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

im thinking 80tflops is a possibility for nexgen consoles?

I don't believe so. Unless some new manufacturing technique develops that can significantly reduce power consumption. To get that TF number, I think we would need to be at a node that is beyond 3nm and it would likely have to be MCM based. The bandwidth requirements would like have to be on the order of multi-terabyte/sec as well. I don't see a feasible technology for that on the horizon yet. This is even way beyond what HBM is currently capable of.

I still think next gen will be 20-30 TFlops plus a boat load of dedicated HW acceleration for raytracing and ML techniques (resolution up sampling, frame rate interpolation).

It seems the current gen will be severely lacking on the ML inference side. When you look at the peak tensor ops of Turing for the various flavors of INT8 and FP16 using the tensor cores, they are far beyond what the GPU cores can do and can be done in parallel with the GPU.

At least what has been disclosed so far from MS/Sony is that RDNA will simply allow for FP16 and INT8 ops to run on the GPU cores. Which is ok, but already behind what the state of the art is.
 
Yes, and I would say, that IF we do go down the route of mid-gen consoles again, in 2023/2024, the main draw for them would be greatly increased ray tracing performance over the base consoles, with Zen 4 or Zen 5 CPUs and RDNA 4 GPUs
It would not be for pushing 8K.

It would also allow Sony to do 60fps versions of all their big games, with ray tracing at 4K.

That said, I'd still prefer there to be no mid-gen consoles and we go straight to PS6 and Future Xbox in late 2026. All the software ecosystems and your services/games will just roll right onto the new hardware easier than ever.
WRT mid-gen vs not I think Sony would rather have the PS6 being the next step up with updated everything and hopefully something PCs won't match for a few years in some specific area regarding exclusives. Whether or not that comes to pass depends on sales in 3 or 4 years which know one knows anything about.

With a refresh I would assume still Zen2 and some RDNA 2/3 amalgam with more CUs chasing RT performance, as you say, and maybe ML. Some better GDDR6 bandwidth but not much of an update regarding the SSD except for size. If co-processors become a thing maybe they just increase the CU count and memory bandwidth and add whatever specific cores needed.

A PS6 that can somehow leverage what the ARM could do with emulating X86 is an outside possibility but that would likely see it out in 2027 or maybe later. At that point you could see something very interesting coming from Sony as it would allow them to differentiate themselves from regular PC hardware increasing the exclusivity of the brand.

More likely a PS6 would have Zen 4 with RDNA 4 or 5 with a fair amount of very fast memory tied to Crosspoint type Non-volatile storage if Sony were to double down on reducing bottlenecks from storage and the cost wasn't as insane. Starting with 1 TB of the stuff and a piece of gum sized SSD with 4 TB on it to hold everything else you would need or just use the network at that point. No optical disk at that point I would assume.

The Series X would just bump up specs every 3 or 4 years or so to keep up with what PCs are doing to minimize developer multiplatform friction.
 
Last edited:
I don't believe so. Unless some new manufacturing technique develops that can significantly reduce power consumption. To get that TF number, I think we would need to be at a node that is beyond 3nm and it would likely have to be MCM based. The bandwidth requirements would like have to be on the order of multi-terabyte/sec as well. I don't see a feasible technology for that on the horizon yet. This is even way beyond what HBM is currently capable of.

I still think next gen will be 20-30 TFlops plus a boat load of dedicated HW acceleration for raytracing and ML techniques (resolution up sampling, frame rate interpolation).

It seems the current gen will be severely lacking on the ML inference side. When you look at the peak tensor ops of Turing for the various flavors of INT8 and FP16 using the tensor cores, they are far beyond what the GPU cores can do and can be done in parallel with the GPU.

At least what has been disclosed so far from MS/Sony is that RDNA will simply allow for FP16 and INT8 ops to run on the GPU cores. Which is ok, but already behind what the state of the art is.
Tensor cores can not run parallel with shader cores.
 
Tensor cores can not run parallel with shader cores.

interesting. I thought that was the case as I based the assumption off this slide.

dxr-on-old-gpus.jpg


Maybe this just means most of the shaders are doing tensor work and only some are doing the FP32 work.
 
interesting. I thought that was the case as I based the assumption off this slide.

dxr-on-old-gpus.jpg


Maybe this just means most of the shaders are doing tensor work and only some are doing the FP32 work.
Everything I've read in the architecture threads here has stated tensor cores can not run concurrently with shader work due to lack of shared resources. Tensors saturate them all.
 
From my perspective, both console makers will look to release a higher end pro console with 5nm along with or even instead of just the regular shrinking of the existing design.

These consoles are pretty powerful but still aren't really strong enough to really push RT fully. Also SSD speeds are increasing massively and MS will probably increase their speed as the cost goes down. I think MS will be looking to turn their console to be less generational and more just tiered anyways.

New tech they could add like tensor cores or just general AI processing elements. HBM might finally become cheap enough for consumer products which would be a pretty big boon.
 
Maybe we will see entirely different architectures (both ISA and µarch) on next-gen GPUs, where they can handle divergence more gracefully so that they can accelerate ray tracing works more "natively"?
 
So with the Ampere / RTX 30 series reveals out of the way (we have yet to really understand what Ampere has, does not have, what it does, what it does not do, etc. that will come in the weeks and months to follow). I was going to ask again how do you feel AMD/Sony/Microsoft will go about designing the silicon for future next-gen consoles, perhaps as early as 2026, but more likely not before Holiday 2027.

What seriously awesome/impressive stuff is not outside the realm of possibility, silicon and RAM wise? Within reason and that could be feasible for $500-$600 consoles ,at some point during the back half of this decade?
 
Also I think we need to define what ray tracing means, do we mean as with an offline renderer casting multiple rays from each pixel? How many bounces do we trace for it to count? Is BVH good enough to qualify as "fully ray traced"? I expect for the next platform quality on par with whatever the 3080Ti on high quality RT offers with performance fast enough for a 4K image based off at least a 1440p base image (I still don't think we'll get to 4K "native" and RT by 2027/28).

I want RT for audio too with proper environmental occlusion, reverb, etc and I want much better AI. Hilariously of course I am saying I want "better" AI without defining that either, what I mean here is I don't want to be able to do what I can do today in Destiny/Division etc which is know that if I run up to an AI in cover I can predict what the AI will do next. Of course it is entirely possible I don't actually want that and that deterministic behaviour is a vital component of creating a fun consistent challenge but increasingly when I play games I don't feel like "I understand the Zar'k-lej warriors and can beat them now" it's more "oh that's what triggers ChangeCover(), if I do this I should be able to trigger HunkerDown() instead"
 
I've been thinking about this some since Ampere was announced. And I'm a little more pessimistic now. While the perf of Ampere looks great, it comes at a cost of an increased TPD. This is mostly fine for PC, but I don't think NVIDIA can really push much farther with their next line. For consoles, the power budget will be a big, big problem.

The node scaling from 7-5 and 5-3 only offers 30% power savings with each jump. So just a very rough estimate would put something like a 3090 (375 GPU power according to NVIDIA) out of the realm of possibility for a console APU without significant architectural improvements. Console manufacturers might have to wait until a process node beyond 3nm just to be able to get the needed performance at a reasonable power budget.
 
I've been thinking about this some since Ampere was announced. And I'm a little more pessimistic now. While the perf of Ampere looks great, it comes at a cost of an increased TPD. This is mostly fine for PC, but I don't think NVIDIA can really push much farther with their next line. For consoles, the power budget will be a big, big problem.

The node scaling from 7-5 and 5-3 only offers 30% power savings with each jump. So just a very rough estimate would put something like a 3090 (375 GPU power according to NVIDIA) out of the realm of possibility for a console APU without significant architectural improvements. Console manufacturers might have to wait until a process node beyond 3nm just to be able to get the needed performance at a reasonable power budget.
I'm not sure how we get around this. This is an issue for all computing here. Trying to continually do more work with less power is very difficult to engineer. The amount of processing power that is occurring in such a small die is really ramping up the watts/cm^2 and therefore the TDP. We're getting to the point that if you're up to 50-80 TF of power that's a lot of transistors in such a small space, the heat is going to keep going up and so will the power. There lies a point that no amount of cooling will be able to keep up with the amount of watts/cm^2. I'm not sure at what TF level that is at, but I suspect it's happening soon. There's no magic solution I see coming here except to
a) have newer engineering materials/methods
b) rethink the way we program to use less energy
c) go extremely wide

We're trying to parallelize the processors as much as possible by increasing the core count and reducing the clock rate, but you're going to run into performance issues trying to go so wide and eventually and die size issues. Memory isn't quite caught up yet with the rate in which we can increase compute power. IMO Major changes would have to occur on both the software and hardware side of things for us to get to the next break through.

IMO, this move to AI processing is a step in that direction. We're going to see more dedicated processors to increase performance while reducing the power footprint like RT cores. The future can't be just plain old compute and lots of it; while flexible, the limit for performance is scalability.

TF will soon be an outdated way to look at performance. Unless of course, we keep converting all these dedicated units into (TF of power like we saw with RT)
 
Back
Top