Console hardware: what Sony/MS went with versus what they should have

[Fjodor2001]

they could have still used an amd cpu, 4 amd desktop cores at 3.2ghz would give vastly better game performance than what they are using now

At what TDP, die area and price? And how does that compare to the current XBONE/PS4 dies?

 

[ShintaiDK]

If you are willing to count all expanses then why ignore some income sources? Include PSN and xbox live subscriptions on the other side aswell as royalties from games.

Because we talk about hardware. the entire console business is build upon getting the money back later in terms of service and games.

 

[ShintaiDK]

3Ghz 4T SR would be about as fast as 8 cat cores at 1.5Ghz.

While there wasn't SR back in 2013 it probably would require more die area and more power to run than jaguar cores.

Also, jaguar core was designed with portability in mind, which is very important for console chips. Allowing easy and swift fab or process switching can be very profitable in the console lifecycle.

You confuse CPU efficiency and performance.

http://anandtech.com/bench/product/1200?vs=1223

 

[hawtdawg]

Developers will eventually figure it out. This will force them to figure out how to properly multithread.

 

[Sweepr]

Either a +3GHz 2M/4C PD/SR or an 8C Jaguar running at 2.0-2.4GHz coupled with a 2.5 Teraflops GCN iGPU. Ideally I'd choose an early 2014 console with 3 Haswell cores and a custom Nvidia Maxwell GPU with 1280 CUDA cores. GTX680/280X like performance and very power efficient.

 

[monstercameron]

A die shrink and clock speed bump would also be doable. As for compatibility think about it like the dsi to the ds or the new 3ds

 

[AnandThenMan]

Because we talk about hardware. the entire console business is build upon getting the money back later in terms of service and games.

It took Sony about 3.5 years to be profitable on the PS3 hardware. Given this round the console generation will be shorter than previous it's easy to understand why Sony didn't want to lose money on the hardware for the majority of its production.

 

[Erenhardt]

The target is 1080p for this gen. So maybe a gpu frame buffer of 2gb gddr5 would work with the cpu getting 8gb ddr3.

2GB is not enough for high rez textures today.

they could have still used an amd cpu, 4 amd desktop cores at 3.2ghz would give vastly better game performance than what they are using now

What about background tasks? You need to dedicate something to background tasks to assure the same experience no matter if there is stream/download/camera etc running in the background.
That would leave the console with 3cores.

But I expect amd didn't propose the big cores on purpose. They wanted consoles to have lots of small cores, which will help their situation in desktop where they offer cheap 8 core cpus that struggles in games to beat 2 core intel counter parts because of how low the cpu utilization is.

This will force devs to make a good use of more weaker cores, which is reflecting in recent games.

 

[turtile]

Like I already posted in another thread, there was no other choice at the price point.

1. Steamroller wasn't ready -> impossible
2. Piledriver was designed on 32nm -> impossible (unless they want to pay more for no difference - GCN was designed on 28nm)
3. AMD's modules would be difficult to share between the OS and games/other tasks
4. ARM's 64-bit wasn't ready -> impossible
5. Neither Intel or Nvidia could design a SOC
6. Both Intel and Nvidia demanded a higher margin and they would have to be separate chips
7. Two chips adds even more costs

 

[Erenhardt]

You confuse CPU efficiency and performance.

http://anandtech.com/bench/product/1200?vs=1223

Nope, you don't know what this graph shows. My fault as I didn;t provide a link. Here: http://www.extremetech.com/computin...o-take-a-page-from-intel-and-dump-steamroller

The simplest way to measure the efficiency of the two chips is to divide their respective benchmark scores in a given application by (CPU Frequency * Core Count). This normalizes both variables and gives us a measure of intrinsic core performance. The next step was to turn each of these clock-and-core normalized figures into a percentage. In a test like Cinebench, a score less than 100% indicates that Kabini is less efficient than its big-core rival, while a score of greater than 100% means Kabini is more efficient.

This efficiency gives you the base to extrapolate performance in applications of X frequency cpu (100% frequency scaling). Kind of like IPC but is based on actual performance.

 

[ShintaiDK]

Nope, you don't know what this graph shows. My fault as I didn;t provide a link. Here: http://www.extremetech.com/computin...o-take-a-page-from-intel-and-dump-steamroller



This efficiency gives you the base to extrapolate performance in applications of X frequency cpu (100% frequency scaling). Kind of like IPC but is based on actual performance.

That didnt help one bit. Its completely irrelevant since jaguar cant scale to steamroller performance (specially per thread), and jaguar is deemed too slow.

 

[Fjodor2001]

Developers will eventually figure it out. This will force them to figure out how to properly multithread.

+1. And they already have in most cases. This will propagate to PC games too, where more cores will be utilized going forward. Go 8 cores or go home...

 

[Erenhardt]

That didnt help one bit. Its completely irrelevant since jaguar cant scale to steamroller performance (specially per thread), and jaguar is deemed too slow.

It does.
Graph showed that 1 jaguar core at 1.6GHz is equal to 1 SR core at 1.6GHz. If so, then 8 cores jaguar @1.5GHz = 4 cores SR @3.0 GHz if the scaling is there.

 

[ShintaiDK]

It does.
Graph showed that 1 jaguar core at 1.6GHz is equal to 1 SR core at 1.6GHz. If so, then 8 cores jaguar @1.5GHz = 4 cores SR @3.0 GHz if the scaling is there.

If that was how it worked. We all sit with Xeon Phis or something similar.

So no, its not how it works and the graph is useless in this context.

I already told you about scaling and games.

 

[BSim500]

"Raise the console TDP"
"Put 1536 stream processors at the same clock"
"What about background tasks?"

If x86 consoles are going to need 300w ATX PSU's, R9 equivalent dGPU's and handle "background tasks", I don't know why they don't just call it a day, scrap them altogether and sell consoley looking Steamboxes that come with a controller and a pack of "XBox" and "Playstation" stickers you can put on the front and pretend it's still a console... :biggrin:

 

[futurefields]

The industry is obviously going compute... Not simpler. Cerny did allegedly speak about tuning the design towards compute.

Won't work in PC land. Too many configurations to rely on Compute for game-critical functions.

Also... Cerny, who designed PS4, and then made the brilliant game Knack.

 

[seitur]

they could have still used an amd cpu, 4 amd desktop cores at 3.2ghz would give vastly better game performance than what they are using now

^^ This.

Thing is FX series is 32nm and GCN is 28nm.

The target is 1080p for this gen. So maybe a gpu frame buffer of 2gb gddr5 would work with the cpu getting 8gb ddr3.

2 GB is nowhere near enough. If you wanna employ certain techs and get rid of heavy compromises that were made in times of previous gen multiplaform AAA games then you need much more than that.

 

[monstercameron]

Thing is FX series is 32nm and GCN is 28nm.


2 GB is nowhere near enough. If you wanna employ certain techs and get rid of heavy compromises that were made in times of previous gen multiplaform AAA games then you need much more than that.

2gb is more than enough for 1080p gaming. The bloated memory demand of newer games is somewhat due to current gem console unified memory.

 

[futurefields]

2gb is more than enough for 1080p gaming. The bloated memory demand of newer games is somewhat due to current gem console unified memory.

this x1000

2gb would be plenty of dedicated VRAM in a console aiming @ 1080p

 

[Shivansps]

It took Sony about 3.5 years to be profitable on the PS3 hardware. Given this round the console generation will be shorter than previous it's easy to understand why Sony didn't want to lose money on the hardware for the majority of its production.

This is the thing, we are moving to 4K now, and these consoles cant properly handle 1080p yet, im not even asking for 60 fps.

Once again we have the consoles dragging everything back like it did happen with the PS3...

Maybe there was not a more profitable option for Sony, but that does not mean its the best they could have done, and personally i dont care about the company, i do care about the games, i dont want devs wasting half the budget trying to figure out how to use the damn jaguar.

 

[Xpage]

I think they should have had the Jag cores made with SOI, and increased the frequencies, and save some of the power budget to put towards GPU. Plus SOI is allegedly cheaper toproduce.

I also think a triple channel memory using DDR3L @1.3 volts would have been a better bet than using more expensive DDR5. A 3x4Gb DDR3L or DDR3 depending on the speed you wanted and thermal budget. Granted this would increase tracing costs on the PCB but that shouldn't be too expensive.

Savings from using DDR3 instead of DDR5 can go to a large CPU or GPU portion of the CPU, or putting in a hybrid hard drive to slightly speed up the system.

Personally I would have had something like intel's Intel Smart Cache Technology to speed up the system, have 64Gb MLC, on a small add in card. So if the MLC goes bad you can just replace the card (think M.2 like). Or can always sell larger cards for an upgrade like the original memory sticks for playstation.


Hopefully the nextgen consoles will be SSD with optional HDD for additional space

 

[podspi]

I think everybody is missing something in this conversation:

We are not building a PC. These are not PCs in shiny boxes that lots of people buy because they're stupid and could have just built their own. You cannot build an Xbone, you cannot build a PS4. Both contain custom silicon not found in any PC. Now, the usefulness of that silicon is debatable and outside the scope of this comment, but it IS there.

Given that, really want MSFT and Sony needed was CPU and GPU IP to plug into their designs. Given this, the only logical choice was Jaguar. And if you're going to AMD for CPU IP, might as well get the volume discount on GPU IP as well.

Is Excavator highly portable like Jaguar? Somewhere I heard it was supposed to be. If so, the new Nintendo console could be really awesome. Otherwise expect more Jaguar+ cores if they're smart (for easy ports).

 

[Arkaign]

It would undeniably be more expensive, but a ~2ghz Intel Haswell Quad + 4GB DDR3 + decent midrange Nvidia GPU (maybe 760-level) + 3GB GDDR5 would have been far preferable.

 

[cytg111]

What game(port) was it that wouldnt run on a dual core cause it was hardcoded not to run on core 0 and 1?
Anyway, I suspect that what console devs *need* is some sort of a guarantee on performance, it is constant, consistent, does not throttle, some annoying Operating System will not lay claim to a 50% timeslice even for a millisecond.. and for a console dev that operates close to the metal 6 cores jaguar at 1.6 may just be what the doctor ordered.
Suppose you had half the cores but twice the clocks - in theory you should be able to slice and dice those clock cycles to fit the 6-core scenario just nice right? Or will you have a thread eating more that its designated 50%? Less maybe?
Just saying there might be advances to having a single core for a single task.

 

[Tuna-Fish]

But I expect amd didn't propose the big cores on purpose. They wanted consoles to have lots of small cores, which will help their situation in desktop where they offer cheap 8 core cpus that struggles in games to beat 2 core intel counter parts because of how low the cpu utilization is.

At the time of the console deals, AMD was desperate for revenue. They had very little room to bargain for uncertain future benefits. If either Sony or MS had wanted large cores, they'd have gotten them.