• Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.

Next-Gen PS5 & XSX |OT| Console tEch threaD

Status
Not open for further replies.

SonGoku

Member
Disabling only each faulty CU would be preferred. But RDNA slides show dual compute units. It may mean that you can only disable the dual CU unit rather than individually. So for each faulty CU, you'll likely be taking it's conjoined healthly twin one out of action too? This is assuming that the semi-custom layouts used for PS5 and Anaconda are very similar to the released RDNA info which may not be the case.
Precisely, that's why you can't have a 60CU chip
 
Precisely, that's why you can't have a 60CU chip

60CU active is a 64 CU die with 2 dual compute units disabled.

64 - 2x2

So if the stars align you're taking out 4 faulty CUs in those 2 groups. But you're really only likely to be disabling (1 faulty+1heathy) x 2.

So you're really only being able to work around a max of 2 faulty CUs in each 64 die.

But you still disable 4 in total.

Because for each of the two faulty CUs, you're forced to disable it's twin as well.

Even if all 64 CUs are healthy, 4 are disabled regardless to always give you 60 active.
 

ethomaz

Banned
60CU active is a 64 CU die with 2 dual compute units disabled.

64 - 2x2

So if the stars align you're taking out 4 faulty CUs in those 2 groups. But you're really only likely to be disabling (1 faulty+1heathy) x 2.

So you're really only being able to work around a max of 2 faulty CUs in each 64 die.

But you still disable 4 in total.

Because for each of the two faulty CUs, you're forced to disable it's twin as well.

Even if all 64 CUs are healthy, 4 are disabled regardless to always give you 60 active.
If you have 4 SE you need to disable 8 CUs (1 DCU per SE).

Navi 10 engine of RX 5700 has 2 SE with 10 DCUs... the XT uses it fully 40CUs but the Pro model disable 2 DCU (one in each SE) decreasing the CU count to 36.

With 4SE the minimum you are suppose to disable is 8CUs... 3 SEs is 6 CUs (if that weird chip will even exists because the layout of SE seems to hint at even numbers).
 
Last edited:
Who said any of that sort? You disable one DCU per SE

I know. I said it as an example of pushing the imbalance too far - having everything (4 CUs) disabled in a single shader engine. Although the system will still chug along quite happily I expect.

In practice you're unlikely to have identical jobs on every CU, each running for the exact same duration which is where this would show the imbalance or having a single shader engine (with the 4 disabled CUs) struggling behind.

From a game perspective, you could probably quite easily live with and never really notice having the active CUs unevenly spread across the shader engines. No idea how it would affect the lifetime of the hardware though.

You haven't answered my question btw, what total power consumption are you toying with?

Not my field, so I'd just be guessing.
 

CrustyBritches

Gold Member
Went through the benchies for AMD's 11 games they've shown to compare 5700 XT with RTX 2070 and made some observations. If you're bored have at it and chat me up if you get a chance...

I went through the AMD Navi 5700 XT slides for their 10-game 1440p batch + initial Strange Brigade demo and came to 6% faster than RTX 2070. Radeon VII used 3 of the same games in comparison to the 2080 and came out to 7% faster than 2080. Real world comprehensive benchmarks show the 2080 beating the Radeon VII by ~10%. We'll see where Navi 5700 XT lands in comparison to RTX 2070.

Using the 1440p 11-game comparison with percentages you can look at Guru3d and TPU 1440p benchmarks and compare the 5700 XT to other cards like Radeon VII and Vega 64. Using TPU card only power consumption you can get an idea of real world average and peak gaming consumption compared to listed TBP. Lastly, the 5700 XT 50th Anniversary gives you a look at draw past their base/boost clocks.

I'm baked out of my brains, so it could be me, but I came to the 5700 XT being around 14% faster than Vega 64 and 6% faster than RTX 2070.

RTX 2070 is ~11% faster than Vega 64 and about the same over 2060 in comprehensive 1440p benchmarks. If that's the case, how is 5700 Pro showing 11% increase over 2060 when that would be within 5-6% of 5700 XT by AMD's own numbers? There's some funky marketing shenanigans going on here.

My guess is just like Radeon VII, Navi ends up about the same or weaker than their Nvidia comparison GPU. Also, AMD real-world power consumption historically runs higher than listed TBP in both Average and Peak gaming consumption. Wouldn't be a bit surprised if 5700 XT's real world gaming average is ~235W and peak ~250W while performing about the same as 195W Average/203W Peak RTX 2070.

PS4 Pro total system power consumption = 155W
Xbox One X total system power consumption = 175W

Slides:
AMD-Radeon-RX-5700_Official_2-740x416.jpg

AMD-Radeon-RX-5700_Official_1-1480x833.jpg
 
Last edited:

Ar¢tos

Member
10 TFs is not bad. It's more than 5x current (base gen). If we consider how much got squeezed out of base Ps4 and that even a toaster cpu managed to handle VR games, Zen2 + 10tf gpu can create amazing games. We don't really need 4k native if we can have high quality CB from 1800p, there might even be hw customizations to reduce the cost of CB and AA + AF.
 
I would take a fairly modest graphics upgrade with great image quality, Ray tracing and 60fps in most games. Instead of a big push in graphics but yay our games still run at sub 30fps.
 

llien

Member
Monolithic dies are more performant because information has to travel a shorter distance to it's destination.

But Zen 2 has chiplet "IO Die" yet boasts record performance.

Are AMD's upcoming APUs monolithic?

If chiplet penalty isn't big (assuming it isn't, because, well, Zen 2) by cramming GPU and CPU core together, one gets worse yields and bigger heat dissipation problems. Since GPU and CPU normally work with different areas of RAM, why even bother?

RT cores, implementing DXR, would be "interested" in different blocks of memory, than CPU/rest of GPU. And if so, again, why cram that onto the same chip?

It is one of the reasons for sure but here are a lot of things nVidia does that AMD not.
There are also things AMD does that nVidia don't (on outside at least). The lag reducing feature, as well as Radeon Chill.
As for what is going on inside chips, one can only speculate.

I went through the AMD Navi 5700 XT slides for their 10-game 1440p batch + initial Strange Brigade demo and came to 6% faster than RTX 2070. Radeon VII used 3 of the same games in comparison to the 2080 and came out to 7% faster than 2080. Real world comprehensive benchmarks show the 2080 beating the Radeon VII by ~10%. We'll see where Navi 5700 XT lands in comparison to RTX 2070.

That's a lie/misleading statement. Here are "real world benchmarks" by 3dGuru, showing Vega VII definitely beating 2080:

m1JUea8.png


The "but cherry picked games != real world performance" argument is valid, however in case of 5700/5700XT AMD has shown a wide range of benchmarks, including "supersponsoredtocripplenvidiacompetitor" games.
 

archy121

Member
  • MS work towards beating the supposed PS5 12.9 perf target
  • Sony slow drip information to collaborate 12.9
  • MS trumpets their 13.5 tf masterpiece expecting they've hit a home run
  • Sony comes in and does a mic drop of 14.2

How about simple strategy to trick the competition to over spec and as a result over price their product out of the market ?

We all know well what happens if you go over the mass consumer sweet spot pricing levels - Self destruction.

Speculating fans need to take account of the fact that companies are answerable to shareholders. They try and deliver the best tech within both financial and technological constraints. The console has to deliver profit through mass consumer buy in. Too easy to forget this and just dream up super tech for a console.
 
Last edited:

SonGoku

Member
10 TFs is not bad. It's more than 5x current (base gen). If we consider how much got squeezed out of base Ps4 and that even a toaster cpu managed to handle VR games, Zen2 + 10tf gpu can create amazing games. We don't really need 4k native if we can have high quality CB from 1800p, there might even be hw customizations to reduce the cost of CB and AA + AF.
I mean its not bad... but 12TF is better specially considering both 4k and rt are resource hogs
For 10TF they will need a bigger chip anyways (56CU enabled) at that point 11TF is more realistic for a $400 price point i can see it happening.
I would take a fairly modest graphics upgrade with great image quality, Ray tracing and 60fps in most games. Instead of a big push in graphics but yay our games still run at sub 30fps.
A PC is perfect for you.
If chiplet penalty isn't big (assuming it isn't, because, well, Zen 2) by cramming GPU and CPU core together, one gets worse yields and bigger heat dissipation problems.
  1. One of the advantages of a monolithic die APU listed on this thread is the perfect design for Raytracing
  2. Isn't a big monolithic die cheaper to manufacture on the long run compared to multiple smaller ones?
 
Last edited:

ethomaz

Banned
This news today.


What that means? That means the Scarlett chip is already done and the guy didn't have more work with MS... so he is now at Intel.
So don't expect changes in design on the Scarlett's SoC... clocks are still subject to changes.

Ohhh and of course Intel wants to enter in that GPU market :D
 
Last edited:

FranXico

Member
This news today.


What that means? That means the Scarlett chip is already done and the guy didn't have more work with MS... so he is now at Intel.
So don't expect changes in design on the Scarlett's SoC... clocks are still subject to changes.

Ohhh and of course Intel wants to enter in that GPU market :D
I'm sure he transferred knowledge to others. Changes are still possible.
 
Last edited:

ethomaz

Banned
I'm sure he transferred knowledge to others. Changes are still possible.
Hardly.

Cerny said before PS4 release the SoC is locked about 2 years before the console release.

Of course others changes can happen like clock, amount of memory, etc... but the SoC is basically done for both now (CPU cores, CUs, secret sauce units, etc).

So you finalized the hardware two years ago?

We finalized the spec of the hardware. The process of creating the hardware is about four years. Two years into that it’s locked enough that you can start talking about all this other stuff that’s going to surround it. It’s not like your work actually finishes at that time, or that was any particular milestone. But it was sort of an “ok guys, we know what the hardware is now.”
 
Last edited:

Ar¢tos

Member
Not surprising, the process of physically making the chip takes at least 3 months, then all the testing, making changes, making new chips, retesting, etc. They need to be locked (design) a long time before release. The only thing that can change is clocks because those are dependent on cooling / power draw.
 

McHuj

Member

Pretty much. To design and verify an SOC is about a 3-4 year process and most of that time is verifying that it works pre and post tapeout.

A typical design would probably last 12-15 months
Then another 3-6 of verification after the features are locked down.
Then a couple of months for the chip layout
Then tapeout which is another 3-4 months on the new processes
Finally silicon comes back and you spend another several months validating everything for the first version.

Ideally, during the validation everything goes well and no major bugs are found so the second time through the process all the steps are shorter as it's just tweaks to performance, power consumption, and area. The tricky part is what happens when you find major bugs close to tapeout or during the validation period that require big redesigns or major changes. That can really screw with schedules.

My guess is that at this point, like MS showed in their video, both Sony and MS have their first silicon and they're in the validation processes. The designs are being tweaked for final clocks and bug fixes. The production version of the chip will tapeout in the next quarter or so and be ready for production in the first half of next year (first final devkits and then consumer units).


At this point I would only expect changes to related to the processes and yields like clocks going either up or down, CU/Cores being disable or not due to bad yields, and maybe RAM changes, but only if they are completely compatible (like doubling capacity or a faster RAM if it becomes available).


People really underestimate the amount of time it takes to bring something like this to market. And frankly, the majority of the time in developing an SOC is verifying that it works and not designing it.
 

SquireDalbridge

Neo Member
From the 10 to 13 TF how many will be used to get 4K Resolution on the new GPU. My thougts are about minimum 4 TF? The Big question is How MAny TF Will be left after getting 4K. One question stil lingers in my mind which CPU is PS5 and Xbox Scarlett using. Is it the Zen2 3700x or the Zen2 3800x ???
 

TeamGhobad

Banned
From the 10 to 13 TF how many will be used to get 4K Resolution on the new GPU. My thougts are about minimum 4 TF? The Big question is How MAny TF Will be left after getting 4K. One question stil lingers in my mind which CPU is PS5 and Xbox Scarlett using. Is it the Zen2 3700x or the Zen2 3800x ???

most definitely the CPU will be zen 2 3700 which means about 4x the speed of what we have now.
 
From the 10 to 13 TF how many will be used to get 4K Resolution on the new GPU. My thougts are about minimum 4 TF? The Big question is How MAny TF Will be left after getting 4K. One question stil lingers in my mind which CPU is PS5 and Xbox Scarlett using. Is it the Zen2 3700x or the Zen2 3800x ???
4.5TF for next gen visuals then whatever is left to 4k-ify them. That way, graphics will look 4x better than last gen.
 

stetiger

Member
How about simple strategy to trick the competition to over spec and as a result over price their product out of the market ?

We all know well what happens if you go over the mass consumer sweet spot pricing levels - Self destruction.

Speculating fans need to take account of the fact that companies are answerable to shareholders. They try and deliver the best tech within both financial and technological constraints. The console has to deliver profit through mass consumer buy in. Too easy to forget this and just dream up super tech for a console.
Are we sure that is the case, a 599ps3 was still sold out in the first 6 months. In the end the ps3 sold more than the 360 despite the higher price. I don't like higher prices either but I do think that to some degree a high price of 499 may not completely cripple a machine.
 

rəddəM

Member
12TF of RDNA Navi + Zen 2 for standard rendering metrics is more than enough for native 4k60 I would assume.
Put that secret HW accelerated Ray Tracing sauce on top of it...
... BOOM!
We got ourselves a MONSTER.
Edit: 10.1TF even.
 
Last edited:
But Zen 2 has chiplet "IO Die" yet boasts record performance.
I understand that the IO Die is 14nm and 12nm is planned for consumer products. There's little info on the IO Die downsides but snippets here and there drop hints about it drawing a good chunk of power and heat. I guess it kind of makes sense that a 12nm / 14nm part would be the weak point in a 7nm system.

Are AMD's upcoming APUs monolithic?

No idea and I would doubt it as chiplets make more sense for AMD as a business.

But console Navi is semi custom and I imagine the customer gets to call the shots. If the customer wants chiplets or a monolith, that's what they get.

If chiplet penalty isn't big (assuming it isn't, because, well, Zen 2) by cramming GPU and CPU core together, one gets worse yields and bigger heat dissipation problems. Since GPU and CPU normally work with different areas of RAM, why even bother?

Performance. "If the chiplet penalty isn't big", it's still a penalty. In a system dedicated to high performance why take on any hindrance at all?

RT cores, implementing DXR, would be "interested" in different blocks of memory, than CPU/rest of GPU. And if so, again, why cram that onto the same chip?

Same reason, performance.

Design a monolith to last 7-8 years, you need to pay attention to every tiny detail. You also benefit from node shrinks as time goes on 7nm, 7nm+, 6nm etc

Compare that to MS approach with putting the same tech in Anaconda/Lockhart/Data centres and you can easily see why a chiplet makes a lot more sense for their implementation. The IO Die overhead would just be thought of as another "cost of doing business" weighed against the freedom to easily drop in different IP blocks for each environment or adding a chunk of cache to the IO Die more easily.

Neither setup is wrong just different design goals and priorities pulling them in different directions.
 

stetiger

Member
I mean its not bad... but 12TF is better specially considering both 4k and rt are resource hogs
For 10TF they will need a bigger chip anyways (56CU enabled) at that point 11TF is more realistic for a $400 price point i can see it happening.

A PC is perfect for you.

  1. One of the advantages of a monolithic die APU listed on this thread is the perfect design for APU
  2. Isn't a big monolithic die cheaper to manufacture on the long run compared to multiple smaller ones?
No not even close. Monolithic dies are incredibly expensive. APUs don't look like a good way to go for next gen based on what we know for AMD. It might be cheaper for AMD to have a zen2 chiplet and two 5700 chiplet for next gen then a monolithic dies. However, as AMD pointed out themselves chiplet GPUs don't perform well for gaming. Chiplets>>>>Monolithic Chips/APUs
 
How about simple strategy to trick the competition to over spec and as a result over price their product out of the market ?

We all know well what happens if you go over the mass consumer sweet spot pricing levels - Self destruction.

Speculating fans need to take account of the fact that companies are answerable to shareholders. They try and deliver the best tech within both financial and technological constraints. The console has to deliver profit through mass consumer buy in. Too easy to forget this and just dream up super tech for a console.

The rumour mill already seems to suggest that Anaconda performance is struggling against the PS5 if it is "just" 12.9 tf.

The whole 14.2 scenario I put forward would definitely be one of those Black Swan theories. It's interesting how the 14.2 number keeps popping up in "leaks" and misinformation trails. The CUs and clock speeds required to hit it are what just seems to be a stone throwaway from 12.9. Couple that with the dev summit leak and other obvious pointers all makes for some interesting speculation. But that's all it is, speculation.
 
Last edited:

demigod

Member
Folks, we are now at 12 cores for the nextbox per windowscentral on a rumor they heard.

"Hopefully Scarlett will be the most powerful console than the PS5. However, only time will tell. There are rumors floating around that Scarlett features a 12-core processor compared to the 8-core one found in the PS5, but nothing has been confirmed yet. As with any such leak, take it with a grain of salt until we receive official confirmation and both parties confirm specifications."
 

ethomaz

Banned
Folks, we are now at 12 cores for the nextbox per windowscentral on a rumor they heard.

"Hopefully Scarlett will be the most powerful console than the PS5. However, only time will tell. There are rumors floating around that Scarlett features a 12-core processor compared to the 8-core one found in the PS5, but nothing has been confirmed yet. As with any such leak, take it with a grain of salt until we receive official confirmation and both parties confirm specifications."
Looks like just guesstimations lol

He needs some fresh news after all.
 
Last edited:

CrustyBritches

Gold Member
That's a lie/misleading statement. Here are "real world benchmarks" by 3dGuru, showing Vega VII definitely beating 2080:



The "but cherry picked games != real world performance" argument is valid, however in case of 5700/5700XT AMD has shown a wide range of benchmarks, including "supersponsoredtocripplenvidiacompetitor" games.
Meh, not really.

Why'd you quote me, "real world benchmarks", and leave out the word "comprehensive"?

Real world comprehensive benchmarks show the 2080 beating the Radeon VII by ~10%. We'll see where Navi 5700 XT lands in comparison to RTX 2070.

What AMD showed is far from comprehensive. It's the exact opposite, and should be when marketing a new product.
---
It's neither here nor there. That wasn't why I posted and you can see I use their marketing as best-case scenario to get frame rate and compare perf/watt with other GPUs like Radeon VII and Vega 64. By their own metrics, the 5700 Pro cannibalizes the 5700 XT. 5700 XT consumes 25% more power while being only 5% more powerful and having the same amount of VRAM.

It's apparent what they've done, but again, that's not the point. Using their own best-case perf/watt numbers you see where a 150W/175W/200W total system power consumption console would land for GPU performance, at best, and start dropping from there to account for a SFF, power-capped console with tiny cooler and single blower fan.
 

xool

Member
Folks, we are now at 12 cores for the nextbox per windowscentral on a rumor they heard.

"Hopefully Scarlett will be the most powerful console than the PS5. However, only time will tell. There are rumors floating around that Scarlett features a 12-core processor compared to the 8-core one found in the PS5, but nothing has been confirmed yet. As with any such leak, take it with a grain of salt until we receive official confirmation and both parties confirm specifications."

They could get 12 cores relatively easily considering how small the Zen cores are (though the issue of Zen2 being in 8 core "CCX" chunks complicates things)

.. the bigger issue - that I keep repeating - is that memory bandwidth is going to be an increasing problem as core counts increase - that's why Zen (1) parts with 12 or more cores have quad channel support (Epyc have octachannel at 8+ cores)

[everyone seems to be expecting next gen CPU + GPU to run on what is essentially single channel 32 bit GDDR6 (2 channels at only 16bit data width) ]
 
Last edited:
Folks, we are now at 12 cores for the nextbox per windowscentral on a rumor they heard.

"Hopefully Scarlett will be the most powerful console than the PS5. However, only time will tell. There are rumors floating around that Scarlett features a 12-core processor compared to the 8-core one found in the PS5, but nothing has been confirmed yet. As with any such leak, take it with a grain of salt until we receive official confirmation and both parties confirm specifications."
I even joked about 16 cores on the XII but... I mean... that would be prohibitively expensive even two years from now. I have a really hard time believing that.
 

Ar¢tos

Member
Scaling up physics effects (ie remember the litter and leaves in early Watch Dogs demo) is relatively easy to do ..
That yes, but if a game has a lot of npcs/enemies and complex AI then it becomes trickier to scale down the CPU load.
 

demigod

Member
I don't think 12 cores is going to happen. Given the performance boost from Jaguar it would be overkill, and if the other console has 8 cores then games will be made with 8 cores in mind. Its not like GPU power that you can with relative ease increase/reduce resolution or effects.

I just find it funny. Navi is in trouble. MS is using VEGA!!! Navi reveal is good, MS back to using Navi! Developers says PS5 is more powerful, Nextbox has 12 cores, what now!
 
  • LOL
Reactions: TLZ
Status
Not open for further replies.
Top Bottom