Next Generation Hardware Speculation with a Technical Spin [2018]

[steveOrino]

What sort of cores would provide a generational improvement with 4 cores compared to the 8 jaguars?

Something in the nehalem/sandy bridge range. Assuming Zen 2 in new consoles its going to be more than 2x. I highly doubt 12 cores though due to GPU being dominate % of total area and thermal considerations. Regardless, Zen Is going to be like a generation and a half upgrade over jaguar.

 

[turkey]

I don’t think they’ll do anything less than 8 cores due to their backwards compatibility patents. They talk about slowing down and restricting features on a granular basis to match old architectures, and I doubt they want to attempt that on virtual cores.

Isn't the Xbox already presenting vCPU cores due to it all being behind the hypervisor, one core is shared with the OS and a variable resource.

 

[anexanhume]

Isn't the Xbox already presenting vCPU cores due to it all being behind the hypervisor, one core is shared with the OS and a variable resource.

Possible, but I was specifically referencing Sony’s parents. Sorry, should have been more clear.

 

[Deleted member 13524]

Routing, I am not expecting 8 core ccx, 4 cox's for more likely. Hopefully with Improvements to coherency handling / distributed home agents.

What routing?
8 cores in a single CCX would mean shared L3 cache for all of them, for coherency.

As mentioned before, Intel only found necessary to partition the L3 cache into multiple segments (with a ring bus to connect them) when they went over 10 cores and 25MB L3.
And they probably did it to increase the amount of salvageable chips with defects in the L3, not necessarily because there was a bottleneck from L3 access requests.

Similarly, the dual 4-core CCX in each Zeppelin probably exists to salvage chips with defects in a full/half CCX.
Come 7nm with each element occupying less than half the die area, I can't really think of a reason not to increase the core count in each CCX. Increasing the number of CCX modules in each Zeppeling successor (assuming AMD will keep sticking to ~200mm^2 dies due to existing form factors) would force them so spend a disproportionate amount of die area on CCX interconnects, not to mention the routing nightmare for inter-die IF.


I honestly thought AMD was going to increase the CCX core count to 6 cores, but rumors are now pointing to AMD going with 8-core CCX for 2019's Zen 2:

https://www.extremetech.com/computi...fer-higher-core-counts-major-ipc-gains-report

 

[Recop]

At least a 7 Tflops machine to get PS4 graphics at 4k/30fps : https://www.resetera.com/threads/am...4-image-quality-requires-7-4-teraflops.55382/

Just obvious.

 

[Xbat]

That's why 12 TFlops using temporal or checkerboarding upscaling will allow for a nice improvement over current gen.
If they can squeeze 14 TFlops into the next machines that would be fantastic and a true next gen leap in my opinion.

 

[goonergaz]

At least a 7 Tflops machine to get PS4 graphics at 4k/30fps : https://www.resetera.com/threads/am...4-image-quality-requires-7-4-teraflops.55382/

Just obvious.

Yes, didn't Cerny say 8TF min

 

[turkey]

Yes, didn't Cerny say 8TF min

he said closer to 8 when asked is 6 enough.

All seems somewhat marketing, the 7.4 number is just PS4 flops times 4, it does not seem to be evidence based.

I am sure Microsoft released some figure from their testing on how many times flops you needed to go from 1080 to 2160 and it was not linear (4). It was 3.6 or something.
Also any new console also would have other GPU effeciencies baked in, so work per TF will be higher so again it is unlikely to require a pure 4 times increase in flops for the same screen output at 4k.

I don't know numbers but I am sure you must be able to find third party titles on the X at 4k that are 1080 on the PS4. This is possibly harder as many enhanced titles also push higher setting and 4k. In thoes cases 6tf is over and above the requirement. There are also going to be cases where it is not enough to raise an equivalent title to 4k.

 

[Recop]

I don't know numbers but I am sure you must be able to find third party titles on the X at 4k that are 1080 on the PS4. This is possibly harder as many enhanced titles also push higher setting and 4k. In thoes cases 6tf is over and above the requirement. There are also going to be cases where it is not enough to raise an equivalent title to 4k.

There is no third party game that runs at 900p on XB1 and native 4K on XBX.

All evidence we have seem to suggest the AMD engineer is right.

You can't have PS4 level of computation at 4k/30fps with a 6Tflops machine.

 

[turkey]

There is no third party game that runs at 900p on XB1 and native 4K on XBX.

All evidence we have seem to suggest the AMD engineer is right.

You can't have PS4 level of computation at 4k/30fps with a 6Tflops machine.

Good point.

I see why you might want to stipulate 900p for Xbox to try and remove wasted performance on the PS4 but dynamic res and other techniques seems to taken hold so it is not so clear.
Wolfenstein 2 goes from dynamic 810p to 2160p which is a huge leap but it deviates a lot.

I had a very quick Google and Far Cry 5 seems a good third party test case. If we assume esram is no longer a performance limit this game shows at least resolution and tf does not scale linearly in some cases. X the version is 5.3 time the pixels of the base release. That's over the tf differential. The pro release only matches the tf differential but with its meager ram bandwidth increase its not clear if that held it back.

Certainly real world results dont support my proposition, but I do think newer GPU revisions will do more per flop so it's not as straight forward as simple multiplying by 4.

Anyone have some good examples for or against?

 

[Shifty Geezer]

Surely the best place to test that is PC? Pick a GPU of x TFs and run a game at whatever res, and then a GPU x*4 TFs and compare.

 

[Recop]

I had a very quick Google and Far Cry 5 seems a good third party test case. If we assume esram is no longer a performance limit this game shows at least resolution and tf does not scale linearly in some cases. X the version is 5.3 time the pixels of the base release. That's over the tf differential. The pro release only matches the tf differential but with its meager ram bandwidth increase its not clear if that held it back.

Indeed, but the same thing happens between the PS4/XB1 where the performance gap is often higher than the TFlops difference. The XB1 seems to be an underperforming machine relatively to its GPU power.

"PS4 uses a dynamic resolution with the lowest resolution found being approximately 1792x1008 and the highest resolution found being 1920x1080 [...] Xbox One X uses a dynamic resolution with the lowest resolution found being approximately 3136x1764 and the highest resolution found being 3840x2160. All four consoles often render at or near their maximum resolution, but the base PS4 in particular rarely ever deviates from its 1920x1080 peak resolution"

1008p vs 1764p = 206%

1840 Tflops vs 6000 Tflops = 226%

XB1 lowest vs 1764p = 433% while the GPU difference is only 358%

By the way, the most demanding games simply run at straight 900p on XB1.

 

[function]

Not all processing work scales with resolution e.g. vertex transform; newer hardware is generally more efficient (higher utilisation, known bottlenecks reduced); and despite all kinds of advancements and enhancements in hardware your bottleneck may be somewhere that doesn't do flops e.g. BW rather than logic.

That said, 4 x PS4 is a fair comment and not intended to be presented as universal law.

Given X1X's BW, regularly pushing 300%+ of the PS4 is no small achievement.

BW is one reason we should probably keep our next gen expectations in check. We're possibly looking at as little as 50% more than X1X.

 

[Recop]

Given X1X's BW, regularly pushing 300%+ of the PS4 is no small achievement.

Since the PS4 almost always hits native 1080p, it's obvious that it's the console that has the most headroom in FC5 and it's confirmed by the lowest possible resolution on all machines.

Lowest resolution on XB1 vs XBX = 433%

Highest resolution on XB1 vs XBX = 433%

VS

Lowest resolution on PS4 vs XBX = 206%

Highest resolution on PS4 vs XBX = 300%

It's obvious that when the X runs at native 4k, the PS4 could easily achieve a higher resolution.

Games that run at a higher resolution than 900p on XB1 are suspicious on PS4 at 1080p if they run with the same graphical settings.

Do you know a third party game with a max resolution of 1600x900 on XB1 that can achieve +300% resolution on X compared to the PS4 version on a regular basis ? I don't know any...

With a similar hardware, you need at least 7Tflops and it's confirmed by actual evidence. Obviously a more recent hardware could achieve more with less Tflops, but it's not the case with the X.

Edit : actually, No Man Sky is the firt third party game running from 900p to native 4k on XBX lol. So there's 1 example, but at the expense of screen tearing and worse performances than on PS4 and PS4 Pro.

 

[function]

Since the PS4 almost always hits native 1080p, it's obvious that it's the console that has the most headroom in FC5 and it's confirmed by the lowest possible resolution on all machines.

Lowest resolution on XB1 vs XBX = 433%

Highest resolution on XB1 vs XBX = 433%

VS

Lowest resolution on PS4 vs XBX = 206%

Highest resolution on PS4 vs XBX = 300%

It's obvious that when the X runs at native 4k, the PS4 could easily achieve a higher resolution.

Games that run at a higher resolution than 900p on XB1 are suspicious on PS4 at 1080p if they run with the same graphical settings.

Do you know a third party game with a max resolution of 1600x900 on XB1 that can achieve +300% resolution on X compared to the PS4 version on a regular basis ? I don't know any...

With a similar hardware, you need at least 7Tflops and it's confirmed by actual evidence. Obviously a more recent hardware could achieve more with less Tflops, but it's not the case with the X.

Edit : actually, No Man Sky is the firt third party game running from 900p to native 4k on XBX lol. So there's 1 example, but at the expense of screen tearing and worse performances than on PS4 and PS4 Pro.

You're using percentages wrong.

100% of PS4 resolution is the same resolution.

200% is twice the resolution.

400% is four times the resolution e.g. 1080p to 4K.

X1X regularly goes beyond 300% of PS4 resolution, even with enhancements. If you re-read my last comment in this thread I don't think you'll find any of it objectionable!

 

[dten]

If the Scarlett generation gets 10+ TF with minimal bottlenecks, we should see 4K, locked or dynamic but mostly locked, with the vast majority of titles. The X already set the stage for that. As mentioned, the bigger bottleneck is likely RAM bandwidth. We didn't get a linear rise of TF to BW in the mid-gens.

 

[Shoujoboy]

There is no third party game that runs at 900p on XB1 and native 4K on XBX.

All evidence we have seem to suggest the AMD engineer is right.

You can't have PS4 level of computation at 4k/30fps with a 6Tflops machine.

This might be a bit off topic but No Man's Sky is 900p on Xbox One and 4k Native on Xbonex

 

[Karamazov]

This might be a bit off topic but No Man's Sky is 900p on Xbox One and 4k Native on Xbonex

but it runs worse at 4K on Xbox One X than at 900p on Xbox One.

 

[BRiT]

but it runs worse at 4K on Xbox One X than at 900p on Xbox One.

CPU limits? Since that's the weakest update and also why it doesnt make sense to target such a pipedream of post-midgen upgrades.

 

[Karamazov]

with all those procedural generations, might be CPU intensive yes.
But the weird thing from the DF video is that it seems to struggle in space stations, which are not procedurally generated. Actually that's where the X runs worse than base. Maybe the last updates or future ones will fix this.
The last update improved performance on my PS4 pro save.