Knightime_X
Member
Nothing pinpoint specific.
Just any and all necessary measures.
If it works, alternative or otherwise. Use it.
-
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.
When next gen hits (Xbox Next, PS6), where will be with node shrinks and feasible performance?
- Thread starter In Cold Blood
- Start date
Just any and all necessary measures.
If it works, alternative or otherwise. Use it.
Black_Stride
do not tempt fate do not contrain Wonder Woman's thighs do not do not
AMD will have an Interpolation solution sooner rather than later.
TSR is already here and looks great.
Other TAAU solutions exist and weve seen devs creating their own like Insomniac.
FSR will continue to improve.
A Reflex rival likely isnt something high on the priority list.
Saying you want DLSS makes it seem you specifically want nextgen consoles to give up die space for AI acelerator.....while im all for that, I think youll find many devs would rather have more horsepower on that die space and rely on TSR and other solutions.
Whether Interpolation to get to 60 is worthwhile, personally I dont think so, but for getting to 120 yeah sure.
Mr.Phoenix
Member
Completely agree. That's why I referred to TF calculations now as rudimentary.
What I do believe will happen is that CUs won't quadruple but instead only double. So PS6 has 72CU. I also feel that focus would instead be on other components of the APU, like having significantly better RT, AI core within the CU, or even straight-up temporal accelerator, More advanced IO, More cache, advancements in dual issue scheduling, and innovations in packaging, like the PS6 may not even use a monolithic APU but maybe a chipset design...etc.
Last edited:
winjer
Gold Member
TSMC plans to have N2 ready in 2026. So in 2027, it should be mainstream.
The big thing about it, is that it will be the first node using GAA. This is a huge step, only comparable to Finfet.
N7 to N5, was a 80% increase in transistor density. N5 to N3, was 70%. So N3 to N2 should have a similar value.
But the big question is price. TSMC has increased prices for all it's nodes in 2021. Even for older ones.
A waffer in N3 costs 20k. So a wafer in N2 should be around 25K.
The big thing about it, is that it will be the first node using GAA. This is a huge step, only comparable to Finfet.
N7 to N5, was a 80% increase in transistor density. N5 to N3, was 70%. So N3 to N2 should have a similar value.
But the big question is price. TSMC has increased prices for all it's nodes in 2021. Even for older ones.
A waffer in N3 costs 20k. So a wafer in N2 should be around 25K.
Bry0
Member
I expect this Gen to be shorter because next gen will be architected around dedicated ai accelerators. Would line up nicely with amd roadmaps. Current gen came out at a weird transitionary period and I think in a Year or two that will be obvious in hindsight. Since there will also be a big need for software support for new hardware features I’m expecting that focus to shift to next gen and not within a refresh if one even happens.
Last edited:
Crayon
Member
Not sure but I'm not too worried about it the way games are now. Let's face it. The best looking games have been mostly cross gen even with a relatively large gap between the 8th and 9th gen capabilities.
I'm still expecting slow but steady improvement as game software itself catches up.
I'm still expecting slow but steady improvement as game software itself catches up.
Mr.Phoenix
Member
Go_Ly_Dow
Member
PS5 Pro will be able to run graphically intensive current gen games in 4k 60fps modes with perhaps some more RT features
PS6 wi be able to run graphically intensive current gen games at 4k 60fps with full RT features and make use of these new engine bells and whistles like Nanite and Lumen.
These are my expectations.
PS6 wi be able to run graphically intensive current gen games at 4k 60fps with full RT features and make use of these new engine bells and whistles like Nanite and Lumen.
These are my expectations.
Knightime_X
Member
You can have/use both.
But ai tech is much more efficient then brute forcing with raw power.
Not to mention a lot cheaper and less power hungry.
Those line up as 84%, 70% and 10% (which he didn't say). However I think he is saying that N2 would be the first to use GAA and result in a similar performance increase value as the other node changes, not similar density increases. It's just worded a bit funny.
Thick Thighs Save Lives
Gold Member
According to techpowerup the PS5 APU built on TSMC 7nm has a transistor density of 34.4M / mm². On the other hand the RTX 4070 is built on TSMC 5nm and has a transistor density of 121.4M / mm². I picked the RTX 4070 as it has similar die size to the PS5 APU (295 mm² for the former and 308 mm² vs for the latter). That's around 4x the amount of transistors going from 7nm to 5nm on TSMC on roughly the same die space.
Last edited:
YOU PC BRO?!
Member
Microsoft should release their next console as a handheld. Somethimg like Steam Deck or Asus ROG Ally. Give it a dock like the Switch which connects to the TV. The dock would contain a powerful external GPU. They have been doing something similar for years already with the Surface product line.
The new AMD Z1 Extreme in the ROG Ally packs a 8/16 Zen 4 CPU with a 12CU RDNA3 GPU. This is already very capable but could be a step up for the next Xbox. They could allow devs to utilise the IGPU when docked to offload lower priority tasks I.e. HUD etc. The external GPU in the dock could be as powerful as they want. They could sell the handheld only with a basic dock (no eGPU) or one version with it included. They could also sell a separate disc drive for those who want it.
It would be such a home run for Xbox if they did something like that.
The new AMD Z1 Extreme in the ROG Ally packs a 8/16 Zen 4 CPU with a 12CU RDNA3 GPU. This is already very capable but could be a step up for the next Xbox. They could allow devs to utilise the IGPU when docked to offload lower priority tasks I.e. HUD etc. The external GPU in the dock could be as powerful as they want. They could sell the handheld only with a basic dock (no eGPU) or one version with it included. They could also sell a separate disc drive for those who want it.
It would be such a home run for Xbox if they did something like that.
Mr.Phoenix
Member
It would be an all-around very complicated and expensive mess.
In Cold Blood
Banned
Did you mean to say the PS5 APU at 7nm?According to techpowerup the PS5 APU built on TSMC 5nm has a transistor density of 34.4M / mm². On the other hand the RTX 4070 is built on TSMC 5nm and has a transistor density of 121.4M / mm². I picked the RTX 4070 as it has similar die size to the PS5 APU (295 mm² for the former and 308 mm² vs for the latter). That's around 4x the amount of transistors going from 7nm to 5nm on TSMC on roughly the same die space.
Mr.Phoenix
Member
TheThreadsThatBindUs
Member
GAAFET will save us all.
Combined with microarchitectural improvements, we'll be looking at a 8x - 9x improvement in raw single precision compute. In general, that's the least interesting part of the next-gen question, imho.
The pink elephant in the room is main memory bandwidth and the complete lack of major advancements in this area compared to microprocessors. HBM was a miserable dud, especially on consoles, adding oodles of on-die SDRAM as L4 cache is a non-starter; not least because of the die size limitations but also the impact on production yields.
The more ideal approach would be to pursue a fully chiplet-based design, but even more granular in concept than AMD's recent high-end CPU offerings; i.e. something that takes the entire GPU/CPU pipeline + AI ASIC and splits it across multiple tiny chiplet dies with plenty SDRAM cache and a high-bandwidth interconnect fabric between each. It would be a major paradigm shift and require a potentially whole new programming model. One of the major drawbacks would be the complete infeasibility of BC.
Realistically, I can't see the above idea taking off. Instead, I think we'll have similar traditional APU designs to the current consoles. I still don't see any major silicon being devoted to dedicated AI cores, unless AMD reveals something revelatory within the next two years. Regardless, memory bandwidth, or lack thereof will be the albatross weighing the next-gen console designs.
Combined with microarchitectural improvements, we'll be looking at a 8x - 9x improvement in raw single precision compute. In general, that's the least interesting part of the next-gen question, imho.
The pink elephant in the room is main memory bandwidth and the complete lack of major advancements in this area compared to microprocessors. HBM was a miserable dud, especially on consoles, adding oodles of on-die SDRAM as L4 cache is a non-starter; not least because of the die size limitations but also the impact on production yields.
The more ideal approach would be to pursue a fully chiplet-based design, but even more granular in concept than AMD's recent high-end CPU offerings; i.e. something that takes the entire GPU/CPU pipeline + AI ASIC and splits it across multiple tiny chiplet dies with plenty SDRAM cache and a high-bandwidth interconnect fabric between each. It would be a major paradigm shift and require a potentially whole new programming model. One of the major drawbacks would be the complete infeasibility of BC.
Realistically, I can't see the above idea taking off. Instead, I think we'll have similar traditional APU designs to the current consoles. I still don't see any major silicon being devoted to dedicated AI cores, unless AMD reveals something revelatory within the next two years. Regardless, memory bandwidth, or lack thereof will be the albatross weighing the next-gen console designs.
winjer
Gold Member
You have the same values. as I. 1.7x means 70%.
Anandtech says there is an improvement of 80% from N7 to N5. Not 84%.
I got my data from TSMC site. Example:
You have one error. It's not 10% from N3 to N2. It's 10% from N3E to N2.
N3E should have some minor improvements over N3.
Still, having only 10% higher density from N3E to N2 is a very small step for a new node.
TSMC must e going very conservative on their first GAA process, to ensure good yields.
YOU PC BRO?!
Member
Why?
Mr.Phoenix
Member
Yes, I am aware that you and I have similar values (just written differently)... what was making me question anything actually was the 1.1x of N3 to N2. Cause I saw you suggest similar improvements as the other node shrinks.
And the N3 vs N3e thing is kinda weird. Compared to N5, albeit N3e (1.6x) is coming after N3 (1.7x), it offers less density.
Last edited:
winjer
Gold Member
I still hadn't seen the forecast for N2. So I thought it would be similar to previous nodes.
Even if N3 to N3E has another 10% in density, it will still be much lower improvement that any other generation.
We are not even in "using any of it's next gen features" phase of this gen yet.
Awkwardly attached Ray Traycing to cross gen rendering pipeline does not count.
No direct storage, no mesh shaders no nothing that suppose to give big perfomance gains...
This gen might stay with us for quite allot longer then we expect
Awkwardly attached Ray Traycing to cross gen rendering pipeline does not count.
No direct storage, no mesh shaders no nothing that suppose to give big perfomance gains...
This gen might stay with us for quite allot longer then we expect
crackajack
Member
i see that also as a realistic baseline; an undervolted 4090 at a future node plus a then modern 8+8 core CPU should be possible within the same power budget as today. PC progress will probably grind to a halt with insane prices leaving many people behind, so progress in efficiecy and cheap tech will get more important than trying to milk the few whales for high end. Game devs must target mass market which not even a 4060 is anymore.
I tend to believe MS might be convinced by something intel is cooking, but ARM from nvidia would also be a quite interesting possibility.
Mr.Phoenix
Member
I don't think so, I believe we are at the very least going to see something like an infinity cache in the next-gen consoles. Its already being used in AMD CPUs (3d-V cache) and in their GPUs(infinity cache), so I do not think its a stretch to say that technology will be very mature in the next 4-7 years and make its way into consoles.
Thick Thighs Save Lives
Gold Member
Yup, my bad.
Here's the TPU links with the relevant details:
AMD Playstation 5 GPU Specs
AMD Oberon, 2233 MHz, 2304 Cores, 144 TMUs, 64 ROPs, 16384 MB GDDR6, 1750 MHz, 256 bit
NVIDIA GeForce RTX 4070 Specs
NVIDIA AD104, 2475 MHz, 5888 Cores, 184 TMUs, 64 ROPs, 12288 MB GDDR6X, 1313 MHz, 192 bit
TheThreadsThatBindUs
Member
It's not about the maturity of the technology. Infinity cache is just a large on-die cache. For consoles, it's not an ideal solution because anything that increases die size or process steps (i.e. 3d-V cache requires an advanced packaging process, which further impacts yields) is non-ideal for non-binned volume devices like console APUs.
More cache is not a solution for console APUs. It's not as simple as you think. Consoles aren't like discrete CPU/GPU parts that have the benefit of using binning to increase yields.
saintjules
Member
I wonder if in the future we can just bring a PS Console to a Sony approved/licensed shop and do incremental upgrades and/or changes to the hardware.
Mr.Phoenix
Member
Ok, forget 3D V cache, that won't work... but the infinity cache, which is pretty much just a lot of cache sitting between the VRAM Bus and the cores can be done. And has far less impact on yields than say 3d V cache.
Furthermore, according to AMD, using Nfinity cache was better all round than having a larger bus on the chip, higher PCB complexity, and higher power draw+ heat. That's why they went with infinity cache in their GPUs. I mean, they already have 128MB worth of it in the RX6800, or 80MB worth of it in the RX6700 (a PS5 equivalent GPU). And those GPUs cost (at launch MSRP) under $570 and $450 respectively... and you know the kinda margins OEMS put on their GPUs, this means they probably cost AMD under $400 and $350 to make respectively.
And while SRAM isn't shrinking as fast as core logic, its still shrinking. I honestly can't see how the next-gen consoles don't come with an infinity cache. I can see them still even going with a 256bit bus, GDDR7 and then adding in like 128MB of infinity cache on top of that. Or even as little as 80MB as seen in the 6700 right now.
Schmendrick
Member
not gonna happen. modular hardware and production cost scaling and ofc build complexity / size don`t go well together.
Last edited:
TheThreadsThatBindUs
Member
Again you're missing the point.
Those GPUs you list cost as much as a console and that is taking into account the fact they are binned parts. Consoles don't have that luxury. And even just having a bunch more SDRAM on die can have a deleterious impact on console APU yields, since the higher density circuitry means defects in the SRAM cells will more likely kill the whole APU, whereas with logic circuitry APU designers can and typically do include redundant cores that can be disabled completely to improve yields.
So no, I don't think a bunch of L4 cache on die is the solution for consoles that you think it is.
Last edited:
LordOfChaos
Member
To be fair they'll also keep getting games years into the next gen
But yeah, I don't really feel a need for new hardware any time soon, even a PS5 Pro, I'd rather see things continue to take more advantage of the hardware we have as the first three years were weird.
HundredThousandTrillion
Member
There won't be more hardware. Not in the way we imagine it anyway.
In Cold Blood
Banned
So with the most likely PS5 Pro configuration being a 72cu GPU, will the feasible node for a 2024 release be able to fit that many CUs on a die size of around 320-350?
In Cold Blood
Banned
There was the rumour that MS was working with AMD on a GPU with a massive ML block on it.
DeepEnigma
Gold Member
Chiplet stacking.
Mr.Phoenix
Member
Well... technically. No reason why not. Especially if a lot of the other things in the die are left as is, i.e Bus, L2 cache, CPU cores...etc.
The Current PS5 die, @308mm2 @7nm has a transistor density of 34.4M/mm2.
The RX 6700XT (the spiritual twin to the PS5 GPU, has the same 40CU though none are disabled)), while lacking a CPU but having 96MB of L3 cache, is a 330mm2 die, also 7nm, and has a density of 51.3M/mm2.
The RX7xxx series are on 5nm Nodes and have a density of over 100M/mm2. That would suggest, that at the very least, going from 7nm to 5nm would mean you can at least tripe the transistor density in the PS5 APU.
All you need to go from 36CU to 72CU... would be around 2x the transistor density. This obviously changes if they want to start adding in stuff RDNA3+ CUs that may be bigger than those in the OG PS5 relatively, increasing the L2 cache from 4MB in the PS5 to say 6MB in a pro, Using Zen 3/4 CPU cores instead the OG Zen 2 (highly unlikely but yeah)...etc. Stuff like that could eat into the relative transistor density. But there should be enough room to double the size of the GPU.
thicc_girls_are_teh_best
Member
I know everyone here is focused on the processing capabilities of 10th-gen systems, but honestly, as I've thought about things more...aren't they going to need some type of hook outside of "MOAR POWAH!!" in order to justify themselves in a market where graphical gains (or leaps in graphical fidelity) are getting smaller gen-over-gen? I mean look at HFW: Burning Shores and the final boss battle there; that seems comparable almost to a CGI film from a decade ago in terms of fidelity and scope, fully playable as a video game, for expansion content built on the base of a game that was made to run on a PS4. It's 100% possible Horizon 3 will have battles, fidelity and scale that puts even Burning Shores to shame.
At that point it's like, "how does this even get any better visually?" in a way where it makes a big impact on most of the market. I don't think it does. So, I think 10th-gen systems need some type of immersion hook beyond just being more powerful than the current systems, to really stand out and galvanize the market IMO. I do have some ideas on that, and in some ways current product strategies by Sony & Microsoft show hints at which way they will go, but in them playing to their own strengths, I think they'll actually take very divergent paths with 10th-gen systems, to the point where I don't feel we'll be able to point to them as directly competing with each other anymore.
Just my two cents.
At that point it's like, "how does this even get any better visually?" in a way where it makes a big impact on most of the market. I don't think it does. So, I think 10th-gen systems need some type of immersion hook beyond just being more powerful than the current systems, to really stand out and galvanize the market IMO. I do have some ideas on that, and in some ways current product strategies by Sony & Microsoft show hints at which way they will go, but in them playing to their own strengths, I think they'll actually take very divergent paths with 10th-gen systems, to the point where I don't feel we'll be able to point to them as directly competing with each other anymore.
Just my two cents.
sachos
Member
AI acceleration hardware, hardware accelerated reconstruction techniques a la DLSS, way better RT harwadre to enable industry wide adoption of Pathtracing a la Cyberpunk 2077.
Answering OP, this gen came out in 2020 with GPU performance around high end PC GPUs from 2 years ago (~2080 from 2018). Totally wild speculation but if the current trend stays, they will manage to release in 2027 (7 year cycle we are used to) with 2025 GPU performance. RTX 4080 released in 2022, 2024 should be RTX 5000 and 2026 should be RTX 6000. So they either release with RTX 5000 series performance or they wait one more year and release in 2028 to reach performance similar to RTX 6000. Of course im only talking about raster performance here, but i i said earlier i suspect the most important will be the RT performance, lets hope AMD can keep up.
Last edited:
ShadyAcshuns
Member
When WokeStation 6 and Xbox: We Haven't Learned a Thing, come out, they can keep them. From this point forward I'm only buying Nintendo systems.
Last edited:
sachos
Member
This is why i have high hopes for Generative AI and it being able to bring back dev cycles down to PS2 era speeds.
Last edited:
I think folks are missing that the PS6 will most likely have GPU chiplets and not a monolithic die. Similar to AMD CPUs.
Mr.Phoenix
Member
AMD is not that far behind. Even with their RT only real issue is that their current-gen RT doesn't accelerate every part of the RT pipeline, it leaves one part out. Thats something that can be easily remedied. As for AI, that is even easier, the whole AI thing is just a case of adding tensor cores. Which contrary to what some may think, is not owned by Nvidia. Another name for this is Matrix cores/accelerators. The only issue with that is that AMD seems to not want to grow their die size by what it would take to add them. Started putting them in RDNA3 though.
But you are spot on with getting hardware equivalent to whatever high-end thing was released two years prior.
Doesn't really matter much, just means certain components in the APU are decoupled, so some stuff gets built on the more expensive N2 process and some stuff on the cheaper N5 or something. The latter would usually go to stuff that has to do with IO, Cache, Buses....etc.
Last edited:
Panajev2001a
GAF's Pleasant Genius
While on the PC market they seem to be having issues taking advantage of their synergies across their divisions the success of mobile handheld and other semi custom designs like XSX|S and PS5 (and XBox One/One X and PS4 before that) show why I am still excited about what they bring to the table for their semi custom partners and in consoles boxes (where everyone is restricted by similar size and power constraints).
The expertise they have within their CPU division, their expertise with custom interconnects and memory solutions (you can probably bet the next generation SoC’s for consoles will all have VCache), the experience they have in their graphics and compute divisions (CDNA splintering off was a good choice and they are free to advance compute and AI, the Matrix cores you mentioned, freely for a market that makes use of them), and recently also FPGA’s with their Xilinx acquisition which now has had enough time within AMD to allow some cross pollination of talent and ideas (FPGA’s / programmable logic in the next generation consoles could be a quite good idea… you can offer built in features but also allow devs to provide their own custom remapping if they do choose).
Mr.Phoenix
Member
An infinity cache... yes. But Vcache, not so much. Chance of making a console APU too expensive. Not from an architecture point of view, but from a yield perspective.
I value single player games and I don't care which generation we getting through, single player games haven't changed its identity, still original, yes ,we are getting fewer, but at the same time we get to access alot more than what we used to play back in the day, nobody listed 20+ games back then and anyone like that considered weird , but not in modern days, 20 games is nothing and we get to like more than what we used to, thanks to technology and the individual creative minds from everywhere.
It should be feasible on 5nm. A 80CU GPU should take up ~250mm^2 as we know the GCD on Navi 32 will be 200mm^2 with 60CU. A PS5 Pro can remove some stuff from that like extra display outputs, a full 16x PCIe interface etc for die size savings.
A stripped down 8-core Zen4 or Zen4c cluster will come in at less than 50mm^2. Zen4 chiplet is 70mm^2 with 32 MB cache and consoles will probably reduce that to 16 MB like the Phoenix APU, using Zen4c cores rather than the full Zen4 should reduce the size a bit too.
Real question is what they do memory controller-wise (fully integrated or chiplet-style, L3 cache etc). They'd probably need at least a 320-bit wide GDDR6 memory controller. If they reuse the RDNA3 MCD chiplets they'd need 5 or those so 37.5*5=187.5mm^2 of 6nm silicon.
Last edited:
It does matter because it would be significantly cheaper than a large monolithic die. In part because of much higher yeilds.
EverydayBeast
thinks Halo Infinite is a new graphical benchmark
PlayStation and Xbox basically grew up together I got to play ps1-5 and thought ps5 was the best.
Panajev2001a
GAF's Pleasant Genius
Let’s see it is a few years away and getting the bandwidth they will need otherwise would put an enormous strain (huge power consumption too) on the external memory / main RAM.
Similar threads
- Poll