Middle Generation Console Upgrade Discussion [Scorpio, 4Pro]

[Deleted member 11852]

So if not for the poor connection between GPU and CPU on PC, we'd have PCs with 20+ GBs of unified GDDR5 accesed by both processors?

If there was never a necessity to develop a faster bespoke memory type (GDDR) and the opportunity to do this in isolation of DDR used by the CPU, it's likely DDR would just have developed in a different direction. Existing technologies generally evolve to solve new problems, new technologies (GDDR) rarely gain traction except in a problem vacuum.

 

[function]

So if not for the poor connection between GPU and CPU on PC, we'd have PCs with 20+ GBs of unified GDDR5 accesed by both processors?

I doubt it. You'd need a 384 bit bus and it would blow through any sensible power limit. More cost, more power, and a huge chunk of die taken up by a huge interface. For a super high end gaming PC (where cost and power aren't the primary concerns) it would make more sense, but for the other 99% of PCs out there it wouldn't appear to be a balanced approach.

GDDR5 doesn't fit with the current focus on power efficiency, DDR4 will never be fast enough for high end gaming, and HBM2 hasn't even appeared and yet its 32GB limit is started to look less and less enormous in the context of PC main memory. Perhaps DDR4 or LPDDR4 could be further developed into a reasonable compromise for most roles ...

There are such wildly differing demands on PCs in different roles and with different costs, that I think split memory pools are going to continue on for a while.

Edit: can GDDR5X be run in 16 wide mode, with two modules to 32 lanes?

 

[AlNom]

Edit: can GDDR5X be run in 16 wide mode, with two modules to 32 lanes?

Yup
https://www.micron.com/~/media/documents/products/data-sheet/dram/gddr5/8gb_gddr5x_sgram_brief.pdf

 

[function]

Yup
https://www.micron.com/~/media/documents/products/data-sheet/dram/gddr5/8gb_gddr5x_sgram_brief.pdf

So, x16 in clamshell gives you ... max of 32GB on a 256-bit bus at current densities.

 

[Deleted member 11852]

I doubt it. You'd need a 384 bit bus and it would blow through any sensible power limit. More cost, more power, and a huge chunk of die taken up by a huge interface. For a super high end gaming PC (where cost and power aren't the primary concerns) it would make more sense, but for the other 99% of PCs out there it wouldn't appear to be a balanced approach.

For faster bandwidth you can go wider and/or faster. We have narrow slow buses only because that is all they you need to sustain today's processors and I/O devices over PCI. And that is PCI's problem. If PCI never happened wider/faster buses would have developed to solve the GPU problem. A wider faster standard bus isn't that crazy, PS2 had a 2,048-bit wide bus and because it's wide you can partition lanes for specific low-latency access. No queuing!

There are such wildly differing demands on PCs in different roles and with different costs, that I think split memory pools are going to continue on for a while.

They will continue for as long as the PCI bottleneck remains. That's not gong anywhere in consumer gear anytime soon but not many are building server architectures around PCI now and that's a start.

 

[function]

For faster bandwidth you can go wider and/or faster. We have narrow slow buses only because that is all they you need to sustain today's processors and I/O devices over PCI. And that is PCI's problem. If PCI never happened wider/faster buses would have developed to solve the GPU problem. A wider faster standard bus isn't that crazy, PS2 had a 2,048-bit wide bus and because it's wide you can partition lanes for specific low-latency access. No queuing!

I certainly like the sound of that, and can see what you mean about PCI artificially limiting the need to have faster access to main memory. PS2 kept things on-chip though, where as the PC needs to be able to communicate off-chip with other, data hungry processors and off-chip memory pools. In that case, couldn't tasking one processor with feeding another data hungry processor be seen to have some power and complexity downsides to balance out the strong upside of having a single unified pool of memory?

This is a bit academic for the console side of things I suppose, as they will hopefully continue to use SoCs so you don't have the off-chip thing to worry about. At some point, isn't there always going to be a tension between power, cost, and speed though with differing memory standards for different roles?

 

[Silent_Buddha]

So if not for the poor connection between GPU and CPU on PC, we'd have PCs with 20+ GBs of unified GDDR5 accesed by both processors?

That's a case where a split memory pool is cheaper. It is more expensive to have a full GDDR pool of memory in use by both the CPU and GPU than it is to have DDR for the CPU and GDDR for graphics.

In less performance critical situations (most notebooks, all mobile "smart" devices, etc.) then a single pool of DDR is cheaper.

You can also see this in effect with cheaper discrete graphics cards where DDR is used instead of GDDR due to cost reasons.

Basically, once performance requires go above a certain level, a split memory pool is cheaper and more effective (no memory contention) than a single memory pool if there is a large cost discrepancy between performance (GDDR in this case) and commodity (DDR in this case) memory types.

Hence you see that while the majority of Notebook computers have a single pool of DDR (cheap and good enough for most generic tasks) gaming and hence performance oriented notebooks have split pools of DDR and GDDR memory due to performance reasons even though having GDDR split off is more expensive than just having a pool of DDR3. And having a unified pool of GDDR would be even more expensive than having split pools of DDR and GDDR, despite the added board complexity as well as potential for less performance than split DDR/GDDR memory pools.

Of course, going forwards there are some things that blur some of that. Heterogeneous computing has benefits it can leverage from a unified memory pool, for example. But does that outweigh potential memory contention between the GPU and CPU for gaming?

Regards,
SB

 

[GrimThorne]

I've seen rumored leaked specs of the 1080Ti having 12GB GDDR5X. Which seems totally plausible to me. And people think new Xbox will have 32GB HBM2?!

Well I don't even think it's even going to have 12GB GDDR5X. I'm thinking 16GB DDR4 with 4GB's dedicated to the OS/UI/dashboard.

 

[mpg1]

Well I don't even think it's even going to have 12GB GDDR5X. I'm thinking 16GB DDR4 with 4GB's dedicated to the OS/UI/dashboard.

Really? I think if they decide to use DDR4 it would be in combination with video memory. Something like 8GB DDR4 + 8GB GDDR5/X

 

[Silent_Buddha]

For the rumored XBO-T (Xbox One-Two), I could see them going with a more traditional split DDR/GDDR memory pool. Except with a fast path between the memory pools (rather than going through PCIE like a PC). Either that or I would imagine they'll bite the cost and go with a unified GDDR memory pool. There's also a potential for them to go with a split DDR/HBM memory pool, although I'd imagine that would be too expensive for a console.

I can't see them going with a massive unified DDR memory pool again, however. Especially not with the rumored increase in graphics performance. Although perhaps there's a potential for a unified DDR memory pool combined with a relatively small (say 1-4 GB) HBM graphics only memory pool. Not all graphics ops require fast memory, I believe.

Still I believe it's more likely it'll either be split DDR/GDDR or unified GDDR.

Regards,
SB

 

[Akumajou]

It's 1-2 financial quarters to get chips back from manufacturing, depending on what stepping is good enough. Then months to ship components, manufacture, and distribute for launch.

AMD probably wouldn't mind a little of that. It's got HPC APU slides with big GPUs and new cores that could be used as a basis, and a sequence of GPU architectures it might want to use.
If a war of designs results, it gets paid either way--with the caveat that it can only put so much work into false starts before its other projects become constrained.

Conspiratorially, AMD would benefit from a cascade of leaks that freaks out its console customers into paying into a succession of hand-over-hand design leapfrogs.

Indeed the rumor specs are confusing enough...one says 6Tf the other 10Tf and 32GB ram...even using dual density Blu Ray you're gonna need a blazing fast SATA III controller and SSD drive as a standard feature.

 

[Rangers]

10TF and 32GB RAM sounds like a pipe dream for sure.

http://www.anandtech.com/show/10389/amd-teases-radeon-rx-480-launching-june-29th-for-199

http://wccftech.com/amd-radeon-rx-480-polaris-10-graphics-card-leak/

Anyways, AMD R480X mooted, launching late June, 14nm Finfet, Polaris architecture, $199, 5.5 teraflops, low power 110-135W TDP. If nothing else sounds like a 6TF Scorpio would be pretty viable.

 

[Akumajou]

That AMD R480X at 150W TDP...is making me think about the old GeForce GTS 250 which also had 150W TDP at 55nm.

Basically that R480X looks to be an aggressive move because I'm assuming it will only require a single six pin connector and either match or destroy most of AMD R9 3X0 cards while running relatively (not literally) ice cooler, nearly silent and saving money on the electric bill.

Indeed a Xbone 6TF is definitely a reality that Sony should start worrying about.

Definitely a new image quality watermark is possible on future exclusives that may mark current Xbone obsolete.

Console may even be smaller...Microsoft better put a full blown SATA III and include SSDs to sweeten the deal and they will have a monster...

Plus if the cpu is Zen...that helps. Still not sure about 32GB unless they believe they can actually do native 4K games.

 

[Rangers]

Big question for me is if they are going to show Scorpio at E3. Personally I doubt it. Always bet the delay in videogames.

Now we have rumors Nx isn't going to launch in early 17 (it's already seen one delay). I'll bet PS4 Neo wont make early 17 either (it will be late 17).

Video games are always delayed.

Maybe we will get some sort of very vague tease IMO.

 

[Shortbread]

I'll bet PS4 Neo wont make early 17 either (it will be late 17).

Nope. Sony is going after this years holiday season with PSVR and NEO for sure. Makes no sense on not doing so...

Sure MS could have better hardware within late 2017, but Sony will more than likely have an uncontested user-base of 60-65 million when that happens. That's more impressive to a developers/publishers, than new hardware.

 

[BRiT]

Can we please not turn this thread into threads that already exist, such as Console Sales or the What should Sony or MS Do Next?

Kthxbi!

 

[rockaman]

It all comes down to One!

I mean Two!

 

[Shortbread]

Hrm, that would be interesting to see how the market works out assuming all 3 models are software compatible. Curious how the market would act with the following lineup:

Xbox One at $200-250 for 1.xTF.
Xbox One Two at $400 for 5-6TF.
Xbox One Two Three at $700-$750 for 10-11TF.

Hum...

 

[AlNom]

Hum...

So good, it'll melt his body.

 

[iroboto]

Heat kills everything. If you want the console to last we have to be realistic about its thermals.

Xbox One in particular is about quiet functionality and designed to be always on for 10 years.

Xbox One with its TV capabilities are locked to that design goal for its lifetime, only Scorpio could be designed with other more gaming focused intentions.

MS' play should be to make XBO as cheap as possible, to get users to enter the ecosystem or return to their Eco system.

Any consoles added subsequently, whether Scorpio or the one after Scorpio will operate off the XBO library as well. Ideally.

That being said the goal is to get consumers to buy into Slim and begin building their library. When they are not satisfied with the performance of the console anymore they can upgrade to a more powerful version.


Sent from my iPhone using Tapatalk