Next Generation Hardware Speculation with a Technical Spin [pre E3 2019]

[lefantome]

Targeting full 4K is not a matter only of GPU, you have to bump up HD (all is bigger), HD speed if you don't wait 15 min for each loading, RAM quantity, Bandwidth... Net speed. How much cost today a full 4K capable PC ?

Xbox One X does this well and you don't need a huge bump on the other aspects, you should compare consoles with consoles not PCs.

Yes a even better HDD and better and more ram are necessary but not that much, at least not because of 4K.

GPU power to have at the same time great graphics and native 4K is needed.
The biggest ? this gen imho is real time lighting.

Are we going to see raytracing? Probably not.
So what else? huge lightmaps again or something like those voxels technologies shown years ago?

I want next gen to be reallt next so they have to have powerful GPUs for whatever thing they want to do otherwise this gen will be a gen of transition before the next one with full 4K ray tracing that will be probably the last one.
After that we will see small improvements that don't justify a "next-gen" label, at least on the graphical side.

 

[anexanhume]

I felt like I seen that slide pre-Polaris launch.

It originated from March 2016 according to Tineye reverse image search.

So vega is per/wat so much better than polaris as on this picture?

At the time they didn’t expect to ship it with broken features

 

[BRiT]

Calculating the growth for 4 years using the growth from PS4 to PS4Pro (2.28) the result is 12.5TF

But the rumors are only 3 years of growth, because it was supposed to be 2019, so that would be 9.58 TF.

 

[mpg1]

It originated from March 2016 according to Tineye reverse image search.

right.. so before Polaris launch..

 

[function]

Whether it's 8, 10 or 12 TF, that's not terribly exciting. It's okay, but it's really just a notch on the slider or small percentage increase in the P.

60 hz, ray acceleration, blistering int8 performance for AI, fast CPU for massively populated worlds and scaled up sandboxes ... that's something to get excited about.

IMO.

 

[anexanhume]

right.. so before Polaris launch..

Yes, so someone would have to dig deeper to see if their claims bore out. The way that they’ve backtracked on perf/Watt gains in 7nm would make it unsurprising if Navi wasn’t as big a jump as they hoped.

 

[mpg1]

Yes, so someone would have to dig deeper to see if their claims bore out. The way that they’ve backtracked on perf/Watt gains in 7nm would make it unsurprising if Navi wasn’t as big a jump as they hoped.

IIRC...those claims did not bare out..

 

[3dilettante]

It'll be curious to see, but I imagine it would be an RT unit (whatever that may be ????) per CU or a group of CUs like how nV has an RT per SM - so maybe it ends up as one RT block per group of CUs, which in Polaris is up to 3 CUs with shared L1 as opposed to the max of 4 in the rest of GCN (including Fiji/Vega I assume).

Wonder if they'd be inclined to go down to 2 CUs per L1 grouping (or @3dilettante can bust my hypotheticals :V )

The RT unit in Nvidia's architecture seems like it would work best if matched with the memory pipeline, and there are some indicators that it does a fair amount of work with the buffers and caches. Traversal can generate multiple memory accesses, so that portion of the workload could stress the memory hierarchy readily. If similar thinking were applied to AMD, I think it would best fit by linking RT hardware to the vector memory path in the CUs. The scalar and instruction caches have varying amounts of sharing between CUs, but this path has relatively low bandwidth, weaker memory consistency, and limited latency tolerance compared to the per-CU vector memory units.

Oh. So the damn Vega has additional x16 PCIe lanes (and likely an onboard NVMe controller)... it makes more sense to me now.

I'm not sure about extra x16 PCIe links. There are Vega 10 die shots that don't show a significant amount of additional IO.

This roadmap is post Polaris launch, so unless AMD misrepresented or used an obscure metric, the answer is yes.

If going by the position of Polaris in mid-2016, the 28nm data point is before 2015 and may be the late 2014 launch of Tonga. That one in particular would serve as a lower bar to clear than other 28nm GPUs, though if I recall Polaris at launch didn't necessarily clear that either.

 

[anexanhume]

IIRC...those claims did not bare out..

FP32 went from 2TF to 4.7TF for a 150W card. Seems pretty close to me.

 

[MrFox]

But the rumors are only 3 years of growth, because it was supposed to be 2019, so that would be 9.58 TF.

That rumor claimed it was delayed because the parts were too expensive for Q4 2019. Implying the specs were too optimistic and they decided to delay to reach the price point instead of lowering specs/clocks/memory.

I still don't believe that guy, too many red flags.

 

[Shifty Geezer]

By the time consoles release RT would have been out in the wild for 2...

But only a small corner of it at the moment. If RT was mainstream in every card and every engine, definitely, but as it's not, it's still a jump. A jump probably worth doing, but I don't think "it's been out for two years already" argument is the valid justification.

 

[lefantome]

But the rumors are only 3 years of growth, because it was supposed to be 2019, so that would be 9.58 TF.

fake rumors? who cares.

 

[Shifty Geezer]

...could this be possible?

Yes.

 

[Kaotik]

So vega is per/wat so much better than polaris as on this picture?

Just because Vega 56 and Vega 64 don't offer better perf/watt than Polaris 10/20 doesn't mean Vega couldn't offer much better perf/watt than Polaris (see MacBook Pro Vega vs Polaris for example)

 

[iroboto]

But only a small corner of it at the moment. If RT was mainstream in every card and every engine, definitely, but as it's not, it's still a jump. A jump probably worth doing, but I don't think "it's been out for two years already" argument is the valid justification.

Fair enough.
A fully supported API feature set from MS designed in conjunction with multiple vendors on both DirectML and DXR which imo, has a lot more weight than just a sectioned cornered customization like most Nvidia game workshop items are.
Let's be real here, it's not like AMD, Intel, and other chip manufacturers had zero say in the development of the APIs. These large scale items, like the introduction of DX12 take time for agreement and consensus.
On an aside, AMD has indicated they are working on RT. Which means both MS and Sony should have access to it. They have to make a choice of whether they want to support it. I think that path is much more predictable for MS.

 

[wco81]

Whether it's 8, 10 or 12 TF, that's not terribly exciting. It's okay, but it's really just a notch on the slider or small percentage increase in the P.

60 hz, ray acceleration, blistering int8 performance for AI, fast CPU for massively populated worlds and scaled up sandboxes ... that's something to get excited about.

IMO.

Well if it’s not a jump in power, what then?

MS tried something different with Kinect 2 and suffered.

OTOH, Nintendo didn’t try to pursue performance for the sake of performance and was rewarded for the Wii and the Switch, but not so much for the one in between.

 

[mrcorbo]

Well if it’s not a jump in power, what then?

MS tried something different with Kinect 2 and suffered.

OTOH, Nintendo didn’t try to pursue performance for the sake of performance and was rewarded for the Wii and the Switch, but not so much for the one in between.

Delivering a higher TF rating isn't the only way to increase performance. People are getting way too hung up on that single metric.

 

[Metal_Spirit]

Delivering a higher TF rating isn't the only way to increase performance. People are getting way too hung up on that single metric.

No... It is not. You are correct.
The problem here it is not the people that know that. It's the global less informed market that having heard about that metric as a separator in this generation, will use it on the next one, even in a case where other methods to increase performance might be available.
Since both systems will be AMD, people, right or wrong, will use Tflops as a metric. That's a given! So Tflops, at least for a big, big, group of people, will be an effective marketing weapon.

 

[vipa899]

Sony could just use some 'bits' again, like the PS2's 128bit processing powers. This is an area Sony is much better at then MS.

 

[mrcorbo]

No... It is not. You are correct.
The problem here it is not the people that know that. It's the global less informed market that having heard about that metric as a separator in this generation, will use it on the next one, even in a case where other methods to increase performance might be available.
Since both systems will be AMD, people, right or wrong, will use Tflops as a metric. That's a given! So Tflops, at least for a big, big, group of people, will be an effective marketing weapon.

That might be OK, if there aren't major, performance-impacting differences between the hardware. I'm more bothered by people comparing Polaris and Navi given we don't know what Navi's performance is relative to this metric.

TFlops can be a useful metric for comparing performance when you can view it in context. We don't have the proper context for Navi or the consoles based on it, yet.