Sony PS6, Microsoft neXt Series - 10th gen console speculation [2020]

That's very good news! So it's actually very possible they could do a PS5 Slim design on 5nm and clock the GPU up (I gave 2.5 GHz as an example) while still getting much better power consumption performance than what they have currently. I wasn't crazy after all xD.

I'm actually pretty surprised at how aggressively AMD are able to push GPU clocks so soon. 3+ GHz for RDNA 3 without overclocking would be very good, I'm curious how far they can push the clocks with future architecture iterations, particularly for new consoles while keeping respectable power consumption targets.



Good points here as always, tho with the info @ToTTenTranz provided in their post apparently there's a LOT of wiggle room for clocks on RDNA architecture. If PC RDNA 2 cards are able to hit around 2.8 GHz already with overclocks on 7nm EUV, does that necessarily make a PS5 Slim with a more modest upclock on the GPU (say to 2.5 GHz) on 5nm look like such a tough thing to pull off while still enjoying the power consumption reduction benefits, more refined 5nm node process, and higher margins? It gives some hope ;)
The benefit would be marginal tbh. The games will still be designed around the original PS5 so a higher clocked GPU in the slim model may definitely suffice but it won't be substantial. For a pro model that makes more sense. Notice the PS4 slim maintained the same GPU clocks as the original PS4? Yet the PS4 pro got a 12% boost? I would expect the exact same if they were to release a slim and pro model. Both of them would be substantially smaller than the current PS5 though. But the pro could possibly double the compute units and have a slightly higher clock. Probably the GPU would be equivalent to a 6800 XT.
 
If PC RDNA 2 cards are able to hit around 2.8 GHz already with overclocks on 7nm EUV, does that necessarily make a PS5 Slim with a more modest upclock on the GPU (say to 2.5 GHz) on 5nm look like such a tough thing to pull off while still enjoying the power consumption reduction benefits, more refined 5nm node process, and higher margins? It gives some hope ;)
The PC RDNA2 cards are not made on 7nm EUV, they're 7nm DUV
 
The benefit would be marginal tbh. The games will still be designed around the original PS5 so a higher clocked GPU in the slim model may definitely suffice but it won't be substantial. For a pro model that makes more sense. Notice the PS4 slim maintained the same GPU clocks as the original PS4? Yet the PS4 pro got a 12% boost? I would expect the exact same if they were to release a slim and pro model. Both of them would be substantially smaller than the current PS5 though. But the pro could possibly double the compute units and have a slightly higher clock. Probably the GPU would be equivalent to a 6800 XT.

The irony in PS5 Pro going for a 72 CU design when "narrow and fast" was spoken of so positively earlier on xD. And make no mistake, it has its benefits, it's just more me being an ass at the expense of Cerny's and, more specifically, certain circle's focus on that design philosophy. FWIW what's even considered "narrow" changes over time for the architectures, at one point 40 CUs was considered "wide" for AMD but nowadays that's clearly not the case.

I don't know how based the RDNA 3 dual 80 CU chiplet rumors are but if that turns out true then a 72 CU design becomes the new "narrow" I guess (but for Sony it will absolutely impact PS6 and that's something they will want to be considerate of due to increasing costs on smaller nodes along with whatever new tech comes with that baked into the silicon increasing size and therefore price).

The PC RDNA2 cards are not made on 7nm EUV, they're 7nm DUV

Really? Even better then; they've pulled off some really impressive stuff at AMD then with these clocks. Bodes very well for future designs on EUV, let alone 5nm or 3nm combined with further architectural refinements. Will keep them competitive with Nvidia and Intel certainly (one way to phrase that).
 
The irony in PS5 Pro going for a 72 CU design when "narrow and fast" was spoken of so positively earlier on xD. And make no mistake, it has its benefits, it's just more me being an ass at the expense of Cerny's and, more specifically, certain circle's focus on that design philosophy. FWIW what's even considered "narrow" changes over time for the architectures, at one point 40 CUs was considered "wide" for AMD but nowadays that's clearly not the case.

I don't know how based the RDNA 3 dual 80 CU chiplet rumors are but if that turns out true then a 72 CU design becomes the new "narrow" I guess (but for Sony it will absolutely impact PS6 and that's something they will want to be considerate of due to increasing costs on smaller nodes along with whatever new tech comes with that baked into the silicon increasing size and therefore price).
You generally get better performance from increasing parallelism. In this case increasing the number of compute units instead of increasing the clock speed. You wont hear too much about the narrow vs wide argument if they make a PS5 pro. They'll simply increase the number of compute units and slightly increase the clock speed. The PS6 will probably be the last physical console. I wouldn't be surprised if they use ARM for the CPU tbh. Much better performance per watt.
 
You generally get better performance from increasing parallelism. In this case increasing the number of compute units instead of increasing the clock speed. You wont hear too much about the narrow vs wide argument if they make a PS5 pro. They'll simply increase the number of compute units and slightly increase the clock speed. The PS6 will probably be the last physical console. I wouldn't be surprised if they use ARM for the CPU tbh. Much better performance per watt.
Generally yes, although not all algorithms lend themselves to ever-increasingly parallelism. The problem that increasing parallelism brings is that you also need a commensurate increase in memory bandwidth per clock - or more cache. In terms of semiconductor density cache is expensive and bandwidth to RAM isn't an infinite resource either.
 
Generally yes, although not all algorithms lend themselves to ever-increasingly parallelism. The problem that increasing parallelism brings is that you also need a commensurate increase in memory bandwidth per clock - or more cache. In terms of semiconductor density cache is expensive and bandwidth to RAM isn't an infinite resource either.
Agreed! The trend seems to be increasing cache sizes, optimizing the memory heirarchy, increasing memory bandwith, using RISC architecture and accelerators. A lot of machine learning algorithms work well with processors that have given parallelism according to Jim Keller. And for gaming workloads, shader programs, ray tracing algorithms all these work best with processors that are highly paralell in nature. If you can get more compute units, more accelerators for certain workloads then you don’t need to clock your processors so high. I think Apple’s ARM chips has been one of the best examples of this. They don’t have the highest clocked processors but a lot of architectural choices they’ve taken have put them at the top
 
Generally yes, although not all algorithms lend themselves to ever-increasingly parallelism. The problem that increasing parallelism brings is that you also need a commensurate increase in memory bandwidth per clock - or more cache. In terms of semiconductor density cache is expensive and bandwidth to RAM isn't an infinite resource either.

True and keeping an eye on developments of memory will play a really big role in all of this; what's the point of pushing TF monsters if the memory bandwidths and capacities needed to keep them fed end up creating $1,000 consoles? The same issue comes up if you want to supplement main memory with on-chip cache; Infinity Cache for example works wonderfully for rasterization tasks (seemingly), but it's not of much benefit for other things like ML and RT, though maybe that's due more to lack of hardware accelerators like Tensor Cores being present (and Tensor Cores seem to work very well for their tasks with much less on-chip cache, though maybe they could have even better performance with additional large block of L3$).

And all of that comes with the extra concerns for increasing prices on smaller nodes; letting nodes mature in years helps with this somewhat though. 10th-gen consoles on N3P will be a lot cheaper in terms of wafer costs vs. PS5 and Series X's costs on 7nm, because by the time 10th-gen systems would be ready to go into production TSMC N3P will be a very mature process (~5 years) and personally I think for the performance targets Sony and Microsoft will aim for 10th-gen N3P gives them that as long as sufficient architectural advancements are made on RDNA.

Agreed! The trend seems to be increasing cache sizes, optimizing the memory heirarchy, increasing memory bandwith, using RISC architecture and accelerators. A lot of machine learning algorithms work well with processors that have given parallelism according to Jim Keller. And for gaming workloads, shader programs, ray tracing algorithms all these work best with processors that are highly paralell in nature. If you can get more compute units, more accelerators for certain workloads then you don’t need to clock your processors so high. I think Apple’s ARM chips has been one of the best examples of this. They don’t have the highest clocked processors but a lot of architectural choices they’ve taken have put them at the top

Out of those things we can say the current-gen systems have essentially taken care of the memory hierarchy part and mostly taken care of the memory bandwidth; on both metrics they hit very good ratios where they aren't bandwidth-starved or data-starved for new data in RAM when it's needed. It's the other three points you mention which 10th-gen systems can particularly improve on.

If RAM doesn't see significant improvements (either in terms of GDDR, HBM or some new memory technology coming along competitive with them) within the next 4-5 years then Sony & Microsoft will make up for it with focusing on larger on-chip caches but smaller increases in memory capacity and bandwidth. If power consumption still remains too heavy to fit in ever-tightening TDP budgets for newer consoles and RISC still isn't an option, then Sony & Microsoft will use that as one of several reasons to focus on more hardware accelerators (wafer pricing and BOM pricing would be the other major reasons).

For 10th-gen I personally think Sony will again focus on a narrower design but heavier focus on hardware accelerators and lower parallelism so they can keep costs lower to move more unit volumes and probably shift some of the excess BOM budget to other things (like an even tighter, more integrated VR as standard). They will probably also focus more on chip cache size and a slightly smaller memory capacity bump, but fast main memory.

Microsoft, honestly I see them going for an even more PC-esque design approach for several reasons, more modularity and flexibility in the design in terms of upgrading components (similar to a PC), by in default having a wider GPU with more raw TF computational potential, smaller on-chip caches (but what they have can potentially run faster), larger pool of main memory but segmented with SAM/BAR acting as a way to mitigate nUMA drawbacks, and likely a more generic, softer focus on hardware acceleration. Whatever they design will do well by Azure/cloud though, obviously.
 
True and keeping an eye on developments of memory will play a really big role in all of this; what's the point of pushing TF monsters if the memory bandwidths and capacities needed to keep them fed end up creating $1,000 consoles? The same issue comes up if you want to supplement main memory with on-chip cache; Infinity Cache for example works wonderfully for rasterization tasks (seemingly), but it's not of much benefit for other things like ML and RT, though maybe that's due more to lack of hardware accelerators like Tensor Cores being present (and Tensor Cores seem to work very well for their tasks with much less on-chip cache, though maybe they could have even better performance with additional large block of L3$).

And all of that comes with the extra concerns for increasing prices on smaller nodes; letting nodes mature in years helps with this somewhat though. 10th-gen consoles on N3P will be a lot cheaper in terms of wafer costs vs. PS5 and Series X's costs on 7nm, because by the time 10th-gen systems would be ready to go into production TSMC N3P will be a very mature process (~5 years) and personally I think for the performance targets Sony and Microsoft will aim for 10th-gen N3P gives them that as long as sufficient architectural advancements are made on RDNA.



Out of those things we can say the current-gen systems have essentially taken care of the memory hierarchy part and mostly taken care of the memory bandwidth; on both metrics they hit very good ratios where they aren't bandwidth-starved or data-starved for new data in RAM when it's needed. It's the other three points you mention which 10th-gen systems can particularly improve on.

If RAM doesn't see significant improvements (either in terms of GDDR, HBM or some new memory technology coming along competitive with them) within the next 4-5 years then Sony & Microsoft will make up for it with focusing on larger on-chip caches but smaller increases in memory capacity and bandwidth. If power consumption still remains too heavy to fit in ever-tightening TDP budgets for newer consoles and RISC still isn't an option, then Sony & Microsoft will use that as one of several reasons to focus on more hardware accelerators (wafer pricing and BOM pricing would be the other major reasons).

For 10th-gen I personally think Sony will again focus on a narrower design but heavier focus on hardware accelerators and lower parallelism so they can keep costs lower to move more unit volumes and probably shift some of the excess BOM budget to other things (like an even tighter, more integrated VR as standard). They will probably also focus more on chip cache size and a slightly smaller memory capacity bump, but fast main memory.

Microsoft, honestly I see them going for an even more PC-esque design approach for several reasons, more modularity and flexibility in the design in terms of upgrading components (similar to a PC), by in default having a wider GPU with more raw TF computational potential, smaller on-chip caches (but what they have can potentially run faster), larger pool of main memory but segmented with SAM/BAR acting as a way to mitigate nUMA drawbacks, and likely a more generic, softer focus on hardware acceleration. Whatever they design will do well by Azure/cloud though, obviously.

Yeah the "infinity cache" especially on a GPU just doesn't provide as much performance gain over increasing memory bandwidth. You'd do better by having larger on chip caches on the CPU and higher memory bandwidth on a unified memory architecture.Thats why neither machine has it.

And as you correctly pointed out with the advancements in the nodes, the 10th gen will have great hardware regardless. Maybe we'll have hardware capable of full path tracing at run time? With regards to RAM I would be surprised if neither companies(and especially MSFT) didn't go with HBM since they have serious cloud gaming ambitions. You simply get higher memory bandwidth and lower energy consumption than going with whatever GDDR RAM becomes available during that time. So you want to do this for your servers. The other thing again is ARM provides much better performance per watt than x86. Again a major advantage if you are running cloud computing operations. I would be surprised if both companies aren't looking at what Apple is doing with their processors and reconsidering x86 CPUs. MSFT already has translation software from x86 to ARM. I bet you they have Xbox software running on ARM processors. Hell I would be surprised if Sony doesn't as well since they're the kings of console gaming.

If MSFT went with an upgradable console, they might as well just shut down the Xbox hardware division. If there's anything, they should continue with what they did with the Series X in particular. A highly custom well thought out computing device for gaming. Everything from the processors all the way up to the OS and cloud computing services has been functionally better than their competition this time. They should and most likely will maintain this. Only qualm people have had are with the UI being a bit old and the lack of gaming content. The only thing they need to work on are their studios. They need more AAA third person titles imho and develop sdlc to ensure that they can efficiently create AAA software for their subscription services. They already have a solid outlook for first person and indie titles.

Basically the hardware for the next gen machines is going to be pretty powerful but pretty predictable. minimum 8x GPU performance in traditional GPU tasks. Accelerators for ML and Ray Tracing, some form of high bandwidth memory to support these new processors(So HBM memory), minimum doubling of RAM, and at least a 16 core CPU. The only question is if there will be disc models from MSFT.
 
My expectations for the PS6 is

CPU - 8Core, 16 thread Zen 4 @4.5ghz
(Most games don't max out 8 core CPUs nowadays anyway. I think 8 core with increased clock and IPC will do the trick)

GPU - RDNA 4 72CU x 2.925ghz 27tflop (2nm)
Ray Tracing cores, ML cores
(Alot more GPU die space will be taken up with tensor and RT core equiv than PS5)
Will follow the PS4 Pro revision method of butterfly GPU for backward compatibility)

RAM - 24gb GDDR6X
384 bit bus
Bandwidth - 936gbs

2TB SSD
12gbs raw - 24gbs compressed.

Playstation is far easier to guess than the Next box is for sure. I have no doubt Sony will go for a 72cu GPU for Backward compatibility ease.
 
My expectations for the PS6 is

CPU - 8Core, 16 thread Zen 4 @4.5ghz
(Most games don't max out 8 core CPUs nowadays anyway. I think 8 core with increased clock and IPC will do the trick)

GPU - RDNA 4 72CU x 2.925ghz 27tflop (2nm)
Ray Tracing cores, ML cores
(Alot more GPU die space will be taken up with tensor and RT core equiv than PS5)
Will follow the PS4 Pro revision method of butterfly GPU for backward compatibility)

RAM - 24gb GDDR6X
384 bit bus
Bandwidth - 936gbs

2TB SSD
12gbs raw - 24gbs compressed.

Playstation is far easier to guess than the Next box is for sure. I have no doubt Sony will go for a 72cu GPU for Backward compatibility ease.
Well that's just not going to happen, Zen 4 and even the hypothetical RDNA4 will be ancient when we get to 2nm.
 
My expectations for the PS6 is

CPU - 8Core, 16 thread Zen 4 @4.5ghz
(Most games don't max out 8 core CPUs nowadays anyway. I think 8 core with increased clock and IPC will do the trick)

GPU - RDNA 4 72CU x 2.925ghz 27tflop (2nm)
Ray Tracing cores, ML cores
(Alot more GPU die space will be taken up with tensor and RT core equiv than PS5)
Will follow the PS4 Pro revision method of butterfly GPU for backward compatibility)

RAM - 24gb GDDR6X
384 bit bus
Bandwidth - 936gbs

2TB SSD
12gbs raw - 24gbs compressed.

Playstation is far easier to guess than the Next box is for sure. I have no doubt Sony will go for a 72cu GPU for Backward compatibility ease.

I think I mostly agree with this except the RAM. I don't think it will be GDDR6X. From a power consumption view, I think they will have to go with HBM or some other similar technology. I don't think GDDR6X on a 384-bit bus is really feasible for a console. Maybe there's a GDDR7 with vastly improved GB/s/Watt, but I don't know
 
My expectations for the PS6 is

CPU - 8Core, 16 thread Zen 4 @4.5ghz
(Most games don't max out 8 core CPUs nowadays anyway. I think 8 core with increased clock and IPC will do the trick)

GPU - RDNA 4 72CU x 2.925ghz 27tflop (2nm)
Ray Tracing cores, ML cores
(Alot more GPU die space will be taken up with tensor and RT core equiv than PS5)
Will follow the PS4 Pro revision method of butterfly GPU for backward compatibility)

RAM - 24gb GDDR6X
384 bit bus
Bandwidth - 936gbs

2TB SSD
12gbs raw - 24gbs compressed.

Playstation is far easier to guess than the Next box is for sure. I have no doubt Sony will go for a 72cu GPU for Backward compatibility ease.

The thing is will a 72 CU RDNA 4 GPU provide the next gen performance? By 2027 it will be old tech by a mile. And If they decided to have a PS5 pro in the next 4 four years they'd likely have to use something close to the 6800 XT which has 72 CUs. The PS6 should have a GPU thats significantly more powerful than a 6800 XT. 40 shader tflops should be the target for such a machine.

For the CPU I agree about 8 cores not being maxed out right now. I think I remember the Dirt developer saying they were only utilizing between 50-80% per core. But don't you think by mid gen this won't be the case? By 2027 most entry level laptops will come standard with 16 core CPUs. One would expect the next gen machines to at least have 16 core CPUs. Also with larger accelerators for Ray Tracing and ML 16 cores at a minimum makes more sense. You don't want the CPU being a bottleneck again.

For the RAM what is the outlook on GDDR7? I think it's going to be easier to predict what MSFT will do since they have more serious cloud gaming ambitions. They'll definitely go with HBM. Otherwise 936GB/s on the next gen Sony machine seems low. The accelerators and larger GPU would definitely be bandwidth starved. You've at least doubled the GPU CUs, added larger areas for RT and ML as well so you can't simply double the memory bandwidth. They'd need closer to 2TB/s. And thats why HBM seems the best DRAM configuration.

I completely agree with the SSD. 12GB/s before decompression would be enough. They can already do 2.5 decompression ratio so the performance would be incredible.
 
With regards to RAM I would be surprised if neither companies(and especially MSFT) didn't go with HBM since they have serious cloud gaming ambitions. You simply get higher memory bandwidth and lower energy consumption than going with whatever GDDR RAM becomes available during that time. So you want to do this for your servers.

I've been tracking timing of HBM version introductions into the market and as a memory technology it's seen a lot of gains in a relatively short amount of time. Much more than GDDR-based memories have, for sure. With rumors that the next HBM (HBM3) will be an architectural redesign (I hope it takes some aspects from the FGDRAM ideas because that would make it quite better in virtually every aspect), I see a similar rate of progress continuing over the next few years.

By the time of 10th-gen systems (assuming they stay on track for a 2026/2027 release...though there's a chance they could come later than that), HBM4 or HBM4E-style variants pushing 6+ Gbps should be the norm, and I expect it to allow 4 GB module capacities by then with 16-Hi stacks being a new upper limit. That's a pretty easy path to 32 GB capacities and any mixture of bandwidth between 800 GB/s to nearly 2 TB/s. The only thing that could limit this would be, of course, pricing. HBM prices have to stay affordable for future designs for mass-market devices or else Sony & Microsoft could bite the bullet and jack up the MSRP for their systems.

The other thing again is ARM provides much better performance per watt than x86. Again a major advantage if you are running cloud computing operations. I would be surprised if both companies aren't looking at what Apple is doing with their processors and reconsidering x86 CPUs. MSFT already has translation software from x86 to ARM. I bet you they have Xbox software running on ARM processors. Hell I would be surprised if Sony doesn't as well since they're the kings of console gaming.

Personally it's either between ARM and RISC-V as the future; I prefer RISC-V because it's an open platform, but it's not as mature as ARM (and lacks its financial backing). Many companies are taking baby steps in making RISC-V designs though so that is a good path forward. In either case, getting x86 compatibility on ARM or RISC-V designs will always have the two issues of reduced performance (by how much will definitely improve as time goes on, though it will mean potentially more initial processor overhead) and the x86 legal license itself. Only Intel and AMD are legally able to emulate x86 microcode, AFAIK.

Out of the two companies, while Sony definitely enjoys more marketshare in the console space, I think Microsoft are quite further ahead in terms of designing ARM processors in the R&D phase. They have more reasons as to where it'd be an initiative. If Sony still had their VAIO line around I could picture them being more invested in ARM-based design prototyping and serious R&D, but I think they'll wait until AMD gets there before planning deeper on that front.

If MSFT went with an upgradable console, they might as well just shut down the Xbox hardware division. If there's anything, they should continue with what they did with the Series X in particular. A highly custom well thought out computing device for gaming. Everything from the processors all the way up to the OS and cloud computing services has been functionally better than their competition this time. They should and most likely will maintain this.

Ah this one's my mistake for not phrasing it eloquently; I agree that they will continue forward with Series X, if it sounded like I was saying they'd just "make a PC" that was bad wording. Microsoft is in a really good position, though, to synergize the console and PC designs in a way no one else really can, and 10th-gen would be a great chance to do it. We're still talking strongly custom GPU, CPU etc. designs, memory I/O features etc. The difference in this case though being they can open up aspects of the design for (tightly controlled) upgradability through official, specific hardware components.

The comparison that might work better here, ironically, is their Elite controllers; you can customize though but the parts still need to be officiated and you can't just throw any random brand of buttons or sticks in there and expect the same quality. So them continuing what they've done with Series X, they can keep that highly custom focus while having a planned path for component upgrades that stay within the capabilities of the managing chipset, designed in tandem with AMD and for other things, with Micron, Seagate etc. A fusion of proven and new technologies in one package that allows for user-level scalability depending on desired use-cases.

This...kind of goes into another aspect of 10th-gen I think Microsoft will try pursuing though; since the Series X already serves a dual-purpose for both gaming and Azure, their next system will continue that but likely also feature a design that can serve console gaming and more general PC work/gaming etc. So two SKUs where the base hardware is the same, but some of the expandability is cut back on one vs. the other, two selling models (the console selling at a loss, the other version selling at a profit), two variant OSes (Xbox OS on one, Windows on the other), same base levels of performance etc. I think there's quite a few ways they can make this work but I'll try talking about it a bit more when I can put up some more ideas on what I think 10th-gen systems might try doing.

Basically the hardware for the next gen machines is going to be pretty powerful but pretty predictable. minimum 8x GPU performance in traditional GPU tasks. Accelerators for ML and Ray Tracing, some form of high bandwidth memory to support these new processors(So HBM memory), minimum doubling of RAM, and at least a 16 core CPU. The only question is if there will be disc models from MSFT.

8x general GPU perf. increase might still be asking a tad much IMO, but that's where accelerators come into the picture, to offload on those kind of things. There's also potentially a chance they still stick with 8C CPUs; whether they increase the thread count though would be interesting. All the other stuff though is definitely in the pipes because, as you say, it's a bare minimum.

2024 is when TSMC is talking about production of 2nm.

Those are the plans, but will they be able to hit them? I think if Sony and Microsoft are okay with pushing 250 watt systems 10th-gen, they can actually get pretty sizable performance increases on N3P, and the node itself would be cheaper than 2nm while being a more mature process.

If they want to push 200 watt or sub-200 watt 10th-gen systems, though, then they'll probably need 2nm, and 2nm will have to offer at least 25% power consumption reduction benefits over N3P...it might also mean slightly later 10th-gen launches (2028 maybe? I've seen some people also mentioning COVID could cause this and that's a factor for sure).

Also funnily enough I think PS6 is in some ways the harder of the two to guess xD; for me it has more to do with chip size than anything. Also I think they are going to want to push even further on VR for 10th-gen and try making it standard; that will necessitate more specialized hardware and lower general GPU performance but still good enough for high-fidelity VR (if they were to include a VR helmet though it'd definitely be the lower/entry-level type with the bare necessities and little else).
 
The thing is will a 72 CU RDNA 4 GPU provide the next gen performance? By 2027 it will be old tech by a mile. And If they decided to have a PS5 pro in the next 4 four years they'd likely have to use something close to the 6800 XT which has 72 CUs. The PS6 should have a GPU thats significantly more powerful than a 6800 XT. 40 shader tflops should be the target for such a machine.

We both agree on the general TF target, but I still can't shake the idea out of me that they'll stick with a very narrow design. AMD's architecture plays very friendly with clocks, and if Sony can exploit that going forward while leveraging hardware accelerators for other types of heavy lifting not tied to general GPU tasks, it helps in getting great performance at a lower price. It's basically what they've already done for PS5.

So I guess that's my way of speculating they will go for yet-higher clocks next time but they don't need a 72 CU chip to do that. And if they stick to, say, some 36 CU - 40 CU design, it doesn't mean the CUs stay the same size; they'll likely be bigger again, as PS5's were to PS4's, and there's a few things AMD and/or Sony could add in to facilitate that actually such as more shader cores.

For the CPU I agree about 8 cores not being maxed out right now. I think I remember the Dirt developer saying they were only utilizing between 50-80% per core. But don't you think by mid gen this won't be the case? By 2027 most entry level laptops will come standard with 16 core CPUs. One would expect the next gen machines to at least have 16 core CPUs. Also with larger accelerators for Ray Tracing and ML 16 cores at a minimum makes more sense. You don't want the CPU being a bottleneck again.

That's a good point; is there maybe any way something like FPGA cores on the CPU that can be programmed for assisting in drawcalls and other graphics instructions can be leveraged in lieu of 16 core CPU designs? Especially if it can be the cheaper and more power-efficient option?

For the RAM what is the outlook on GDDR7? I think it's going to be easier to predict what MSFT will do since they have more serious cloud gaming ambitions. They'll definitely go with HBM. Otherwise 936GB/s on the next gen Sony machine seems low. The accelerators and larger GPU would definitely be bandwidth starved. You've at least doubled the GPU CUs, added larger areas for RT and ML as well so you can't simply double the memory bandwidth. They'd need closer to 2TB/s. And thats why HBM seems the best DRAM configuration.

It's a bit funny because if you take the timing track of previous GDDR releases, GDDR7 should theoretically come to market around 2024 and probably around 24 Gbps for the first iteration or so. But if they'd want GDDR7 to work for a 10th-gen system at non-obscene capacity amounts or bus widths (due to associated costs), they'd need 32 Gbps modules, and it would be some years for that to happen.

So it'd come down to, will GDDR7 be more affordable than whatever HBM-based memory is there by the time? And would it be by enough? It's pretty difficult to say. They wouldn't even need the then-latest HBM memory to match GDDR7 performance depending on the capacity for memory they'd want, IMO.

I completely agree with the SSD. 12GB/s before decompression would be enough. They can already do 2.5 decompression ratio so the performance would be incredible.

Yeah, 12 GB/s (or 16 GB/s) would be all needed for 10th-gen systems.
 
We both agree on the general TF target, but I still can't shake the idea out of me that they'll stick with a very narrow design. AMD's architecture plays very friendly with clocks, and if Sony can exploit that going forward while leveraging hardware accelerators for other types of heavy lifting not tied to general GPU tasks, it helps in getting great performance at a lower price. It's basically what they've already done for PS5.

So I guess that's my way of speculating they will go for yet-higher clocks next time but they don't need a 72 CU chip to do that. And if they stick to, say, some 36 CU - 40 CU design, it doesn't mean the CUs stay the same size; they'll likely be bigger again, as PS5's were to PS4's, and there's a few things AMD and/or Sony could add in to facilitate that actually such as more shader cores.

I think fundamentally Sony has to simply accept that they didn't build the overall best hardware possible in the PS5. For their long term goals it must have been great though because they'll be able to produce the PS5 with high margins throughout the generation! But if they decide to make a pro model Almost certainly they're going to base it on either the 6800 or 6800XT. The whole narrow vs wide argument will fly out the window with that model. Empirical data shows adding more silicon provides higher performance compared to increasing clocks. And I keep on harping about parallelism because thats just the fundamental nature of computers and the reason why we still don't have 5GHz CPUs for the mass market. You'd get better performance per watt by simply increasing the number of cores instead of increasing the clock speed. And going on to the next gen PS6, if anything they shouldn't clock the GPU too high again unless they want to launch a large underpowered machine like they did this time. Moving forward the focus will be on size as well as performance.

Besides the whole narrow vs wide argument Sony made, the SSD's impact on gaming was not properly explained by Sony while MSFT undersold their SSD.
So basically don't expect a PS5 pro model with very high clock speeds. If made it will simply hit between 60-72 CUs on the GPU with about 2.5GHz or less.

That's a good point; is there maybe any way something like FPGA cores on the CPU that can be programmed for assisting in drawcalls and other graphics instructions can be leveraged in lieu of 16 core CPU designs? Especially if it can be the cheaper and more power-efficient option?

This would be impractical. No developer needs to waste time re-programming FPGA cores.
 
I think fundamentally Sony has to simply accept that they didn't build the overall best hardware possible in the PS5. For their long term goals it must have been great though because they'll be able to produce the PS5 with high margins throughout the generation! But if they decide to make a pro model Almost certainly they're going to base it on either the 6800 or 6800XT. The whole narrow vs wide argument will fly out the window with that model.

Can't deny indeed that prices were a driving factor (arguably THE driving factor) for how Sony's designed the PS5. Good benefits but also a lot of tradeoffs for sure.

Empirical data shows adding more silicon provides higher performance compared to increasing clocks. And I keep on harping about parallelism because thats just the fundamental nature of computers and the reason why we still don't have 5GHz CPUs for the mass market. You'd get better performance per watt by simply increasing the number of cores instead of increasing the clock speed. And going on to the next gen PS6, if anything they shouldn't clock the GPU too high again unless they want to launch a large underpowered machine like they did this time. Moving forward the focus will be on size as well as performance.

I think you have to keep in mind that certain things do benefit from higher clocks though, and not all hardware components need higher scaled parallelized general compute in order to themselves be increased, depending on the flexibility of the architecture. For something focused a lot on data crunching and various compute tasks, wider parallelism is always going to benefit, but you're never guaranteed the maximum increase in certain aspects more beneficial to gaming in particular, like pixel and texture fillrates, with that.

So in that aspect I can see why Sony's made some of the design decisions they made, and that's not at the dismissal of Microsoft's choices because they are very competitive on those fronts regardless. But taking that into account and, again, the relative friendliness with clocks RDNA seems to play with, I don't think a necessarily wider approach is the only option for building a high-performance 10th-gen console. It's one approach, and one I think Microsoft will favor, but not the only one.

However, taking a narrower approach again is going to have to rely on flexibility in other aspects of the architecture design; if those are not very flexible or their flexibility is strictly tied to general compute scaling, then yeah you might not have much a choice but to go wider in the future. I think we'll need more GPU examples from AMD over the next couple of years to see how flexible their architecture really is and how granular that flexibility is, but at the very least I am going to have to consider a wider PS6 as a possibility for 10th-gen xD.

This would be impractical. No developer needs to waste time re-programming FPGA cores.

The thing with the way I'm seeing it though is that the devs wouldn't actually be doing any of the programming themselves; basically having a bank of pre-programmed models to configure the FPGA stored in some private cache (or potentially a bank of NOR flash for XIP functionality, if they are okay with a decent-size NOR flash capacity...maybe a small bank of embedded MRAM if the prices make it feasible at that point) that devs would select from and load to configure the FPGA cores.

Sony and Microsoft could handle all of this themselves, update the stored data as desired, and leave things open for devs who want to implement their own code to do so (NOR flash wouldn't be suitable for this, so either possibly MRAM or just some type of SRAM though in the SRAM case they would need to store the code itself to the SSD or, preferably, some private block of embedded NAND that can be re-loaded as needed).
 
I think fundamentally Sony has to simply accept that they didn't build the overall best hardware possible in the PS5.

Its not Sony thats having a problem with that.
Midrange hw (5700xt perf) was to be experted from Sony since github did a godsent.
 
And Zen 4 is early 2022, (hypothetical since AMD hasn't actually mentioned it on roadmaps yet) RDNA4 should be either late 2022 or early 2023

Well, going by the 16-month intervals that've been established so far for RDNA (granted it's not much of a pattern but it's something), in a best-case scenario RDNA 3 should be ready by early 2022. However, a few things could possibly that: the chip shortages, lower-than-anticipated sales due to lack of supply to satisfy demand with RDNA 2 cards, and ongoing strain on developments due to lockdowns backing up all of the channels in terms of manufacture, distribution etc.

I would just hope that those factors don't push RDNA 3 back to late 2022 because unless they're throwing out a ton of existing things through massive architecture redesigns or chiplets are proving to be a real pain to implement, there's nothing on the R&D front that seems it'd hold back release to late 2022/early 2023. Even if they're working on more hardware accelerators, that shouldn't increase the window between 2 and 3 too drastically (I'd hope).
 
Back
Top