Baseless Next Generation Rumors with no Technical Merits [pre E3 2019] *spawn*

Status
Not open for further replies.
His sources also said Sony was targeting 2080 TI
So around the 13TF mark is crazy talk to you?
I've not settled on a figure personally but that wouldn't be far away from many leaks.
What TF are you expecting?

As for the benefits of 3 vs 2 threads:
Multiplats will see little overall improvement if game accessible core count is the same.
But for exclusives that optimise for it, could see reasonable performance benefits. But would also depend on how the upgraded pipeline works out.

As these apu's would be used for non gaming tasks in the cloud also I could see why they could go this route.
Still find it unlikely though. But not crazy or useless.
For strickly a console may not go that route, but for a multi purpose apu, may make decisions that would otherwise seem not worth considering.
Example could make running more remote desktops on a single apu more viable.
 
But for exclusives that optimise for it, could see reasonable performance benefits.
If you're already hand-crafting code to that level (which strikes me as a rare possibility), you'd already be maximising what the typical hardware can offer from two saturated threads. Again, you'd only get more work per thread with more execution units. This is how POWER8 did it and managed to get something meaningful from 8 threads. You'd only benefit a handful of exclusive games, for a whole CPU architecture.

For strickly a console may not go that route, but for a multi purpose apu, may make decisions that would otherwise seem not worth considering.
Example could make running more remote desktops on a single apu more viable.
Would anyone want an APU in servers? Other than virtual consoles, at which point you'd only be running one game on a machine.

It's worth noting that Intel has reduced threading to increase core count, with the i7-9700 being 8 cores, 8 threads as opposed to the previous 6 cores, 12 threads. The 8700 has a small advantage in Blender CPU raytracing benchmarks. The 9700 typically tends to do better in games. 6 cores, 12 threads is about the same as 8 cores, 8 threads, overall. 4 cores, 16 threads would be inferior to 8 cores, 16 threads in performance. Cheaper and smaller though, unless the cores get fatter. If Zen 3 is marked change in architecture, we'll have to wait and see.
 
Would anyone want an APU in servers? Other than virtual consoles, at which point you'd only be running one game on a machine.
I used remote desktop as an example as I know that's something MS is actively working on providing. Also other reason I used it as an example was I can see it giving a better experience due to better latency.
They have said they intend to use it for more than just games, so do you think the cloud will be a different architecture and not use the APU?

If Zen 3 is marked change in architecture, we'll have to wait and see.
Yea, agree, that's why I think it would really depend on how much the pipeline has been changed. This wouldn't be a minor update, otherwise just wouldn't be worth doing.
 
It's worth noting that Intel has reduced threading to increase core count, with the i7-9700 being 8 cores, 8 threads as opposed to the previous 6 cores, 12 threads.

I think that's about market segmentation rather than a statement about the utility of smt. The architecture still supports it, and it still offers gains when enabled.

If consoles support it there's no reason not to enable it.
 
I used remote desktop as an example as I know that's something MS is actively working on providing. Also other reason I used it as an example was I can see it giving a better experience due to better latency.
They have said they intend to use it for more than just games, so do you think the cloud will be a different architecture and not use the APU?
The Xcloud gaming system is set to be consoles in the sky of some form. That or 'PC's running Windows multiplat versions, for next-gen. Zen 3 isn't just being made for Xbox though. If AMD have eyes on using the architecture in servers, perhaps it's more thread heavy?
 
The Xcloud gaming system is set to be consoles in the sky of some form. That or 'PC's running Windows multiplat versions, for next-gen. Zen 3 isn't just being made for Xbox though. If AMD have eyes on using the architecture in servers, perhaps it's more thread heavy?
That's the bit we don't know much about, we know xcloud is X1S based, but we don't know what the follow up will be. I'm assuming Scarlett, as you say console in the cloud but not solely used for Xbox game streaming.

CPU and gpu performance for an APU should be very good, maybe not discreet level but good nonetheless, so I could see it being used for other work loads near the edge especially where latency is also a factor.

That's one of the reasons I'm saying in console isolation 3 thread may be crazy talk, but if it was also used in cloud then may make sense for console and cloud use cases. Azure & Xbox input into design.
 
To be fair it's also one way I think they can offset costs to console.
By using Scarlett in the cloud, would be purchasing a lot more, possibly different bining on apu, and be able to sell the resource to businesses, so overall higher utilisation and probably charge more to businesses.
 
No-one said that though. It's a possibility.
I'm saying I would think it's out there if it was just for Scarlett console. Not impossible, but the benefits over probably cheaper zen 2....

Never meant it to sound like I'm attributing it to someone else.
 
Next Xbox with Zen 3 and 3 way threads SMT?

SMT can, if used properly, give about a 28% boost. But what about the third thread? 12 to 14% more?
 
Does a quad channel memory get a huge boost in performance from dual channel?

It doubles the bandwidth. If you're interested I recently made a post showing the bandwidth of different channel setups (2-12 channels) with DDR4 3200: https://forum.beyond3d.com/posts/2068064/

How much of that translates into more performance depends on the application and it's usage.

Can next gen APUs use the already cheap 3000+ 32 GB memory in quad channel?

Do you mean in console APUs? Then yes, MS or Sony can choose anything that AMD already has available or can implement.

[...]

Would anyone want an APU in servers? Other than virtual consoles, at which point you'd only be running one game on a machine.

[...]

One advantage I can think of is the higher density you should be able to reach with semi-custom APUs compared to dedicated hardware. AMD had Opteron X APUs for that purpose in the past and is still researching APUs for HPC customers which have CPU and GPU chiplets in combination with HBM.

https://www.computermachines.org/joe/publications/pdfs/hpca2017_exascale_apu.pdf
 
Last edited by a moderator:
Next Xbox with Zen 3 and 3 way threads SMT?

SMT can, if used properly, give about a 28% boost. But what about the third thread? 12 to 14% more?
Optimized titles won't benefit from SMT if they're fully utilizing the core, there's nothing to switch too if the core doesn't stall.
The 3/4 thread bit for Xbox works if you're looking at it from the perspective of doing compute workloads on the servers and also from non-optimized titles/applications; quite a good fit imo.

But it does appear certainly to be an easier way to handle multi-threading. I know the GDC talk by Naughty Dog has gotten a lot of praise by looking in-depth how they managed to get all their 8 cores working in concert. I suppose with more hardware threads, you can rely on the hardware to do the heavy lifting. In a general case perhaps easier overall as a lift for most developers, but I don't see it being better than the absolute best optimized engines.

Outside of that, I don't think you're going to see a large benefit over a 2 thread processor, one might be able to debate that it could be detrimental for extreme optimization cases.
 
To be fair it's also one way I think they can offset costs to console.
By using Scarlett in the cloud, would be purchasing a lot more, possibly different bining on apu, and be able to sell the resource to businesses, so overall higher utilisation and probably charge more to businesses.
someone once commented on the possibility of sending cloud work to local consoles for processing (lol) as an opt-in. That would be interesting if you could be paid in MS rewards or what not to have your console on to grind away compute loads while you're not playing for credit.

I'd be curious to see the opportunity explored.
 
Outside of that, I don't think you're going to see a large benefit over a 2 thread processor, one might be able to debate that it could be detrimental for extreme optimization cases.
We already have real-world examples where HT slows down a game. That's not to say it's useless; only to keep in check any possible ideas of more threads == more performance. It means a different utilisation profile of the CPU and something devs will work (or not) with. My gut feeling is perhaps a 10% basic boost to general workloads with the CPU better able to schedule work and keep the CPU active. But of course, the more active it is, the hotter it runs, so the lower it needs to be clocked per core.

If there isn't already a discussion in the PC architecture forum, it's worth having one there as more knowledgable engineering folk may be able to weigh in.
 
I guess so. Likely the automatic workload dispatch isn't ideal and causes thread stalls. Can't see any other reason except thermal throttling which shouldn't be happening in benchmarks with proper cooling (and would be reported on anyway if the article is competent!). However, that's only what little I've found with Googling. The general impact of SMT on games may be a ways different from that. And no benchmark of dual-threads is going to give any idea what 3 or 4 threads could achieve. Conceptually, what if Zen 3 has an extra SIMD unit per core for running maths tasks? May be possible to keep that populated and active as a third thread. :-?
 
We already have real-world examples where HT slows down a game. That's not to say it's useless; only to keep in check any possible ideas of more threads == more performance. It means a different utilisation profile of the CPU and something devs will work (or not) with. My gut feeling is perhaps a 10% basic boost to general workloads with the CPU better able to schedule work and keep the CPU active. But of course, the more active it is, the hotter it runs, so the lower it needs to be clocked per core.

If there isn't already a discussion in the PC architecture forum, it's worth having one there as more knowledgable engineering folk may be able to weigh in.
This is why I think the cost to benefit is lower than using zen 2 if its for Scarlett console apu only and not using the apu in cloud
 
Is there a lot of workload in a game engine that is so highly parallel that cpu threading is a big advantage, but not so highly parallel that it's better to run it on the GPU?
 
Is there a lot of workload in a game engine that is so highly parallel that cpu threading is a big advantage, but not so highly parallel that it's better to run it on the GPU?
I suspect branching is a big piece to that question.
The reason we don't do more GPU based processing is because most GPUs cannot submit its own work back to itself to do. Ie - a shader cannot call a shader.
The better we get at that particular feature (both API and hardware), we could see more GPU based processing. Otherwise it's a large pain for devs to keep 2 paths/completely separate code bases.

The furthest ahead right now in that space are Kepler and forward GPUs. These require special extensions on vulkan or i imagine game works features/library exploit this feature.

And Xbox Ones and above; this required hardware customization and a xbox variant of DX12. Currently no known game leverages this yet - at least confirmed to do it. I would have suspected that 1P titles would exploit it, but since they are all going PC, this is doubtful as this would now require 2 codebases.
 
Status
Not open for further replies.
Back
Top