Next Generation Hardware Speculation with a Technical Spin [2018]

Status
Not open for further replies.
Tiled resources are an ideal fit. You can also use it more effectively to procedural content creation, which can be created, saved, and streamed very efficiently, rather than constantly procedural creating in realtime or having to save and load as monolithic textures with super slow seek-times between each texture.

And that's off the top of my head. The moment the tech is there, it's going to be used. same as putting in 32 GBs RAM. If the RAM is there, devs will find a way to use it. If the storage is there, devs will find a way to use it. And if the fast storage is there, devs will find a way to use it. What flash provides is a good compromise between RAM and HDD, offering latency which cannot be address any other way beyond more RAM which is both cost prohibitive and volatile.

For the example of tiled resources.... would this really help vs and hdd? wouldn't you just call the tile loads earlier on an HDD vs SDD?
I'm assuming you are talking about loading upcoming resources vs augmenting tiled video memory which even SDD/3d xpoint wouldn't be fast enough. (Again I say this based on the general consensus in the 3d xpoint thread)

Of course if its present someone will find a way to use it, however its a question of best use of $$$. However the $$$ spent may be more useful on something else like cooling for higher clock which would be useful all the time vs only specific scenarios.
 
@BlackAngus I can't find the exact reviews I saw before right now, but here's one by DigitalFoundry where the SSHD gets nearly the performance of a pure SSD in some circumstances. In other situations it's mid-way between the two.

There is a larger caveat with this as it's from 2016 so it is not on an Xbox One X, where the internal stock drive is already 50% faster the Xbox One S, so don't get too hung up on the Stock vs SSHD, but look at the SSHD vs SSD.

https://www.eurogamer.net/articles/digitalfoundry-2016-should-you-upgrade-your-xbox-one-with-an-ssd

upload_2018-7-6_8-29-52.png


upload_2018-7-6_8-31-28.png


upload_2018-7-6_8-40-34.png
 
Last edited:
For the example of tiled resources.... would this really help vs and hdd?
Yes.
wouldn't you just call the tile loads earlier on an HDD vs SDD?
The more preloading you do, the more RAM you need to store it, which is expensive. The intention of tiled resources an streaming is to reduce memory requirements by shifting the data to storage. The faster your storage is, the more variety of tiled resources you can fit in memory and less memory you need, either at less system cost or enabling more other stuff in memory.

However the $$$ spent may be more useful on something else like cooling for higher clock which would be useful all the time vs only specific scenarios.
The specific $$$ balance here is RAM v Storage. We need both. What's the optimal balance with the option of fitting flash somewhere between (or instead of storage). Gimping on RAM or storage to up clock-speeds is going to create an imbalanced system with a tragic storage bottleneck, though of course over-investing in a fancy storage solution would compromise system processing power and what you can do with all that data.

There is a larger caveat with this as it's from 2016 so it is not on an Xbox One X, where the internal stock drive is already 50% faster the Xbox One S, so don't get too hung up on the Stock vs SSHD, but look at the SSHD vs SSD.
For me, this highlight how caching is unpredictable and how a better system would get more value from the same flash. Compare the GTA5 load which is a massive improvement on the SSHD not far off the SSD performance, to the Fallout Vault 111 where SSHD performs closer to HDD than SSD. A system level flash cache+scratchpad seems a very sound idea for giving the best of all worlds.
 
I'm there as well, in having a full control cache area. But unfortunately it all comes down to costs to Sony/MS to see if everyone can benefit. If its a choice between normal HDD and SSHD for minor differences I hope they go SSHD if they're not going to have a fast cache area. It certainly wont be a choice between HDD and SSD for bulk storage.

While I enjoy my 2TB external SSD on my console, I know it isnt being used to the real advantage if the developers could design an entire engine and game around having the speed and fast i/o with low latency.
 
What scenarios would a (say) 16/32 GB SSD really be helpful for? (Thinking this size from a cost perspective)
Why would it be 16 to 32GB, if 128GB SSD NVMe drives that reach ~2.1GB/s reading speeds are selling right now for $45 for the consumer and the price-per-GB for MLC solid state has been steadily declining and is planned to further reduce in the following years?
I find it hard they would ever use less than 128GB in a 2020 console, and that would easily fit entire games. And if today's consoles let us swap their 2.5" SATA HDD, this "solid state cache" could also be replaceable through a M.2 socket.
 
I think that's a very clear direction IMO. Design your $400, then make it $450 and add $50 of flash cache/scratchpad. That'll be sooo much better value than a $400 HDD only SKU, and provide better storage than a $400 only 256/512 GB SSD console.

Although I still entertain the idea of selling a small SSD console for $400 with easy storage expansion options (HDD bay, external storage, flash drives and SD cards), or even $450 with a small SSD, additional storage options, and $50 extras of quality cooling, bit more power. I'd even allocate $2 per console for using Arctic Silver instead of mouse turds. ;)
 
No way.
eDRAM or eSRAM in larger capacities are an area hog that is immensely best spent on more execution units and a wider memory controller, as the Xbone -> Scorpio transition showed.
Another example are the Intel Iris Plus with the eDRAM that are easily beaten by the integrated Vega 8/11 on Ryzen APUs with a fraction of the eDRAM's bandwidth using regular DDR4, compared at similar power levels.

What was interesting to me is how much the eDRAM benefited games when discrete GPUs were in use.
 
Why would it be 16 to 32GB, if 128GB SSD NVMe drives that reach ~2.1GB/s reading speeds are selling right now for $45 for the consumer and the price-per-GB for MLC solid state has been steadily declining and is planned to further reduce in the following years?
I find it hard they would ever use less than 128GB in a 2020 console, and that would easily fit entire games. And if today's consoles let us swap their 2.5" SATA HDD, this "solid state cache" could also be replaceable through a M.2 socket.

Being conservative on the 16-32, as I think spending more would almost alway be way better use of $$$ to put in more CPU/GPU/RAM.

Do you think they will spend 10% of the console cost on storage for cache only?
To me that seemed like a bit much, yes the storage costs will decline over time, but even 5% seems like a bit much considering you still need to add cost for HDD and RAM is king over both and is pretty expensive.
 
Wouldnt 50$ in GPU be alot more useful?

+1 on the AS vs mouse turds =)

On this note its all about cost to performance balance, have we ever seen a BOM with costs for a previous gen console?
I would love to see a price breakdown.
 
Here is a BOM estimate for PS4 that seems to be inline with most of the data I have found. (Please add any corrections if there is something missing or glaringly incorrect)

Code:
Mats:                     Cost:    Percent:
CPU/GPU:                    100       26.25
DRAM:                        88       23.10
Power Supply:                20        5.25
Optical Drive:               28        7.35
Hard Drive:                  37        9.71
Mechanical/Electro:          35        9.19
Other Components:            40       10.50
Console Sub:                348       91.34
Controller:                  18        4.72
Packaging:                    6        1.57
Total for Mats+Pack:        372       97.64

Wow that formating came out badly... any way to make a nice table in the forums?
 
Last edited by a moderator:
Being conservative on the 16-32, as I think spending more would almost alway be way better use of $$$ to put in more CPU/GPU/RAM.

Do you think they will spend 10% of the console cost on storage for cache only?
It won't be just a cache. It'll be somewhere between RAM and HDD, providing a unique massive, fast storage and a very responsive system. If they put in more flash rather than just a small cache.

Wouldnt 50$ in GPU be alot more useful?
It's all relative. If $100 gets you 10 TF, and $150 gets you 15 TF, you get a 50% improvement. Which results in something on screen that'll be hard to perceive as 50% better, probably. If instead you're looking at $50 of 6 TF versus $100 of 12 TF, that's significant.

$50 of flash gets you 10,000 IOPS instead of 100 on an HDD - a 100x improvement. It gets you load times 1/4 of those off HDD. It nets you 64 GBs of fast work space instead of 32 GBs of RAM, meaning lots more room for storing your persistent, progressive data. It means opening apps is instantaneous instead of taking a few frustrating seconds.

Which is better depends on what your market prefers. And of course all the stats can be presented in different ways to make different arguments.

The real take home is it's not possible to gauge the value of a decent SSD in a console until someone's tried it. ;) But just thinking of it as faster load times for games would totally undervalue it. And oif that's all it ends up doing, yeah, more GPU would probably be the better choice.

On this note its all about cost to performance balance, have we ever seen a BOM with costs for a previous gen console?
I would love to see a price breakdown.
There are plenty if you search for BOM and teardown. Already posted in this thread is the PS4 SOC+RAM costing $188. (https://www.engadget.com/2013/11/19/ps4-costs-381-to-make-according-to-hardware-teardown/)
 
It won't be just a cache. It'll be somewhere between RAM and HDD, providing a unique massive, fast storage and a very responsive system. If they put in more flash rather than just a small cache.
..
The real take home is it's not possible to gauge the value of a decent SSD in a console until someone's tried it. ;) But just thinking of it as faster load times for games would totally undervalue it. And oif that's all it ends up doing, yeah, more GPU would probably be the better choice.

There are plenty if you search for BOM and teardown. Already posted in this thread is the PS4 SOC+RAM costing $188. (https://www.engadget.com/2013/11/19/ps4-costs-381-to-make-according-to-hardware-teardown/)

My original point was a hybrid drive doesn't seem very valuable as you cannot really use it in a granular fashioin like dedicated flash. I agree that a dedicated flash scratch pad may be useful (wanted to hear from some devs on how they think it could be useful).
 
What was interesting to me is how much the eDRAM benefited games when discrete GPUs were in use.
But AFAIK eDRAM was only ever used in one system with a discrete GPU (X360), and that was almost 13 years ago when the main GDDR3 memory could only provide 22.4GB/s.
With GDDR6 they could achieve 20x more bandwidth out of a 256bit bus using 8 chips, or a similar bandwidth using 2 chips of HBM2. Or 28x more bandwidth with 2 stacks of HBM2. This already on the level of an L3 cache and wider than any eDRAM implementation I know of.

Being conservative on the 16-32, as I think spending more would almost alway be way better use of $$$ to put in more CPU/GPU/RAM.
Using fast SSD cache would mean a contribution to the memory subsystem, and the question here is if it wouldn't be significantly more cost effective than spending more money on RAM alone, given the current market trends.
The idea isn't new. Both Intel and AMD are implementing this, the first with Optane and the second with Fuzedrive/StoreMI.


Do you think they will spend 10% of the console cost on storage for cache only?
Spending 10% of ~$400 would mean that in 2020 the cost for volume 128GB NVMe drives will actually go up during the next 2 years, which seems unrealistic at best.
Come 2020 the cost for 128GB NVMe should be closer to $20 than to $40.
 
Using fast SSD cache would mean a contribution to the memory subsystem, and the question here is if it wouldn't be significantly more cost effective than spending more money on RAM alone, given the current market trends.
The idea isn't new. Both Intel and AMD are implementing this, the first with Optane and the second with Fuzedrive/StoreMI.
Fully understand however 3d Xpoint (optane) is an order of magnitude or two faster than SSD from a latency perspective, so it starts in a much more useful place than SSD especially if used in the dimm form factor vs NVMe. (Depending on use of course)

I think the question on SSD is - in games is there a use case that having storage 100x faster than HDD but 100x slower than ram is really useful and brings significant benefit? What does this allow developers to do that they cannot do today?

Spending 10% of ~$400 would mean that in 2020 the cost for volume 128GB NVMe drives will actually go up during the next 2 years, which seems unrealistic at best.
Come 2020 the cost for 128GB NVMe should be closer to $20 than to $40.

I did say in the last after you quoted that it would likely be around 5% + the cost of normal mass storage, which is inline with your number =)
 
Last edited:
Optane is also byte addressable and random read latency is an order of magnitude better. I don't see Sony paying for it though, 3GB/s NVMe should be plenty to augment main memory and texture streaming doesn't seem to be that random in nature (or is it?).
 
Status
Not open for further replies.
Back
Top