Rumor: PS5 devkits ~ 13 TFLOPS

Status
Not open for further replies.

SonGoku

Gold Member
Aug 16, 2018
2,028
1,713
320
Sony should just go with the 24gb gddr6 @ 864 GB/s. Call it a day. I'm sure Samsung would cut them in especially when they have started mass production
I was thinking that... but considering Samsung 16Gbps chips are at the sample status, I would be happy with 700GB/s+
For 800+GB/s 18-20Gbps chips are needed and no mention of those on Samsungs site.
 
Last edited:

xool

Member
May 29, 2018
317
211
210
btw this video
(linked before) is really interesting at the 11:30 mark - he says that the interposer (in HBM) used to cost $100 + , whereas now (2019) it's more like 10-20$ - I think this is a big part of the HBM originally being much more expensive compared to traditional RAM - also reinforces my thinking that the "HBM is expensive" idea was right in 2016/7, but is now mostly out of date.

Sony should just go with the 24gb gddr6 @ 864 GB/s. Call it a day. I'm sure Samsung would cut them in especially when they have started mass production
My gut feeling is you are right.
 
Last edited:
  • Like
Reactions: Imtjnotu

mckmas8808

Member
May 24, 2005
40,639
3,595
1,560
Fake.
Would be funny if Sony used ‘Snake Eater’ though since MS is using Anaconda.
Oh snap! That would funny.

btw this video
(linked before) is really interesting at the 11:30 mark - he says that the interposer (in HBM) used to cost $100 + , whereas now (2019) it's more like 10-20$ - I think this is a big part of the HBM originally being much more expensive compared to traditional RAM - also reinforces my thinking that the "HBM is expensive" idea was right in 2016/7, but is now mostly out of date.



My gut feeling is you are right.
Could it be due to cooling issues? Doesn't the heat issues become easier to deal with HBM2, than GDDR6?
 
Last edited:

mckmas8808

Member
May 24, 2005
40,639
3,595
1,560
where do you get your pricing information?

in such a configuration:




DDR4

it would cost like 12 x $15 = $180 for 672-768GB/s



HBM2

2 times 4GB : 2 x $75 = $150 for ~500GB/s

or

4 times 4GB: 4 x $75 = $300 for ~1000GB/s


(if you don't buy in sony bulks of course)
$300?!?! Geez. and the CPU/GPU combo will cost at least $120 right?
 

Racer43

Member
Jun 14, 2018
252
195
195
We don't know how many stacks the "cheap" version allows yet.
Thats correct. But dont count it out. Been in the oven for a long time (projected for 2019 originally I believe?) so should be ready in time for production ramp of next-gen (assuming launch fall 2020).
 
Last edited:

SonGoku

Gold Member
Aug 16, 2018
2,028
1,713
320
Thats correct. But dont count it out. Been in the oven for a long time (projected for 2019 originally I believe?) so should be ready in time for production ramp of next-gen (assuming launch fall 2020).
If it allows 24-32GB HBM3 im all for it. Im leaning more towards GDDR6 atm.
 

LordOfChaos

Member
Mar 31, 2014
8,615
582
575
Didn't know that, can you explain how? or link a vid/article
About GPGPU benefiting? Basically one of the big barriers to GPU compute has been sharing results between the CPU and GPU. Sometimes, copying the results back over would take so much time as to make moving it to the GPU not worth it. With unified memory, there's no copying results over, the GPU can work on something in memory and the CPU can start doing its side on it immediately.

GPU device memory bandwidth is typically 5X to 10X higher than CPU DRAM memory, and the longer latency doesn’t matter in a stream-optimized GPU computing engine architected to tolerate latencies through very efficient hardware-based multithreading. In most accelerated systems, these two memories are connected using a PCI-Express bus, which is a very good I/O bus but not a very good memory bus. At best, PCI-Express can deliver a few tens of gigabytes a second when moving data between the two memories. That makes it a bottleneck, and the CUDA programming model was designed to focus the programming on explicit data management in order to maximize efficiency and performance.



Languages like CUDA will attempt to hide this away, creating a unified pointer space for developers, prefetching, etc, but there's no replacement for physically being on the same pool. I think Sony even said it was their most requested thing from developers when building the PS4

 
Last edited:

Imtjnotu

Member
Nov 4, 2017
738
902
200
san diego
NOICE! Maybe they didn't update their site?
18Gbps makes 800+GB/s speeds possible, that amount of bandwidth on a console would be beast.
It helps to future proof it also in some ways. I remeber in 2013 thinking 176gbps was a huge amount. 800 gives so much headroom
 

SonGoku

Gold Member
Aug 16, 2018
2,028
1,713
320
It helps to future proof it also in some ways. I remeber in 2013 thinking 176gbps was a huge amount. 800 gives so much headroom
Yeah its a game changer really what devs could do with this amount of bandwidth and super fast ssd storage. Should help get the most out of the GPU
 

SonGoku

Gold Member
Aug 16, 2018
2,028
1,713
320
Old Pastebin leak, not very reputable, but I'll take it on this drought
  • CPU 7nm ryzen 8 core 16 threads,unknown speed
  • GPU 7nm Navi arhitecture around 14TF,its gonna be powerful and power efficient,Sony working with Amd for Navi,some sort of Ray Tracing but will not focus on that,more focus with VR and 4k,much better bandwidth overall
  • 24GB Gddr6 + 4gb ddr4 for os,we have 32 gb dev kits
  • 2tb hdd some sort of nand flash

@LordOfChaos
Whats your take on these comments (from reera):
DukeBlueBall said:
You can automatically discredit any rumor that has small pool of DDR4. With the SSD write speeds you can page out any OS + Apps allocations in under a second. 3GB of OS can be written to the SSD in under a second.
BradGrenz said:
Yes, I agree with this. It will be very easy to tombstone/suspend background apps. So much so that the idea of having an entire DDR4 controller just for supporting 4GB of OS RAM feels incredibly wasteful.
BradGrenz said:
Because if you have a lightning fast SSD media apps never need to be stored in memory just to accomplish fast multitasking. You will be able to instantly suspend and resume them on demand like on a smartphone. The OS memory needs will only be for actual background services which require far less memory.
DukeBlueBall said:
you can essentially use the SSD like a small pool of DDR4 when it comes to OS, Apps, and flexible game memory.
 
Last edited:

LordOfChaos

Member
Mar 31, 2014
8,615
582
575
@LordOfChaos
Whats your take on these comments (from reera):

>3GB of OS can be written to the SSD in under a second.

This makes a common mistake when talking about todays SSDs, as it implies everything would be a tidy sequential read/write. We don't live off of fast SSDs with no RAM because the fastest are still almost an order of magnitude slower on sequential than just plain old single channel DDR4 (Even DRAMless SSDs are kind of horrid), and that's without considering latency and random access speeds (the RA part!). Realistically you hit smaller files and less predictable access and cut your transfer rate down several times. So now how does 400MB/s for a mixed batch of files instead of 19200 sound, and add another order of magnitude in latency, it adds up when the OS is hitting it with millions of requests.

It's kind of just a truism that faster SSDs make OS paging less painful, but to run most of the OS and background apps off the SSD and keep loading/offloading them...I think that would at least cancel out how much faster the CPU would be for the OS speed, and I'm already impatient with the PS4 OS speed lol


>So much so that the idea of having an entire DDR4 controller just for supporting 4GB of OS RAM feels incredibly wasteful.


I mean, one of the great things of the 8th gen over the 7th was the ARM coprocessor allowing background updates and downloads, I assume the 9th will still have that, and if you have an ARM coprocessor it needs RAM...So, wasteful to have a tiny memory controller on the coprocessor? Needs it anyways


I thought most rumors were pointing to it having some DDR4 for the OS but none of us know, but I know which one would perform better for getting in and out of the OS, store, browser, etc. Plus 1GB in the main pool is less than we already allocate to the OS this gen, so that sounds instantly out.
 
Last edited:

SonGoku

Gold Member
Aug 16, 2018
2,028
1,713
320
>3GB of OS can be written to the SSD in under a second.

This makes a common mistake when talking about todays SSDs, as it implies everything would be a tidy sequential read/write. We don't live off of fast SSDs with no RAM because the fastest are still almost an order of magnitude slower on sequential than just plain old single channel DDR4 (Even DRAMless SSDs are kind of horrid), and that's without considering latency and random access speeds (the RA part!). Realistically you hit smaller files and less predictable access and cut your transfer rate down several times. So now how does 400MB/s for a mixed batch of files instead of 19200 sound, and add another order of magnitude in latency, it adds up when the OS is hitting it with millions of requests.

It's kind of just a truism that faster SSDs make OS paging less painful, but to run most of the OS and background apps off the SSD and keep loading/offloading them...I think that would at least cancel out how much faster the CPU would be for the OS speed, and I'm already impatient with the PS4 OS speed lol


>So much so that the idea of having an entire DDR4 controller just for supporting 4GB of OS RAM feels incredibly wasteful.


I mean, one of the great things of the 8th gen over the 7th was the ARM coprocessor allowing background updates and downloads, I assume the 9th will still have that, and if you have an ARM coprocessor it needs RAM...So, wasteful to have a tiny memory controller on the coprocessor? Needs it anyways


I thought most rumors were pointing to it having some DDR4 for the OS but none of us know, but I know which one would perform better for getting in and out of the OS, store, browser, etc. Plus 1GB in the main pool is less than we already allocate to the OS this gen, so that sounds instantly out.
Very interesting, thanks for the insight
Would fast NVMe mean less memory needs to be reserved for OS tasks for things such as video recording? I worry that OS memory consumption will rise yet again for the sake of new features, possibly 6-8GB! In that case a 4GB DDR4 pool would be of tremendous help to use as little GDDR6 memory as possible for OS.
 

xool

Member
May 29, 2018
317
211
210
Whats your take on these comments (from reera):
The PS4 OS needs around 1GB ram to store the 15mins of video capture data it continually records - at 4k this is going to at least double. (attempting to stream this to SSD will rapidly wear level the SSD, and reduce SSD bandwidth for games)

Additionally many OS functions need to stay in memory while game is running - ie when a notification pops, when you get an acchievement, when using friends, different types of chat, internet connections - all use OS in game.. needs to be in RAM.

4GB for OS makes perfect sense. - Resetera post hasn't thought this through. I'd bet real world cash on 4GB DDR4 or similar amount slow cheap RAM.
 
Last edited:

David Ricardo

Member
Sep 25, 2008
1,049
142
760
Gijon, Spain
more like $60-80 if my proposed chiplet solution would come true. $60 is around what MS must have paid for x1x APU when it launched.
Sony paid 100$ per APU in the base PS4 at launch.

MS also paid 100$ for the SOC on One S
https://ihsmarkit.com/research-analysis/microsoft-xbox-one-s-makes-a-play-for-the-big-time-for-only-24-more-teardown-shows.html

Where did you get that 60$ figure for the One X? It doesn't sound right for it to be cheaper than the One S SOC.
 
Last edited:

Insane Metal

Member
Mar 11, 2006
20,605
1,542
1,385
Br
HOLY SHIT

What does that even mean?

Edit: that's talking about memory speed... so not necessarily related to consoles. I mean probaby not at all.
 
Last edited:
  • LOL
Reactions: TLZ

SonGoku

Gold Member
Aug 16, 2018
2,028
1,713
320
The PS4 OS needs around 1GB ram to store the 15mins of video capture data it continually records - at 4k this is going to at least double. (attempting to stream this to SSD will rapidly wear level the SSD, and reduce SSD bandwidth for games)

Additionally many OS functions need to stay in memory while game is running - ie when a notification pops, when you get an acchievement, when using friends, different types of chat, internet connections - all use OS in game.. needs to be in RAM.

4GB for OS makes perfect sense. - Resetera post hasn't thought this through. I'd bet real world cash on 4GB DDR4 or similar amount slow cheap RAM.
Damn then i hope we get at least 4GB (hopefully more) of LPDDR4 for OS functions, that way it reserves as little main memory (GDDR6) as posible.
 

LordOfChaos

Member
Mar 31, 2014
8,615
582
575
Damn then i hope we get at least 4GB (hopefully more) of LPDDR4 for OS functions, that way it reserves as little main memory (GDDR6) as posible.

For the RAM and the APU too, I hope as little OS as is needed sits on the main system. The story is, I don't know if it's true, that the PS4 ARM coprocessor was intended to do more but ended up being too underpowered for a lot of it, so they had to keep ~1.5 x86 cores reserved for the OS for everything it's doing in the background. If the x86 cores could only run the OS APIs/libraries necessary and the game, there may also be disproportionate wins for cache coherency and more regular memory access patterns, because remember this:



Nowadays it would be hard to screw up getting an ARM SoC powerful enough to run background downloads, video recording, the UI, etc.
 

SonGoku

Gold Member
Aug 16, 2018
2,028
1,713
320
For the RAM and the APU too, I hope as little OS as is needed sits on the main system. The story is, I don't know if it's true, that the PS4 ARM coprocessor was intended to do more but ended up being too underpowered for a lot of it, so they had to keep ~1.5 x86 cores reserved for the OS for everything it's doing in the background. If the x86 cores could only run the OS APIs/libraries necessary and the game, there may also be disproportionate wins for cache coherency and more regular memory access patterns, because remember this:



Nowadays it would be hard to screw up getting an ARM SoC powerful enough to run background downloads, video recording, the UI, etc.
btw was the ARM CPU memory ddr or lpddr3?

edit:
I don't understand this phrase that's been going around lately: CPU bandwith reduces GPUs disproportionately.
I though the CPU/GPU had dedicated garlic/onion buses, and that CPU uses 20GB/s max.
 
Last edited:

LordOfChaos

Member
Mar 31, 2014
8,615
582
575
btw was the ARM CPU memory ddr or lpddr3?

edit:
I don't understand this phrase that's been going around lately: CPU bandwith reduces GPUs disproportionately.
I though the CPU/GPU had dedicated garlic/onion buses, and that CPU uses 20GB/s max.

I think effectively the different kinds of data access patterns of a CPU vs GPU make it harder for the memory controller to manage, predictable patterns are best, also the CPU has a 20GB bus to the GPU as well, maybe that's partly the CPU bandwidth in the slide, so the CPU talking to the GPU takes some of its memory controller performance.


First, we added another bus to the GPU that allows it to read directly from system memory or write directly to system memory, bypassing its own L1 and L2 caches. As a result, if the data that's being passed back and forth between CPU and GPU is small, you don't have issues with synchronization between them anymore. And by small, I just mean small in next-gen terms. We can pass almost 20 gigabytes a second down that bus. That's not very small in today’s terms -- it’s larger than the PCIe on most PCs!



This is the ARM SoC, looks like DDR3, 256 MB by the way

 
Last edited:
  • Like
Reactions: SonGoku

Negotiator

Member
Jun 28, 2011
3,576
194
510
For the RAM and the APU too, I hope as little OS as is needed sits on the main system. The story is, I don't know if it's true, that the PS4 ARM coprocessor was intended to do more but ended up being too underpowered for a lot of it, so they had to keep ~1.5 x86 cores reserved for the OS for everything it's doing in the background. If the x86 cores could only run the OS APIs/libraries necessary and the game, there may also be disproportionate wins for cache coherency and more regular memory access patterns, because remember this:



Nowadays it would be hard to screw up getting an ARM SoC powerful enough to run background downloads, video recording, the UI, etc.
Actually it's the AMD VCE (which resides in the Radeon GPU) that handles video recording on the PS4, not the ARM SoC in the southbridge.

I don't think it's possible to change that. It needs direct access to the main DRAM pool, where the framebuffer is. ARM has an entirely different memory pool, with different memory addressing/pointers.

Regarding the OS functions, I thought the PS4 OS was written in x86-64 and not ARMv7? Is it possible to have both at the same time?

And yeah, video recording directly in flash/NAND storage is a bad idea (due to wear levelling). Nobody does that, not even Nintendo (they allocate a ~20MB video buffer on Switch DRAM).

Modern SSDs are cheap, but not very durable (SLC is no longer the norm).
 
Last edited:

sinnergy

Member
Jun 16, 2007
2,109
16
855
Can't currently listen? And small highlights?
Video is rehash and amateurish imo. Nothing new and he sounds like a regular forum visitor, no new info in it just, fanboy wishes. But to me it has become clear the last few weeks that if we get anything above 9 TF, would be amazing, 14 TF is off the table with a console power requirement. The rest is wishful thinking.
 
  • Like
Reactions: Imtjnotu

bitbydeath

Gold Member
Nov 25, 2015
7,045
4,166
600
Video is rehash and amateurish imo. Nothing new and he sounds like a regular forum visitor, no new info in it just, fanboy wishes. But to me it has become clear the last few weeks that if we get anything above 9 TF, would be amazing, 14 TF is off the table with a console power requirement. The rest is wishful thinking.
You got that backwards. 14TF would be amazing. 9TF and below is off the table.
 
Last edited:
Status
Not open for further replies.