Predict: Next gen console tech (9th iteration and 10th iteration edition) [2014 - 2017]

[Globalisateur]

Ok as of now it's fair to say it's going to be very hard to ship a console with 64GB of HBM in 2020. But seeing the Pro using 1GB of ddr3 we could have for instance 32GB of HBM dedicated to games and 8GB of ddr3/4 for the OS.

Could the OS be completely isolated on the slow DDR with a sufficiently powerful ARM ?

 

[iroboto]

Ok as of now it's fair to say it's going to be very hard to ship a console with 64GB of HBM in 2020. But seeing the Pro using 1GB of ddr3 we could have for instance 32GB of HBM dedicated to games and 8GB of ddr3/4 for the OS.

Could the OS be completely isolated on the slow DDR with a sufficiently powerful ARM ?

????? When did we skyrocket to 32GB?
Current top of the line video cards are at 4GB. You'd need perfect exponential growth to get to 32GB in 3 years.

I don't see anything on the roadmap that would indicate that the technology is on that path.

Furthermore, im worried about the TDP Rumour values for Vega as per the other threads, will wait and see the power draw, but if it turns out to have terrible performance/watt then I don't see anything happening anytime soon in the console space.

 

[fehu]

There's a thread dedicated just to "Is 4GB enought?"

Anyway, with virtualization it must be almost trivial to dedicate a separated memory pool just to the os.
I'm not a fan of this scenario, but this generation I've mistaken the prediction of almost any single technical aspect.

 

[Blazkowicz]

Furthermore, im worried about the TDP Rumour values for Vega as per the other threads, will wait and see the power draw, but if it turns out to have terrible performance/watt then I don't see anything happening anytime soon in the console space.

Perhaps predictably they'll do a new Fury X : push voltage and clock up to match (roughly) its competition. That sells the flaghship GPU to the half-kilowatt gamers with overvolted CPU at least.
For a console there's more incentive and less downside to put the voltages and clocks at a better, lower level and anyhow a PS5 etc. will likely be on 7nm process with the Vega's successor.

 

[mpg1]

A lot of the memory speculations seem way out of whack...I wouldn't be surprised if PS5 had 16GB (albeit at 400-500GB/s bandwidth).

 

[Prophecy2k]

????? When did we skyrocket to 32GB?
Current top of the line video cards are at 4GB. You'd need perfect exponential growth to get to 32GB in 3 years.

I don't see anything on the roadmap that would indicate that the technology is on that path.

Furthermore, im worried about the TDP Rumour values for Vega as per the other threads, will wait and see the power draw, but if it turns out to have terrible performance/watt then I don't see anything happening anytime soon in the console space.

Aren't you forgetting that the existing PS4 launched with 8GB in 2013. The XB1X will have 12GB in 2017.

Why are you looking at top of the line video cards on PCs which also have a separate pool of GBs worth of cheap system RAM?

 

[Prophecy2k]

A lot of the memory speculations seem way out of whack...I wouldn't be surprised if PS5 had 16GB (albeit at 400-500GB/s bandwidth).

We said this before and then Sony launched the PS4 with 8GB of GDDR5.

Samsung is looking to make low-cost, low-power HBM available by 2019. So the 32GB low-cost HBM speculation is based on just two stacks of this. With an organic interposer, it could possibly be a cost effective solution for consoles in 2019/20... maybe... if the stars align

 

[AlNom]

"low-cost" needs a comparison to conventional DRAM setups.

 

[iroboto]

Aren't you forgetting that the existing PS4 launched with 8GB in 2013. The XB1X will have 12GB in 2017.

Why are you looking at top of the line video cards on PCs which also have a separate pool of GBs worth of cheap system RAM?

It was in reference to HBM on chip. I'm not looking at it, I just saw a prediction of 32GB of HBM for PS5. I don't think we will be on HBM to be honest. It's ideal, but the technology doesn't look like it's there yet.

 

[Kaotik]

It was in reference to HBM on chip. I'm not looking at it, I just saw a prediction of 32GB of HBM for PS5. I don't think we will be on HBM to be honest. It's ideal, but the technology doesn't look like it's there yet.

Actually the current high end has 16 GB of HBM2, either in 4-Hi (P100, (V100 isn't really out yet so take it or leave it)) or 8-Hi (Vega FE)

 

[iroboto]

Actually the current high end has 16 GB of HBM2, either in 4-Hi (P100, (V100 isn't really out yet so take it or leave it)) or 8-Hi (Vega FE)

Correct, I should clarify my stance to consumer grade gaming cards, as that's what I was thinking along as opposed to machine learning hardware.

 

[milk]

Considering the fact that ps4pro has a separate pool of cheap/slow memory for OS and aparently that didn't become such a big overhead in design complexityas I've heard claimed before, wouldn't the smart design solution be to have LOADS of the cheap memory for caching virtual texture and asset streaming, which are big eaters of memory but don't need fast access all the time, and have a cosr effextive 8 or 16GB lightning fast pool dedicated only for what really needs it for? I feel like even the ps4 could have been better If they had gone that route. Just 2 or 4 Gigs of GDDR5 might have been enough to store the data that actually uses its almost 200GB/s while with the saved money they could put an extra full 8Gigs of cheap ram ON TOP of that.
The actual reason sony did not go that rout in my opinion was to give devs the message they were really considering ease of develoment this time.

 

[iroboto]

Considering the fact that ps4pro has a separate pool of cheap/slow memory for OS and aparently that didn't become such a big overhead in design complexityas I've heard claimed before, wouldn't the smart design solution be to have LOADS of the cheap memory for caching virtual texture and asset streaming, which are big eaters of memory but don't need fast access all the time, and have a cosr effextive 8 or 16GB lightning fast pool dedicated only for what really needs it for? I feel like even the ps4 could have been better If they had gone that route. Just 2 or 4 Gigs of GDDR5 might have been enough to store the data that actually uses its almost 200GB/s while with the saved money they could put an extra full 8Gigs of cheap ram ON TOP of that.
The actual reason sony did not go that rout in my opinion was to give devs the message they were really considering ease of develoment this time.

If I understand correctly, Virtual Texturing methods, either hardware or software based, significantly improve with lower latency. This is most obvious with texture pop in for virtual texturing solutions that happen when you teleport or move into a zone that is too complex for asset streaming to keep up. The move to high speed memory would retrieve these assets significantly faster and probably allow for finer grade streaming of assets. The larger the ram pool, the more of the area you can load, making pop in less noticeable. HBM appears is a good fit for virtual texturing. It's large enough to hold a large number of textures for retrieval (and load/unload) what they need from slower RAM (and ram from HDD). But it takes optimization effort.

The high bandwidth slow memory pool which we have today seems to have won out by just loading a shit ton more lol. But it would appear the costs are being put back onto the consumer to have more and more ram to support a larger playing area.

 

[DieH@rd]

A lot of the memory speculations seem way out of whack...I wouldn't be surprised if PS5 had 16GB (albeit at 400-500GB/s bandwidth).

We already know that GDDR will come in 2GB per chip packages in 2018. If Sony repeats 16 chip approach, that will be 32GB with up to 512GB/s speeds.

 

[turkey]

Is ddr4 too slow for texturing if it's not in contention with the CPU and not being used for render targets or other GPU work?

 

[iroboto]

Is ddr4 too slow for texturing if it's not in contention with the CPU and not being used for render targets or other GPU work?

Perhaps it is not substantially faster than GDDR5 but substantially less bandwidth.

Scratch pads/on chip memory on the other hand have that distance advantage that is incredible.

 

[itsmydamnation]

if HBCC is real and works well why do we need 1 pool of really fast ram? Why not say 2x 2 Hi HBM stacks and then 16/32gb of DDR4/5 ? Remember AMD have already demo'd this by limiting Vega's allow memory to 1 or 2gb ( cant remember) and then showing a gaming demo HBCC on/off.


edit:
http://digiworthy.com/2017/02/28/amd-radeon-rx-vega-hbcc/

AMD deliberately capped the amount of addressable memory on an HBCC-aware system to only 2GB versus 4GB addressable by the non-HBCC-aware system. The HBCC-aware system purportedly output gains of up to 50% average FPS and 100% minimum FPS than the other non-HBCC-aware system.

so HBCC had less local memory and was still significantly faster.......

 

[iroboto]

if HBCC is real and works well why do we need 1 pool of really fast ram? Why not say 2x 2 Hi HBM stacks and then 16/32gb of DDR4/5 ?

That's the key I guess. Without substantial testing by developers we aren't sure what will break it, what it's strong at.

Sounds like a good fit for consoles though.

 

[McHuj]

We already know that GDDR will come in 2GB per chip packages in 2018. If Sony repeats 16 chip approach, that will be 32GB with up to 512GB/s speeds.

I agree unless there's a need for bandwidth on the order of 1 TB/sec, they'll stick with GDDR6.

 

[dobwal]

In 4 years we will have moved from 8 GBs to 12 GBs of RAM on consoles, which is a 50% gain. But somehow we going to go from 50% gains in the last 4 years to 150% in the next 3-4 years. Seems like a stretch to me.

Even in the PC gaming space, 8 GB cards in the form of a Hawaii based card showed up in 2014, but if I'm not mistaken, we are at just 12 GB in three years for gaming tailored hardware.

You see 16 GB HBM on a couple of GPUs but those are workstation cards or chips designed for data centers or super computers. HBM2 is suppose to lower costs but those cost savings haven't naturally led to greater utilization in the gaming hardware space.