Could next gen consoles focus mainly on CPU?

What can be cut out from zen cores that is useless/expensive on a console?
Maybe smt and turbo boost, and some infinity fabric interface too?
Mine's an uneducated guess, but from all these threads is obvious that half the commenters thinks that it must be zen based, the other half thinks that zen is too big for a console budget.
 
Last edited:
In an unified memory world too strong CPUs are non so fully used because they steal resurces to GPU... Better spend the silicon budget on bandwidth & GPU (as ONE X did by the way)... I know there are games very CPU intensive... Me too I would like to see Star Citizen on consoles.... yes... But
 
fehu said:
What can be cut out from zen cores that is useless/expensive on a console?
Maybe smt and turbo boost, and some infinity fabric interface too?
Mine's an uneducated guest, but from all these threads is obvious that half the commenters thinks that it must be zen based, the other half thinks that zen is too big for a console budget.
Click to expand...
why would you cut smt?
SMT is a highly useable feature, just because one "thread" can never saturate a whole cpu. Yes it could be a bit hard to count on SMT when you develop for the console, but without you can never max out the whole cpu.

what you don't need in the CPU are all or most of the external stuff (like integrated PCIe controllers). sometimes I wonder why we actually need new mainboards, because everything relevant is already integrated in the cpu. everything else is connected via pcie/usb which is also integrated in the cpu. The mainboards are more or less just there to connect and power all those things.

those boost things ... yes you barely want that in a console, but it would be great if a game could just "turn off" a few cores it definitly does not need, just to save some heat so the cooler can be more quite if the game does not demand so much power. Not every game will max out the cpu/gpu, so different power states would be a great addition. e.g. idle or sleep mode can really need a few tweaks (10+W in sleep mode is to much). That is nothing games really need, but make the product more consumer friendly.
 
It would be interesting a couple of predefined/predictable modes, like increase gpu frequency and reduce cpu's one to allow for better graphic and lower frames, or reduce two cores' frequency to increase one and park there your heaviest thread.
Zen can do this :cool:
ooohmmmmmm
 
fehu said:
What can be cut out from zen cores that is useless/expensive on a console?
Maybe smt and turbo boost, and some infinity fabric interface too?
Mine's an uneducated guess, but from all these threads is obvious that half the commenters thinks that it must be zen based, the other half thinks that zen is too big for a console budget.
Click to expand...

SMT on its own is a modest change to large OoO core, since the engine's ability to rename and track separate sections of one program can be extended incrementally to use that independent tracking to separate threads. The high-performance features that usually come with the OoO engine and larger hardware budget SMT affords by making extra hardware justifiable make SMT's impact take up a smaller fraction of the core.
The hardware impact of the P4's SMT was claimed by Intel to be 5% or so, on a core much more narrow and smaller than today's cores.
The increasing sophistication of SMT to make it more consistent might have added expense in an absolute sense, but this was usually paired with increasing hardware budgets everywhere else.

There can be a power impact on the most power-sensitive devices, due to threads potentially waking up parts of the chip that could sleep during long stalls, and at the high-end systems end due to heavier utilization. Consoles are something of a middle ground where they don't feel either pressure that strongly, though perhaps an argument could be made for their desiring a consistent performance contract for devs, which can get more complicated with SMT.

Turbo is an offshoot of the same DVFS hardware and algorithms on-chip, so the same measures that allow chips to operate more efficiently also underpin turbo. AMD's Jaguar had that capability priced into its architecture, it was just barely functional.
I can see the argument for consistency being made for disabling turbo as for SMT, though in both cases it's more like it's telling hardware that's already doing a lot of other things to not perform this one additional function.

There could be minor elements removed, like perhaps a smaller number of buffers for SMT or fewer sensors here and there for DVFS, but the savings of that versus paying for revalidating a barely smaller custom core don't seem compelling.

The fabric might be able to dispense with some error-checking features, but at the same time if the consoles are like they are today they'll have an unusually higher ratio of memory bandwidth and memory clients per chip, which seems like it would more than compensate in terms of dropping those extra features.

If the goal is to really save on hardware, more significant changes like paring back larger structures like the front end and caches, ALUs, and re-targeting the pipelines to a more modest clock range would be needed to make it noticeable.
 
fehu said:
What can be cut out from zen cores that is useless/expensive on a console?
Maybe smt and turbo boost, and some infinity fabric interface too?
Mine's an uneducated guess, but from all these threads is obvious that half the commenters thinks that it must be zen based, the other half thinks that zen is too big for a console budget.
Click to expand...
I think a majority of us agree it’s going to be Zen. The question which actually seems to be at debate is a matter of how much Zen (size) or how much of Zen (features) should we require for next-gen knowing what we know today and where developers want to bring their games in the future.
 
Last edited:
What will be the importance of AVX 512 etc...? Is it still worth the heat/area expense to improve FP throughput on the cpu with wider vectors, or will that sort of code be moved to the gpu in the future?
 
HBRU said:
yes Mr Fox... This is a good question. As in todays consoles there is not this instruction.
Click to expand...
I'm not talking about that extension specifically, I mean generally improving the performance of very wide vector stuff.
 
It certainly seems unnecessary, but at the same time having them will help with compatibility with libraries etc. I'm not sure any core functionality changes are really worth it in the long run. It's not about making the ultimate efficient, lean console hardware, but making the most productive, functional platform, and the greatest compatibility with PC code is going to be part of that. A few mms silicon savings seem a false economy to me.
 
Focusing indeed
fehu said:
What can be cut out from zen cores that is useless/expensive on a console?
Maybe smt and turbo boost, and some infinity fabric interface too?
Mine's an uneducated guess, but from all these threads is obvious that half the commenters thinks that it must be zen based, the other half thinks that zen is too big for a console budget.
Click to expand...
Why cut anything? Having take a good look at the last Rysen 2400/2200G definitely the "big" cores and the associated cache are not taking a crazy amount of die space. They will run circle around the antiquated Jaguar cores. There is no further integration into the CPU cores hence their size is relatively stable (leaving Intel huge SIMD aside, + See EDIT2) compared to the other elements. If one look at this die shot he will notice how "big" are the multimedia engine, the display engine, etc not to mention the GPU SIMD arrays compared to the CPU (+ what was an incredible amount of cache some years ago).

If manufacturers want to push out a tiered offer to the mass they could focus on the CPU by designing a (custom) SOC shared among various SKU, while they rely on OTS part for the GPU. NUMA works and is not the reason why PC costs more than consoles at least it is clearly not the main factor. UMA is great but providing high bandwidth and lots of ram is complicated and costly.

I know the PS4 (hence the XbX) went the route and not long before release Sony was stuck with 4GB of RAM, then there is the randomness of memory prices: huge bet on more than one factor. It all aligned nicely but lets be rational and consider the whole picture MSFT had rational intensive to pass on a plain UMA design (as most manufacturers PC and console before them).
On the topic of cost looking at the non subsidized industry (aka PC) APU (and so UMA) still have tough time to compete and that include price. Those last Rysen require fast and costly memory to "shine" and the shine is not that "bright". I read Hardware.fr review the TDP AMD is announcing is in line with CPU use once you push graphic you get into the +100 Watts easy.
I disliked the XB1 design (I dislike the XBX design too) yet I would not consider SOC+UMA a given or the more rational choice without serious cost studies. Now even exes are not always free from trends and things like that, you can add public perception.

EDIT
On the topic of cost, I just want to underline that datum: AMD sell the Rysen 2200G (4cores, no smt, 8 SIMD GPU) for 99€, how low would or will they sell hypothetical "Athlon" (with disable GPU)?
Cheap CPU+CHeap memory+cheap GPU vs significantly more expensive APU, more expensive memory (x2 dual channel), etc.
I do not believe that the situation is so different for AND and the surrounding PC ecosystem and console manufacturers.
EDIT 2
Pushing further down that line of thinking, providing enough CPU is getting cheap, the split in silicon attribution should be closer to the usual Nintendo take on the matter (than that of its competitor). Definitely I would consider a <<100mm2 CPU complex bound to an affordable pool of RAM through a single channel.
 
Last edited:
That's for the EUV variations, not what's going to ship late this year/next which should just be same Finfet based technology.
 
From the comments...
Stochastics generated defects is an old news and industry is already working on it. On the first two days there were half a dozen papers showing how we will address defects and there will be many more papers on how to address defectivity over coming days. I will summarize them after the conference. Defectivity is not a showstopper but a difficult challenge which will be addressed. Cheers!
Click to expand...
 
ok... but it may be expensive (on 7nm) or really difficult (on 5nm) as quantic phyisics starts to give effects
 
HBRU said:
ok... but it may be expensive (on 7nm) or really difficult (on 5nm) as quantic phyisics starts to give effects
Click to expand...
I'm not familiar with quantic phyisics effects giving. Can you develop a bit more?
 
You must log in or register to reply here.

Similar threads

Globalisateur
Checkerboard Rendering in Dark Souls Remastered
Replies
10
Views
3K
Johnny Awesome
J
M
Sony PS6, Microsoft neXt Series - 10th gen console speculation [2020]
29 30 31
Replies
601
Views
88K
Seanspeed
S
R
An Nvidia console seems a no brainer to me. Who else could Console?
2 3 4
Replies
75
Views
12K
Deleted member 13524
D
A
Dynamic framebuffer resolution
2 3
Replies
42
Views
16K
chris1515
C
Inuhanyou
Is PS4's focus on Compute a proper substitute for a more powerful CPU?
2 3
Replies
56
Views
13K
Inuhanyou
Inuhanyou
Back
Top