Next Generation Hardware Speculation with a Technical Spin [2018]

Status
Not open for further replies.
Here's an "all else being equal" example. Among the different skus from intel scalable platform we have:
Xeon 8164: 26 cores 2.0GHz (150w tdp)
Xeon 6144: 8 cores 3.5GHz (also 150w tdp)

It means dropping the clock from 3.5 to 2.0, and adding cores up to the same tdp, provides 1.86x the raw total performance and therefore 1.86x perf per watt. That 26 cores looks great on paper but the die is much bigger ($$$) and not all workloads can actually take advantage of it. Network servers are perfect for this because they run hundreds or thousands of processes with little interraction between them. I suspect the typical games on PC would be faster on the 3.5GHz for a much lower cost.
 
No way we'll see more than 8 cores... if we do it'll be a mistake. I suppose amd could sell them some low clocked threadrippers on the "cheap", but those games that have 1 or 2 main threads and the rest are pretty much idle are going to suffer. Of course, such a chip would still be much better than jaguar single thread performance.

Unless there's some coding/engine revolution next gen where it's easy to effectively utilize 32 threads that I can't foresee.
 
I don't think the console should be designed around performance profiles for games that are poorly threaded. Looking at what Mike Acton and others are doing at Unity, I think the general performance and quality of Unity games is going to increase significantly, because it'll be highly optimized for any platform it runs on. I haven't watched the latest Unreal Engine GDC stuff, but I wouldn't be shocked if they're following the same path. I don't think core count is something to really worry about at this point. There may be a number of other things about ARM that makes it unsuitable for a console, but to me, that should not be one of them.

As for gaming on ARM, Nintendo Switch has great results. Obviously Sony and Microsoft would want a higher performance profile than Switch, but it's proof that game studios can handle ARM without issue. Also, look at mobile gaming. Fortnite and PUBG are shockingly good versions of the full games. They obviously pale in comparison to PC and console, but it's remarkable what they've done on handheld mobile devices.

If you put an arm cpu in a console, Guerrilla Games, id software, Unity, Epic, Naughty Dog, Turn10 would all handle it without issue. The real drawback is backwards compatibility and re-optimizing code-bases and tools that were already optimized for x86. Of course, the only real reason to switch to ARM would probably be to save power in the overall console design, so those Watts could either be saved to reduce cooling, or to put those Watts into the GPU.
 
I imagine that the PS5 and XBoxTwo will maintain x86 CPU's - AMD offers an easy, cheap, straightforward upgrade path for the next generation, at least.

But as silicon scaling becomes more difficult over that generation, I can imagine ARM may look quite attractive come ~2028.
 
I imagine that the PS5 and XBoxTwo will maintain x86 CPU's - AMD offers an easy, cheap, straightforward upgrade path for the next generation, at least.

But as silicon scaling becomes more difficult over that generation, I can imagine ARM may look quite attractive come ~2028.
At higher performance targets, I would assume the instruction decoder becomes a more negligible difference. So the reason to go ARM is more about the IP being open, and going x86 is about binary compatibility.
 
AMD/NVIDIA.
We've been through this already, many times... Nvidia likes high margin contracts, so why would they sell their precious GPUs for cheap to a console manufacturer when they could sell them for loads more to someone else?

AMD has limited engineering capacity; PS4 and Xbone APUs overlap somewhat with PC APUs utilizing their own x86 cores and GPUs, and Xbone and PS4 APUs are also each technically similar. A widely dissimilar solution sold to each console maker would mean greater risks for little genuine reward.

Twice the cores, half the per-core performance, will means higher efficiency.
Higher power efficiency; lower computational efficiency. Half the performance per core means jobs take 2x as long to complete, which isn't good in a system where low latency is important (VR especially, with its 90FPS minimum requirement.)
 
AMD has limited engineering capacity; PS4 and Xbone APUs overlap somewhat with PC APUs utilizing their own x86 cores and GPUs, and Xbone and PS4 APUs are also each technically similar. A widely dissimilar solution sold to each console maker would mean greater risks for little genuine reward.
It does bring back and old question, what does next gen need technically which would require a different engineering than what amd/nvidia are already developping for the PC gaming market?

A more integrated APU obviously (semi-custom), and a few additional features, but this gen and even last gen are proving the wild tangents away from standard architectures end up causing more problems than they solve. PS4, PS4P, XB1X, and switch, are bog standard architectures now. No cell, no esram, no fancy stuff messing up future emulation efforts.
 
It does bring back and old question, what does next gen need technically which would require a different engineering than what amd/nvidia are already developping for the PC gaming market?

Moar Tensor Cores!!11!!11!!1111!One1


Is there any remote possibility of having hardware-assisted RayTracing items on next-gen? I think it's still too soon unless AMD/Nvidia are releasing consumer products under $400 in the PC space in the next 12 months.
 
A more integrated APU obviously (semi-custom), and a few additional features,
I could see integration changing either way. Separate pools with HBCC functionality for a smallish HBM stack or integrated directly. May be more feasible with GPU driven approaches. On the other hand straight up using a Ryzen APU with some improved bandwidth solution might be the more economical solution. Whether or not those two overlap is hard to say, but may work considering timelines.
 
Moar Tensor Cores!!11!!11!!1111!One1


Is there any remote possibility of having hardware-assisted RayTracing items on next-gen? I think it's still too soon unless AMD/Nvidia are releasing consumer products under $400 in the PC space in the next 12 months.
Yeah I have no problem waiting until 2021 for next gen if it means they show a 120TF figure at the unveiling.

By that time all the tensor cores experiments will have been done on PC and it won't be a risky solution. Algos will be well tested and integrated in major engines.
 
Moar Tensor Cores!!11!!11!!1111!One1


Is there any remote possibility of having hardware-assisted RayTracing items on next-gen? I think it's still too soon unless AMD/Nvidia are releasing consumer products under $400 in the PC space in the next 12 months.

I was kind of wondering the same. I suppose Ray Tracing could remain exclusive to PC versions and have rasterizer/compute fallbacks for console.
 
I don't think the console should be designed around performance profiles for games that are poorly threaded. Looking at what Mike Acton and others are doing at Unity, I think the general performance and quality of Unity games is going to increase significantly, because it'll be highly optimized for any platform it runs on. I haven't watched the latest Unreal Engine GDC stuff, but I wouldn't be shocked if they're following the same path. I don't think core count is something to really worry about at this point. There may be a number of other things about ARM that makes it unsuitable for a console, but to me, that should not be one of them.

As for gaming on ARM, Nintendo Switch has great results. Obviously Sony and Microsoft would want a higher performance profile than Switch, but it's proof that game studios can handle ARM without issue. Also, look at mobile gaming. Fortnite and PUBG are shockingly good versions of the full games. They obviously pale in comparison to PC and console, but it's remarkable what they've done on handheld mobile devices.

If you put an arm cpu in a console, Guerrilla Games, id software, Unity, Epic, Naughty Dog, Turn10 would all handle it without issue. The real drawback is backwards compatibility and re-optimizing code-bases and tools that were already optimized for x86. Of course, the only real reason to switch to ARM would probably be to save power in the overall console design, so those Watts could either be saved to reduce cooling, or to put those Watts into the GPU.



Basically, the era for consolidation of mobile and console gaming is here. This is from the top boss.

Mobile has a much larger base than PC. Gaming can be consolidated around the ARM architecture with ARM-based next-gen consoles.

Microsoft can then leverage ARM/UWP gaming against Steam/Win32 with ARM-based XBOX strengthening Windows Store at the same time.
 
Moar Tensor Cores!!11!!11!!1111!One1


Is there any remote possibility of having hardware-assisted RayTracing items on next-gen? I think it's still too soon unless AMD/Nvidia are releasing consumer products under $400 in the PC space in the next 12 months.

PowerVR showed that off a couple years ago, and that was with a phone/tablet graphics array. It really disappoints me PowerVR hasn't jumped back into the big GPU game.
 
We've been through this already, many times... Nvidia likes high margin contracts, so why would they sell their precious GPUs for cheap to a console manufacturer when they could sell them for loads more to someone else?

Because.. NVIDIA of 2018 is not NVIDIA of 2004. If it wants PS5 for the PR win it can have it. Masayoshi Son, the 'Warren Buffet' of Japan, 4th largest shareholder of NVIDIA, and CEO of Softbank (owner of ARM), if he wants to partner with Sony and broker a sweetheart deal for PS5, it can happen. PS5 will be a large, high volume contract.

AMD has limited engineering capacity; PS4 and Xbone APUs overlap somewhat with PC APUs utilizing their own x86 cores and GPUs, and Xbone and PS4 APUs are also each technically similar. A widely dissimilar solution sold to each console maker would mean greater risks for little genuine reward.

K12 is done. It's R&D was 'highly leveraged' with Zen x86 and was not a separate development. And if Microsoft wants to leverage K12 ARM XBOX silicon with ARM servers, Surface devices, or something else then there is a synergy.

Higher power efficiency; lower computational efficiency. Half the performance per core means jobs take 2x as long to complete, which isn't good in a system where low latency is important (VR especially, with its 90FPS minimum requirement.)

Multithreading is superior on ARM over x86 when using 2x ARM cores, with lower power consumption. Couple multithreading cores with high performance single-thread cores for overall higher performance AND higher efficiency solution.

Crude mockup of next-gen big.Medium.Little console ARM CPU concept:

DY_5pctV4AA7Jp9.jpg


(2x A75 class = superior MT performance to 1 Zen at lower power consumption. K12 ST cores = >Zen ST performance with higher power efficiency. A55 class cores free 1 Zen core lost for OS/standby solution.) Note: A75 class cores smaller than 1mm2 at 7nm.
 
Last edited:
This really shows that simpler CPUs with less single threaded performance are more efficient for extremely parallel work such as web servers. I doubt it would run a game nearly as well as each of the ARM CPU cores have only about 60-80% of the IPC the Skylake has. I doubt game developers would want 20 ARM64 cores over 8 high performance x86 cores even if the ARM cores are almost 2x as energy efficient as the x86 when doing purely parallel work.

Which is why you can see why you would partner multithreading cores with 'big' dedicated single-threaded cores in a theoretical high performance/high efficiency ARM console CPU. And it would be superior to 8 Zen on all metrics, i.e., tdp, die area, ST/MT performance.

DY_5pctV4AA7Jp9.jpg
 
Because.. NVIDIA of 2018 is not NVIDIA of 2004. If it wants PS5 for the PR win it can have it. Masayoshi Son, the 'Warren Buffet' of Japan, 4th largest shareholder of NVIDIA, and CEO of Softbank (owner of ARM), if he wants to partner with Sony and broker a sweetheart deal for PS5, it can happen. PS5 will be a large, high volume contract.
you have to understand how some businesses like to operate.
high margin businesses will always be more ideal than low margin, as your ability to deal with fluctuations in the market is significantly better.
in a scenario where you cannot sell new video cards like disposable cups, the idea of high margin units especially when you are the leader in that market makes a lot of sense.

It still costs resources to make a cheap, low margin unit, so why bother when you're beating your competition with high margin products?
 
Which is why you can see why you would partner multithreading cores with 'big' dedicated single-threaded cores in a theoretical high performance/high efficiency ARM console CPU. And it would be superior to 8 Zen on all metrics, i.e., tdp, die area, ST/MT performance.

DY_5pctV4AA7Jp9.jpg

What big single threaded ARM core? Big SOC's are multi year development projects, a console in late 2019 is already past the point of no return, a console in late 2020 would have to be close. But none of your arguments are in anyway convincing, because you're still not offering any major performance improvements, you're not addressing any of the major cost challenges going forward.

Your comparison is to yesterday's x86 CPU, on yesterday's process, compared to today's arm cpus on today's process. Well the next console will be tomorrow's cpu on tomorrow's process.

In your proposition, who is going to build what and when is it going to release? ( the more vendors and the closer the integration the longer the devlopment effort)

edit: Also looks like the core you love sucks at games physics(3d mark):
https://www.androidcentral.com/samsung-galaxy-s9-tested-exynos-9810-vs-snapdragon-845
looks like it sucks at computationally heavy latency sensitive workloads.
I expect Falkor sucks at that even more - consoles shouldn't be designed by geekbench/spec rate scores.
 
Last edited by a moderator:
Status
Not open for further replies.
Back
Top