpliest answer is that theory was entirely bullshit.

It's possible that engineers were moved from Vega to assist with the PS5 chip. Doesn't mean they co-designed the main navi line, perhaps that was just wishful/poor interpretation of what happened. Likely the wrong guess. But people loved the idea so much that's was the interpretation that circulated.

They may have been building their own things here (for a 2019 launch), in this case, helping to build Aerial, whatever it is (perhaps needing to help developing a specific version of navi that would support their needs). Then later decided to cancel the 2019 launch in 2017 (after seeing the X1X reveal -- IIRC they were surprised by the 6TF spec - though I need to figure out where I read that from edit/ no chance I'll find it - next gen news dominating search results), and then refit it for a 2020 launch with additional features and requirements.
 
Last edited:
Yes, RayTracing with doubling of graphical performance, quadrupling of cpu performance, doubling of memory, and insane increase in resource streaming throughput from storage won't amount to massive graphical leaps. Not without increasing demand on the content creation pipelines. If companies are not willing to increase that aspect, things will not improve. They'll just keep pumping out 4Pro/OneX level games.
 
Status
Not open for further replies.
Back
Top