Megadrive1988
Veteran
I don't believe Sony would go with less than an 8c/16t Zen.
Predict: Next gen console tech (9th iteration and 10th iteration edition) [2014 - 2017]
- Thread starter Shifty Geezer
- Start date
- Status
what do you suspect would be the approximate mid-range CPU in the PC space at the time of delivery for PS5?
I can't imagine they would go with more as well. 8c/16t may be the right spot; however, a 6c/12t with a higher clock may also be viable.
phoenix_chipset
Regular
Doubt they would go with less than 8 cores, they're not going to go backwards after giving developers that many to work with. Well they have 7 to work with anyway 
I wouldn't expect hyper threading though considering 360 is the only console that has it. They should prioritize per core performance they're not workstations.
Thought about it though and they might be able to go with a full fat 8 core desktop zen rather than zen lite ; if the 1700 only has a 65 watt tdp at 14nm then by the time we hit 7nm or lower such a chip could be feasible. Disabling hyper threading would allow for higher clocks if not higher yields.
I wouldn't expect hyper threading though considering 360 is the only console that has it. They should prioritize per core performance they're not workstations.
Thought about it though and they might be able to go with a full fat 8 core desktop zen rather than zen lite ; if the 1700 only has a 65 watt tdp at 14nm then by the time we hit 7nm or lower such a chip could be feasible. Disabling hyper threading would allow for higher clocks if not higher yields.
itsmydamnation
Veteran
I think it will come down to ccx size, if zen2 ccx is 4 core then 8 would be likely, if ccx is 6, I think 6 is far more likely then 12....
If I was amd my target would be to make the 7700k irrelevant(for ports). 6 cores at 3ish does that, 8 cores at 2.4-5 does that. 12 cores at 2.0-2.4 would be insain.
MS will want to align next gen XBox CPU performance as close to PC hardware as possible to enable easier Play Anywhere development.
Considering the Ryzen 7 1700 already has 8 cores @3GHz in a 65W TDP envelope, and TSMC promises 40% power reduction going from 16ff+ to 10nm, I expect 8 cores @3GHz in less than 40W. Spend the remaining 150W of the main SOC on GPU/auxiliary functions.
Cheers
Alternatively, disable one on each hypothetical hexacore CCX for yields. Hard to say what the tradeoff benefits are for a fully enabled 4+4 vs that particular scenario for the time frame we're observing. i.e. are yields going to be good enough that fully enabled cores are the way to go compared to cramming more cores in with a bit of redundancy, bearing in mind one dead CPU core = entirely useless chip for a console. It may have been fine for 2013 Jaguars, but I don't know how yields compare between 28nm and latest nodes at the same point in life.
Could be moot a point if Zen2 is just fixing the stuff that sucks about Ryzen 2017, and they'll stick to quad modules and/or add more cache.
Last edited:
Jedec have now published the official GDDR6 specs, and it has a lot of interesting details.
https://www.jedec.org/document_search?search_api_views_fulltext=GDDR6
Two independent x16 channels (instead of one x32), with fully separate data, addr, and command.
Still allows clamshell mode, which splits channels in two x8. So four x8 per chips turn back into two x16 channels, each channel taking half of each chip.
Capacities planned are 8gb, 12gb, 16gb, 24gb, 32gb.
My guesses:
Maybe the more granular capacities could allow a capacity increase in sync with each memory manufacturer's new node instead of waiting two nodes to double the capacity.
The dual channels per chip would reduce r/w contention.
Being x16 it compensates the 16n instead of 8n prefetch, so it's still 32bytes. Must be important depending on cache lines I guess?
If they launch at 8gb on 16nm and 16gb "soon after", chances are 16gb@16gbps is very possible for 2019 at a low cost and large volume.
Time for a new PS5 prediction:
With all the problems, missed specs, and high prices of HBM, I think GDDR6 is a more reasonable option.
768GB/s
384bit bus
24GB total
- Or -
512GB/s (or a bit less)
256bit bus
32GB clamshell
And for the GPU: Suppose Navi increases IPC by say 20%, solve power issues, maybe improve AMD's AVFS for a higher clock at lower power (like, uh, the xbox one x?), double the PS4 pro configuration for another "easy" BC, increase the clock in the 1200 range. That's a good 14TF (28TF fp16) and can be served by the above 256bit configuration.
Secret sauce: AMD's equivalent "tensor cores" for AI and deep learning based temporal reconstruction.
The tensor cores have already been hinted, Sony's pre-show last E3 had the technology guy talking about new circuitry improving AI processing.
https://www.jedec.org/document_search?search_api_views_fulltext=GDDR6
Two independent x16 channels (instead of one x32), with fully separate data, addr, and command.
Still allows clamshell mode, which splits channels in two x8. So four x8 per chips turn back into two x16 channels, each channel taking half of each chip.
Capacities planned are 8gb, 12gb, 16gb, 24gb, 32gb.
My guesses:
Maybe the more granular capacities could allow a capacity increase in sync with each memory manufacturer's new node instead of waiting two nodes to double the capacity.
The dual channels per chip would reduce r/w contention.
Being x16 it compensates the 16n instead of 8n prefetch, so it's still 32bytes. Must be important depending on cache lines I guess?
If they launch at 8gb on 16nm and 16gb "soon after", chances are 16gb@16gbps is very possible for 2019 at a low cost and large volume.
Time for a new PS5 prediction:
With all the problems, missed specs, and high prices of HBM, I think GDDR6 is a more reasonable option.
768GB/s
384bit bus
24GB total
- Or -
512GB/s (or a bit less)
256bit bus
32GB clamshell
And for the GPU: Suppose Navi increases IPC by say 20%, solve power issues, maybe improve AMD's AVFS for a higher clock at lower power (like, uh, the xbox one x?), double the PS4 pro configuration for another "easy" BC, increase the clock in the 1200 range. That's a good 14TF (28TF fp16) and can be served by the above 256bit configuration.
Secret sauce: AMD's equivalent "tensor cores" for AI and deep learning based temporal reconstruction.
The tensor cores have already been hinted, Sony's pre-show last E3 had the technology guy talking about new circuitry improving AI processing.
Last edited:
2x ps4p
Prophecy2k
Veteran
That's still 6x improvement over 1.84TF GPU in base PS4. And since it is not a wild spec increase, it will probably enable Sony to again reach $399 pricepoint.
With better CPU and faster&more RAM [mass GDDR approach still more likely than any 2.5D solution], that would give a nice boost over games that are today built to run on a 1.3TF console GPUs [in base Xbone].
IMO, 16x 2GB [16gbit] chips for a start, and then move again to 8 chips down the line [8x32gbit].
Also, I hope that devs who make high-end AAA games will avoid going with native 4K rendering. I'm fine with image quality taking a sliiiiiiight clarity hit when 4K checkearboard rendering is used [like in Horizon or TLOU2 unveil vid]. That would leave so much resources available for rendering of more complex assets.
Additionally, I would like for Sony to move entire main OS to the secondary "background" chip, Make it beefier [or even X86 if needed to maintain compatibility with PS4s software... heck it can even be 2 regular Jaguar cores
], place 4GB DDR3 ram next to it, and dedicate entire APU for gaming.
Last edited:
Order of importance IMO for Sony is:
One of the main differences between PS5 and other gens will be development environment regardless of BC. I think 90% of their tools, api's and dev environment will be the same as PS4, and if there isn't any BC, then a game could be rereleased by recompiling with minor porting if required.
This will also allow it to get off to a rolling start instead of devs needing to get to grips with new dev environment, apis & tools.
This OS will also be a port of the PS4 one, so the PS5 will be a fully functional and feature complete OS from the start.
SMT - I think this will be enabled, I don't think it takes up much die space, and even if it did it probably would take more work to reclaim it. I'm assuming the non SMT zens just have it disabled, not that the circuitary isn't there.
May take bit more work for devs to deal with cache lines etc, but the benefits out weigh it imo.
8C 16T @ 3Ghz (max)
Probably would want to go with Navi, or an updated 4pro gpu. 10-12 TF
16GB Gddr6
I'm not sure if the overall spec would come in too much though. 2.5Ghz cpu, 8-10 TF gpu on the lower side?
- Cost
- Time to market
- Performance
- Backwards compatibility
One of the main differences between PS5 and other gens will be development environment regardless of BC. I think 90% of their tools, api's and dev environment will be the same as PS4, and if there isn't any BC, then a game could be rereleased by recompiling with minor porting if required.
This will also allow it to get off to a rolling start instead of devs needing to get to grips with new dev environment, apis & tools.
This OS will also be a port of the PS4 one, so the PS5 will be a fully functional and feature complete OS from the start.
SMT - I think this will be enabled, I don't think it takes up much die space, and even if it did it probably would take more work to reclaim it. I'm assuming the non SMT zens just have it disabled, not that the circuitary isn't there.
May take bit more work for devs to deal with cache lines etc, but the benefits out weigh it imo.
8C 16T @ 3Ghz (max)
Probably would want to go with Navi, or an updated 4pro gpu. 10-12 TF
16GB Gddr6
I'm not sure if the overall spec would come in too much though. 2.5Ghz cpu, 8-10 TF gpu on the lower side?
I could see Jag making another appearance. I mean, hopefully not, but I could see it.
MS have already pushed it to 2.3 gHz with newly engineered latency improvements. I have no doubt Sony could have done the same given the time and money.
If 10 nm and a couple of years work could get you another 30% you'd be looking at 3 gHz with hassle free backwards compatibility. A 4 x 4 core arrangement would take up less space than a 2 x 4 Zen arrangement too.
MS have already pushed it to 2.3 gHz with newly engineered latency improvements. I have no doubt Sony could have done the same given the time and money.
If 10 nm and a couple of years work could get you another 30% you'd be looking at 3 gHz with hassle free backwards compatibility. A 4 x 4 core arrangement would take up less space than a 2 x 4 Zen arrangement too.
Prophecy2k
Veteran
The point I was making was that 72 CUs at 1200 MHz isn't enough to get to 14 TFLOPs.
There might be other ways to get to 14TF, but so far I like the napkin math. I mean, it's possible to raise the clock speeds higher I guess, but with 'easy' BC, this seems to make the most sense.
Diehard is right about 11TF hitting a 399 price point. I think everyone is hoping for 'more', but it's realistic. And it would fall in line with 11 TF GPU prices in a couple years time, much like Scorpio is landing right at the same price point of a ~1070 for me (in canada atm).
- Status