Next Generation Hardware Speculation with a Technical Spin [2018]

Status
Not open for further replies.
screen capture if the job offer is taken down

v6d4iZm.jpg


0EzsnJ1.jpg

9RJiZBl.jpg
 
Parts of HDMI 2.1 features can be used in HDMI 2.0. IMO, VRR is just adaptive sync [which is used in AMD's Freesync over HDMI].

That was never the question... All HDMI 2.0b devices can use some of the future HDMI 2.1 features.
That includes Xbox One X, Xbox One S, PS4 Pro, but also should be able on PS4 due to its DisplayPort to HDMI bridge (The reason we can have HDR on this console). The Original Xbox One is pure HDMI 1.4 and cannot have any of those. So it is limited to Freesync V1.
But thats does not make the HDMI 2.1 since not all features are there. And the bandwidth has no change!
 
Not much relevant because consoles do not use 4:4:4 on 4K.

HDMI 2.1 will be important for PC users much more than for consoles.
This generation doesn’t, but we’re talking next gen, aren’t we? :)

Woohoo! (It really is a no-brainer as per this discussion). Might not be a full SSD and might go with some fairly voluminous fast flash, but at least they recognise the need for fast main storage. And if this is fast enough, RAM pressure will be reduced.

I would guess they are trading all options still given the DRAM job posting too. Interesting they don’t mention HBM by name, though.
 
If 16GB is enough, would HBM be more practical?

If I'm not mistaken, 16GB of GDDR6 would result in 576GB/s of bandwidth if using 18gb/s memory. The X1X seems to do well with its 326GB/s, so presumably, with a doubling of power, it would be preferable to double bandwidth too?
 
If 16GB is enough, would HBM be more practical?

If I'm not mistaken, 16GB of GDDR6 would result in 576GB/s of bandwidth if using 18gb/s memory. The X1X seems to do well with its 326GB/s, so presumably, with a doubling of power, it would be preferable to double bandwidth too?

I don’t know that 16GB wouldn’t have the same cost base as 32GB. I think they can do 16GB in one stack with 8 high, so you’re just adding the cost of those die and interconnects to go from 2 stack 16GB to 2 stack 32GB, not another entire stack which adds more controllers at the bottom and more vias to the interposer.
 
I don’t know that 16GB wouldn’t have the same cost base as 32GB. I think they can do 16GB in one stack with 8 high, so you’re just adding the cost of those die and interconnects to go from 2 stack 16GB to 2 stack 32GB, not another entire stack which adds more controllers at the bottom and more vias to the interposer.

My thinking is that low cost HBM is stated to have a bandwidth of 200GB/s per stack.[1] So four stacks, each 2 high, would give 800GB/s bandwidth - enough versus the 576GB/s provided by the same amount of GDDR6.

This next part, I don't know, although I'm really curious to know if I'm right: I reckon 2-high stacks of HBM must have a better yield than 4-high stacks, which, themselves, must have a better yield than 8-high stacks.

If that's the case, 4 2-high stacks would put them in good stead for future revisions of the console, which could replace them with one stack of HBM3 for a slim, or multiple stacks for silly bandwidth in a mid-generation iteration. [2]

[1] https://www.extremetech.com/gaming/...tions-to-blow-the-doors-off-the-memory-market

[2] https://wccftech.com/hbm3-ddr5-memory-early-specification-double-bandwidth/
 
I still think SSD is somewhat impractical for consoles. Consoles need to be cheap but also require high storage. The next gen will probably need 2TB as the starting point with the ability to still go higher (which is a key point).

In my mind you'd use something like 3d Xpoint (Intel's Optane) memory which is sort of a DRAM/NAND flash hybrid that could act as a high capacity cache for a traditional 2TB-4TB hard drive.

You'd probably only need 32GB of it and right now at retail it that only costs about $60. A couple years from now at wholesale level you'd probably only be looking at $30-$40 at most.

So $40 + cost of traditional 2TB or higher. Or cost of 2TB SDD or higher?....hmmmm seems pretty obvious to me.
 
screen capture if the job offer is taken down

v6d4iZm.jpg


0EzsnJ1.jpg

9RJiZBl.jpg
Interesting they mention gddr6, but not hbm.

At these frequencies, you need some really hardcore engineers. There was an article saying 20ghz is reaching the limit of what's possible on a standard pcb. Samsung might be offering 18gbps parts, and rambus have a gddr6 phy to sell, but there is still the pcb to design inbetween.
 
If 16GB is enough, would HBM be more practical?

If I'm not mistaken, 16GB of GDDR6 would result in 576GB/s of bandwidth if using 18gb/s memory. The X1X seems to do well with its 326GB/s, so presumably, with a doubling of power, it would be preferable to double bandwidth too?

The XboneX may have an overkill of memory bandwidth to achieve "perfect" compatibility with the Xbone's EDRAM bandwidth.

Regardless, with HBCC I doubt they would go with more than 8GB HBM. 2x 4-Hi HBM2 stacks at 2.4Gbps would result in 614GB/s total. Above that they could just use e.g. another 16GB of 128bit DDR4.


My thinking is that low cost HBM is stated to have a bandwidth of 200GB/s per stack.
I wonder where HBM Low-cost is going to.
With HBM2 already doing 307GB/s per stack, we're at a point where 2 stacks of HBM2 will get as much bandwidth as 3 stacks HBM Low-cost.
The price for HBM Low-Cost had better be ~50% the price of regular HBM2 or less, otherwise the price for it won't compensate given the additional cost for a larger interposer.
 
Interesting they mention gddr6, but not hbm.

At these frequencies, you need some really hardcore engineers. There was an article saying 20ghz is reaching the limit of what's possible on a standard pcb. Samsung might be offering 18gbps parts, and rambus have a gddr6 phy to sell, but there is still the pcb to design inbetween.

SerDes is pushing the 40Gbps data rates, and it can be mind-boggling how much compensation they have to apply to the channel to close the link. At that point you almost have to rope in RF designers to model your link rather than rely on signal integrity tools’ approximation of your stackup.

I wonder where HBM Low-cost is going to.
With HBM2 already doing 307GB/s per stack, we're at a point where 2 stacks of HBM2 will get as much bandwidth as 3 stacks HBM Low-cost.
The price for HBM Low-Cost had better be ~50% the price of regular HBM2 or less, otherwise the price for it won't compensate given the additional cost for a larger interposer.

Can anyone shed light and how much lateral mobility MS or Sony would have in adapting a solution? Can they build GDDR6 now and transition to HBM later? Go from HBM2/3 to LCHBM or vice versa? I’m curious what the sticking points are besides bus width.
 
Last edited:
https://www.resetera.com/threads/mi...-gddr6-for-future-xbox-design-projects.37207/

GDDR6 for next Xbox via a job offer from Microsoft :) it begins.. Yeah

other job offer
oh well, I expected HBM2 but anyways... BC means they are more likely going with AMD again, let's see what they come up with now that their main engineer in the GPU division isn't working for AMD anymore. HDD checked, flash memory checked..., the intriguing part is how much space a SSD could have by then and be profitable at the same time, but this has been discussed in the thread already..
 
In my mind you'd use something like 3d Xpoint (Intel's Optane) memory which is sort of a DRAM/NAND flash hybrid that could act as a high capacity cache for a traditional 2TB-4TB hard drive.

You'd probably only need 32GB of it and right now at retail it that only costs about $60.
How long will it take to load 32GBs of data from that HDD?
 
A little longer at game launch, but not so if most data is nearly together. What makes HDDs slow is seek speeds way more than read speed. During gameplay, data can be prefetched a bit ahead of time just like its done today.

Its not like current consoles have supper fast HDDs right now anyway. There is a lot of room for improvement by just enforcing a higher minimum speed as standard on the next machines before going all the way to SSDs. Its probably considerably more economically feaseble.

Having two replaceable HDs in there, one of which of a reasonably sophisticated type... Just doesn't sound like a mass consumer product to me. I can see it in a pro model, but not on the base.
 
I wonder if it'd just end up being a specialized form of SSHD?
¯\_(ツ)_/¯
 
Last edited:
Status
Not open for further replies.
Back
Top