Next Generation Hardware Speculation with a Technical Spin [pre E3 2019]

[MrFox]

Hey I forgot to check, but my account already had access to the hbm2e doc. Not much new stuff but a few interesting things:

While products currently available are 2ch per die, vendors are free to design dies with 1, 2, or 4ch/die. So it's still possible to have fully populated channels with only 2 dies.

The additional base logic die is not part of the standard, vendors are free to remove it if they implements that logic into the memory dies. (low cost hbm?)

Die sizes can also be 6gb or 12gb. So it could even do intermediate stack capacites like 18GB or 9GB.

The total number of banks seems to depend on channel size and stack configuration. It's 8, 16, 32, or 48 banks per channel.

The stacks of 8 and 12 are addressed with a stack ID to select one of the group of 4. In practice it's just like more bank select bits. But that means higher stacks have more banks. So maybe better performance from bigger stacks?

 

[Tkumpathenurpahl]

Something just crossed my mind, and I'd like some input. Some context:

• This generation could probably go on longer than 2019/2020, thanks to the Pro and the X1X, but the base consoles would struggle to be AAA lead platforms for another 4-5 years. Unless they start facing situations like the Switch's 148x83 port of Ark.

• The limiting factor of the base consoles is widely agreed to be their CPU's. They were the best option at the time, but still poor.
• Console launches are expensive and risky.

So, my question is this: could the upcoming generation be designed to last longer?

I'll expand on this later, but right now, my dog really wants to play, and he keeps jumping on me with his squeaky toy, so I'm going to oblige.

 

[anexanhume]

Something just crossed my mind, and I'd like some input. Some context:

• This generation could probably go on longer than 2019/2020, thanks to the Pro and the X1X, but the base consoles would struggle to be AAA lead platforms for another 4-5 years. Unless they start facing situations like the Switch's 148x83 port of Ark.

• The limiting factor of the base consoles is widely agreed to be their CPU's. They were the best option at the time, but still poor.
• Console launches are expensive and risky.

So, my question is this: could the upcoming generation be designed to last longer?

I'll expand on this later, but right now, my dog really wants to play, and he keeps jumping on me with his squeaky toy, so I'm going to oblige.

They may not have a choice. Node revisions after 7nm are going to be very expensive, and we don't even have confirmation of an HPC 5nm variant yet.

 

[Megadrive1988]

Does anyone know the current status of HBM3 in general?

HBM3 was announced August 2016 at HotChips (Samsung, SK Hynix) which was expected to begin manufacturing in 2019 or 2020.

I did see this recent article from Tomshardware:

HBM2
HBM2 debuted in 2016. In December 2018, the JEDEC updated the HBM2 standard. It now allows up to 12 dies per stack for a max capacity of 24GB. The standard also pegs memory bandwidth at 307GB/s, delivered across a 1,024-bit memory interface separated by eight unique channels on each stack.

Originally, the standard called for up to eight dies in a stack (as with HBM) with an overall bandwidth of 256GB/s.
_____________________________________________________________________________
HBM3 and HBM4
While not yet available, HBM3 and HBM4 standards are currently being discussed.

HBM3 is expected to increase density over HBM2 with more dies in a stack, more density per die and higher efficiency. Transfer data rate is being pegged at 4GT/s, and bandwidth is expected to be 512GB/s to 1TB/s. According to an Ars Technica report, Samsung is planning production in 2019 or 2020.

HBM4, meanwhile, will follow, but there’s limited information on the specification available. There is a report from Wccftech claiming a transfer rate of 4TB/s, but it remains unconfirmed.
_____________________________________________________________________________

https://www.tomshardware.com/reviews/glossary-hbm-hbm2-high-bandwidth-memory-definition,5889.html

I do wonder if HBM3 or the future HBM4 might be relevant to potential mid-gen upgrades of PS5 and Scarlett aroun 2023-2024.

 

[goonergaz]

I'm not sure if I should post this here, feel free to delete or move...but a member at resetera (anexanhume) just posted an amazingly long and detailed post of all the info we have to date - it's so large it's over 4 posts!

"Mission Statement

The purpose of this post is to aggregate all the news, rumor, and good ole' speculation from the past several years regarding the release of next generation consoles. I have made a light attempt at editorializing and narrative construction to make the content more digestible and a little less dry in nature. To that end, the purpose of this post is not to predict what next generation consoles will be, but to collect the relevant items surrounding their respective releases in one place, so that we may have a reference for future discussions, and hopefully we spark a few new ones along the way."

https://www.resetera.com/threads/ps5-and-next-xbox-launch-speculation-post-e3-2018.49214/page-395#post-16530248
https://www.resetera.com/threads/ps5-and-next-xbox-launch-speculation-post-e3-2018.49214/page-395#post-16530262
https://www.resetera.com/threads/ps5-and-next-xbox-launch-speculation-post-e3-2018.49214/page-395#post-16530270
https://www.resetera.com/threads/ps5-and-next-xbox-launch-speculation-post-e3-2018.49214/page-395#post-16530281

He has also included references;

https://www.resetera.com/threads/ps5-and-next-xbox-launch-speculation-post-e3-2018.49214/page-395#post-16530287
https://www.resetera.com/threads/ps5-and-next-xbox-launch-speculation-post-e3-2018.49214/page-395#post-16530301

 

[Michellstar]

You could have asked him, look it´s just two post before you

 

[goonergaz]

You could have asked him, look it´s just two post before you

Fair enough, I mean - it's the internet, and there's no mention of restrictions...hell if I'd done the work I'd want people to share it out, but I will ask.

edit - permission granted so post updated with links

 

[AlNom]

I'm not sure if I should post this here, feel free to delete or move...but a member at resetera (anexanhume) just posted an amazingly long and detailed post of all the info we have to date - it's so large it's over 4 posts!

Have only had a cursory glance, but from what I understand (source = Rys), the following graph is actually misleading and very inaccurate (and old ).

https://u.cubeupload.com/anexanhume/7nmdesigncost.png

----

2x bonus points for "Cirrus Business". 1010/1010, would pun again.

----

Regarding Zen config - wonder if 16MB L3 would be the trade-off for a console target. MS had mentioned/hinted at plans for multi-purposing the hardware in the cloud, but 32MB L3 would maybe seem overkill for the console. Maybe. Would be more forward thinking to just roll with the fatter config though.

---

Kinda brought it up a while ago regarding GDDR6, but it operates with 2chan x 16-bit I/O, which may have (good) implications for bandwidth utilization and sharing on an APU configuration.

 

[anexanhume]

Have only had a cursory glance, but from what I understand (source = Rys), the following graph is actually misleading and very inaccurate (and old ).

That’s a fair assertion. It overestimates AMD’s own chart, which claims a 2x increase over 14/16nm. That’s why I provided both for context.

Regarding Zen config - wonder if 16MB L3 would be the trade-off for a console target. MS had mentioned/hinted at plans for multi-purposing the hardware in the cloud, but 32MB L3 would maybe seem overkill for the console. Maybe. Would be more forward thinking to just roll with the fatter config though.

I wonder this too, but it may not save enough on yield or die area to be material enough to care about. Having a bigger L3 would help the CPU tone down its DRAM accesses ever so slightly, given the GPU more consistent bandwidth.

Kinda brought it up a while ago regarding GDDR6, but it operates with 2chan x 16-bit I/O, which may have (good) implications for bandwidth utilization and sharing on an APU configuration.

It is a good point. I have the feeling that feature in the spec was specifically requested by AMD for that reason.

 

[AlNom]

I wonder this too, but it may not save enough on yield or die area to be material enough to care about. Having a bigger L3 would help the CPU tone down its DRAM accesses ever so slightly, given the GPU more consistent bandwidth.

True.

It is a good point. I have the feeling that feature in the spec was specifically requested by AMD for that reason.

Another side implication is that it might actually make for an interesting revision for Durango.

 

[mrcorbo]

@goonergaz Thanks for sharing, since @anexanhume chose to cheat on us with another forum

Seriously though, @anexanhume, that's some outstanding work there and I get why after putting that much effort in you'd post it in the place that it would get the most eyes-on and the most engagement.

WRT Backwards Compatability and why it's desirable and likely, you might have also included some of the stuff Phil Spencer said that I detailed in this post as lifted from this interview with Gamasutra. In general, digital libraries being on the come-up on console and the expectation created by the existing paradigm in other markets (PC, Mobile) of your digital purchases coming with you from device to device also impact the need for BC going forward.

I'm not sure what else I might have added, though. You were amazingly thorough. And the citations? <chef kiss>

 

[anexanhume]

True.


Another side implication is that it might actually make for an interesting revision for Durango.

I think it would be an interesting revision for either. A PS4 super slim with 7nm APU and only 4 GDDR6 chips could help them make a $200 box still profitable, increase their buy on GDDR6 parts, and function as a validation of their GDDR6 controller for a new console design.

@goonergaz Thanks for sharing, since @anexanhume chose to cheat on us with another forum

Seriously though, @anexanhume, that's some outstanding work there and I get why after putting that much effort in you'd post it in the place that it would get the most eyes-on and the most engagement.

WRT Backwards Compatability and why it's desirable and likely, you might have also included some of the stuff Phil Spencer said that I detailed in this post as lifted from this interview with Gamasutra. In general, digital libraries being on the come-up on console and the expectation created by the existing paradigm in other markets (PC, Mobile) of your digital purchases coming with you from device to device also impact the need for BC going forward.

I'm not sure what else I might have added, though. You were amazingly thorough. And the citations? <chef kiss>

Thanks for the kind comments. There are several reasons. First, posts are hard to format for forums. Second, the content covers a breadth of topics that would span no less than three current topics on this forum. Also, the level of information presented is well beneath the understanding of many of the regular participants in this topic, so I feel it would be much less additive comparatively. That forum also allows for a little more creative license. I try to keep my posts here more professional in nature and narrower in scope. I’m much more in a position to learn than to relay knowledge, so brevity serves me well.

Your insight on Spencer’s comments is something I’d absolutely agree with and think it’s good context to have. Each of the topics was somewhat of a struggle to keep to to the length that I did. That’s why I allowed myself a somewhat extemporaneous section near the end to get some ideas out that I didn’t want to let further balloon the sections they would have normally gone in.

 

[AlNom]

I think it would be an interesting revision for either. A PS4 super slim with 7nm APU and only 4 GDDR6 chips could help them make a $200 box still profitable, increase their buy on GDDR6 parts, and function as a validation of their GDDR6 controller for a new console design.

Cer'nly, although you'd be looking at more expensive GDDR6 bins to compensate for speed. Durango is in a different performance bracket.

edits:

It might make sense for a 4Pro revision since there's more to reduce in cost in an absolute sense when looking at power consumption & implications to the chassis. I suppose the other thing to consider is if they predict any shift in sales towards 4Pro since the market is somewhat saturated already with the base model, if that makes sense, so in that regard, it's a game of how much do they spend on R&D for slim2.0 @ the somewhat more expensive 7nm & GDDR6 parts.

hm.......

 

[Globalisateur]

...

---

Kinda brought it up a while ago regarding GDDR6, but it operates with 2chan x 16-bit I/O, which may have (good) implications for bandwidth utilization and sharing on an APU configuration.

Even if they use the clamshell mode ?

 

[anexanhume]

Even if they use the clamshell mode ?

They still have to be addressed separately. They can’t drive the data lines at the same time.

 

[MrFox]

I suppose gddr6 can be seen as the logical reverse of clamshell.

Clamshell is linking two chips together so that the controller only sees one.

Gddr6 is splitting a single chip so that the controller literally sees two fully independent chips with separate command, address and data lines.

It should reduce the stalls from collisions between cpu and gpu accesses to the same channels, since there's twice as many channels for the same bus width.

 

[anexanhume]

I suppose gddr6 can be seen as the logical reverse of clamshell.

Clamshell is linking two chips together so that the controller only sees one.

Gddr6 is splitting a single chip so that the controller literally sees two fully independent chips with separate command, address and data lines.

It should reduce the stalls from collisions between cpu and gpu accesses to the same channels, since there's twice as many channels for the same bus width.

Precisely. So if you’re running GDDR6 in a 128 bit bus on a system that’s used to a 256 bit GDDR5 bus at half the rate, you’re still seeing the same bandwidth per channel.

 

[vipa899]

How much memory do next gen consoles need as a minimum you think? Maybe if we look at how much ram curent games use we get an idea.
Only know of one exempel, thats killzone shadowfall, its supposedly using 3072mb for vram, about 2gb for game logic, audio, physics etc. How does that look like for GoW, Spiderman, lats of us 2, or any of the AAA games?
For shadowfall, 3gb seems alot, as the game doesnt look like a title with that much vram useage. Spiderman, HZD looking much better does see a much larger vram usage?

 

[Shifty Geezer]

That's floored thinking in many ways. Firstly, what does a 3 GB game look like? How can you tell VRAM requirements of a game just by looking? Secondly, if next-gen has SSD, RAM requirements for buffering will be far smaller. Thirdly, what's your minimum requirement to measure RAM usage against? What's your minimum world size, minimum AI complexity, minimum texture res, etc? I'd say current gen world detail with next-gen lighting at high res and framerate and next-gen worlds (AI, physics, interactivity) would constitute a suitable next-gen without needing gobs more RAM but needing gobs more power instead. If you had enough power, you could even procedurally generate everything and need even less RAM than we have now.

The decision is from a triangle of bandwidth, capacity and price, and you need to pick a target that balances them optimally. You can't focus on any of those variables in isolation. Hence a discussion looking at RAM techs and seeing what options are available and which would be the best for a console.

 

[Mobius1aic]

How much memory do next gen consoles need as a minimum you think? Maybe if we look at how much ram curent games use we get an idea.
Only know of one exempel, thats killzone shadowfall, its supposedly using 3072mb for vram, about 2gb for game logic, audio, physics etc. How does that look like for GoW, Spiderman, lats of us 2, or any of the AAA games?
For shadowfall, 3gb seems alot, as the game doesnt look like a title with that much vram useage. Spiderman, HZD looking much better does see a much larger vram usage?

Based on current top end GPUs, I think that gives a good scope for VRAM needs of the next generation.

12 GB seems plausible, maybe 16, plus 8 system memory. So 20 to 24 GB of total RAM is possible.