Next Generation Hardware Speculation with a Technical Spin [2018]

[anexanhume]

That would put a Titan V at how many 'amd tflops'?

I wouldn’t try and equate them, particularly since Titan is not a gaming card. Compare Vega to 1080 gaming performance and scale from there if you have to, I suppose. But comparisons across architectures, development environments and optimization levels will be hard to be precise. Me saying a 2020 console can equal that is educated guessing.

 

[function]

AMD TF =/= NVIDIA TLFOPS. Generally AMD has more TF/Watt but less perf/Watt.

That's on PC, under DX. With Vulcan or on a console utilising async compute heavily, I'm not sure things would be so heavily skewed towards Nvidia.

And console specific optimisation may make a difference, even at the hardware level. MS made around 80 hardware optimisation to X1X based on their profiling. It has no problem outperforming a 580 or even a 1060 at only 180W power draw for the entire system at the wall.

AMD suffer from a number of none hardware factors on PC, in addition to hardware factors (late, bugged, process etc).

 

[Megadrive1988]

This is true, the mid gen refreshes aren't doing all that poorly imo. So 8TF target seems like a mass market target imo. Bring the costs down and the yield up.

8TF GPUs for PS5/Scarlet isn't going to cut it IMO. 8TF is the estimate Cerny gave for doing PS4 games at native 4K,

Next Gen consoles will need around 12TF to seem like a generational upgrade over base PS4.

 

[vipa899]

particularly since Titan is not a gaming card

Gaming card or not, it performs better then a 1080Ti in gaming. Guess the new Volta gpu's will be Titan V level, GTX1180 or something.

It has no problem outperforming a 580 or even a 1060 at only 180W power draw for the entire system at the wall.

Thats indeed very impressive, even if the X wouldnt outperform a 1060.

 

[Shifty Geezer]

That's on PC, under DX. With Vulcan or on a console utilising async compute heavily, I'm not sure things would be so heavily skewed towards Nvidia.

And console specific optimisation may make a difference, even at the hardware level. MS made around 80 hardware optimisation to X1X based on their profiling. It has no problem outperforming a 580 or even a 1060 at only 180W power draw for the entire system at the wall.

AMD suffer from a number of none hardware factors on PC, in addition to hardware factors (late, bugged, process etc).

Yes indeed. Flops is a peak measure of maths. In an ideal environment, with code used to extract maximal performance from the GPU architecture, one Flop is one Flop. Where you have differences in performance is where the hardware isn't be maximised and flops are going unused, waiting for work to do or processing data that isn't ideally suited to the architecture. In an ideal console, you'd hope the GPU will be more ideally targeted resulting in higher utilisation and those Flops all being used. eg. You could have an architecture with twice as many flops achieve half the work of another in PC because the scheduling and overheads are all out of whack, whereas the same component in a console could run twice as fast because the workloads are tailored for it.

 

[vipa899]

the same component in a console could run twice as fast because the workloads are tailored for it.

Yeah in a ideal world a 970 in a console could perform as fast a as a 1080Ti.

 

[Wynix]

My money is on $449 launch price, Q4/2020 and then price drop to $399 announced at E3/2021.

 

[dobwal]

I still feel launching a year earlier trumps any feature a competitor could implement by launching a year later. I still don't buy the end of 2019 7nm chips being much more exspensive than 2020. 2020 will probably be able to have a 10-15 percent faster clock speed.

TMSC 7nm+ will increase transistor density by 20% and lower power consumption by 10% compared to its 1st gen 7nm process.

 

[mpg1]

TMSC 7nm+ will increase transistor density by 20% and lower power consumption by 10% compared to its 1st gen 7nm process.

I'd take the year advantage honestly.

 

[metacore]

8TF GPUs for PS5/Scarlet isn't going to cut it IMO. 8TF is the estimate Cerny gave for doing PS4 games at native 4K,

Next Gen consoles will need around 12TF to seem like a generational upgrade over base PS4.

Unofrunately it looks like 12TFflops won't be enough for "generation jump" in graphics/shading at native 4k, more like twice of that even at 30hz. Some nice posts about the matter by Timothy Lottes.

https://timothylottes.sourceforge.io/20160909.html
https://timothylottes.sourceforge.io/20170516.html

"...The difference between PS4 and Xbox360 is roughly a 3.4x bump in Kflop/pix..."


Another thing is memory bandwidth glass celling which is probably 500-600GB/s for 2019/20 console using 256bit gddr6. This memory bandwidth reality also shoots down hopes for any much faster gpu because that would be paper flops without ample increase in bytes. Temporal reconstruction plz send help ;/

 

[anexanhume]

Unofrunately it looks like 12TFflops won't be enough for "generation jump" in graphics/shading at native 4k, more like twice of that even at 30hz. Some nice posts about the matter by Timothy Lottes.

https://timothylottes.sourceforge.io/20160909.html
https://timothylottes.sourceforge.io/20170516.html

"...The difference between PS4 and Xbox360 is roughly a 3.4x bump in Kflop/pix..."


Another thing is memory bandwidth glass celling which is probably 500-600GB/s for 2019/20 console using 256bit gddr6. This memory bandwidth reality also shoots down hopes for any much faster gpu because that would be paper flops without ample increase in bytes. Temporal reconstruction plz send help ;/

Do you think 384-bit GDDR6 is that out of the question? I’d say MS opened that door when they did it with XB1X.

 

[iroboto]

Yes indeed. Flops is a peak measure of maths. In an ideal environment, with code used to extract maximal performance from the GPU architecture, one Flop is one Flop. Where you have differences in performance is where the hardware isn't be maximised and flops are going unused, waiting for work to do or processing data that isn't ideally suited to the architecture. In an ideal console, you'd hope the GPU will be more ideally targeted resulting in higher utilisation and those Flops all being used. eg. You could have an architecture with twice as many flops achieve half the work of another in PC because the scheduling and overheads are all out of whack, whereas the same component in a console could run twice as fast because the workloads are tailored for it.

https://mynameismjp.wordpress.com/2...-barriers-part-3-multiple-command-processors/

Welcome to Part 3 of the series! In this article, I’m going to talk a bit about how multiple command processors can be used to increase the overall performance of a GPU by reducing the amount of time that shader cores sit idle.

Indeed.

In Part 4, we’ll discuss how having multiple command processors can be useful for scenarios where multiple applications are sharing the same GPU.

can't wait honestly

 

[Megadrive1988]

Unofrunately it looks like 12TFflops won't be enough for "generation jump" in graphics/shading at native 4k, more like twice of that even at 30hz. Some nice posts about the matter by Timothy Lottes.

https://timothylottes.sourceforge.io/20160909.html
https://timothylottes.sourceforge.io/20170516.html

"...The difference between PS4 and Xbox360 is roughly a 3.4x bump in Kflop/pix..."

Sounds similar to what DF recently posted in their article about the possibilities for more than one next gen Xbox.

https://www.eurogamer.net/articles/...ory-will-there-be-more-than-one-next-gen-xbox

The issue is that while AMD can offer a generational leap over the graphics performance of the base PlayStation 4 (effectively the main target current-gen target platform), delivering the same increase in performance over PS4 Pro and especially Xbox One X is far more challenging. Doubling X's GPU power is achievable, but if the aim is also to double frame-rate, the ability to deliver richer graphics is therefore more limited. If the objective is to lock to native 4K and deliver hugely improved visuals, again, 2x Xbox One X performance isn't enough.

My best guess? The so-called 'FauxK' upscaling techniques seen today will be back, refined and improved for next-gen - be it through developers' ingenuity or via dedicated hardware. A 6x to 8x leap in GPU performance over the base PlayStation 4 can be delivered, but servicing a 4x leap in resolution doesn't leave a huge amount of remaining overhead for radically improved visuals. The jump from OG Xbox to Xbox 360 required a bleeding-edge GPU to deliver a 3x increase in pixel density with enough headroom for much improved graphics. Meanwhile, the leap from PS3 to PS4 saw only a 2.25x increase in pixel-count. In this respect, expecting native 4K across the board from next-gen doesn't seem likely.

More probable is innovative use of custom hardware. There are already rumours of Sony collaborating directly with AMD on its Navi architecture, and PS4 Pro offered some fascinating technology in allowing a console GPU to punch well above its weight in supporting 4K screens

Another thing is memory bandwidth glass celling which is probably 500-600GB/s for 2019/20 console using 256bit gddr6. This memory bandwidth reality also shoots down hopes for any much faster gpu because that would be paper flops without ample increase in bytes. Temporal reconstruction plz send help ;/

Do you think 384-bit GDDR6 is that out of the question? I’d say MS opened that door when they did it with XB1X.

I think both Scarlet and PS5 need to go with 384-bit GDDR6 (24GB) with 800+ GB/s bandwidth, if they're not going with any HBM2/HBM3.

I agree MS opened that door with XB1X.

 

[Xbat]

TMSC 7nm+ will increase transistor density by 20% and lower power consumption by 10% compared to its 1st gen 7nm process.

Will 7nm+ be ready by the end of 2020? I think 7nm is a lock for the first machines.

Regarding native 4k there's no point, they definitely going to be using temporal reconstruction upscaling. Going native is a massive waste of resources that will have a much bigger impact being used elsewhere.

 

[dobwal]

Will 7nm+ be ready by the end of 2020? I think 7nm is a lock for the first machines.

Regarding native 4k there's no point, they definitely going to be using temporal reconstruction upscaling. Going native is a massive waste of resources that will have a much bigger impact being used elsewhere.

7nm+ is slated to ramp up in 2019. Zen3 and AMD’s next gen gpu are slated for 7nm+ and 2020 releases.

 

[ultragpu]

Unofrunately it looks like 12TFflops won't be enough for "generation jump" in graphics/shading at native 4k, more like twice of that even at 30hz. Some nice posts about the matter by Timothy Lottes.

https://timothylottes.sourceforge.io/20160909.html
https://timothylottes.sourceforge.io/20170516.html

"...The difference between PS4 and Xbox360 is roughly a 3.4x bump in Kflop/pix..."


Another thing is memory bandwidth glass celling which is probably 500-600GB/s for 2019/20 console using 256bit gddr6. This memory bandwidth reality also shoots down hopes for any much faster gpu because that would be paper flops without ample increase in bytes. Temporal reconstruction plz send help ;/

Yeah very good of Tim to put that in math terms, finally somebody spent some effort to bring us closer to the harsh reality. More reason we need to mainstream Resolution, Visual and Framerate mode come next gen. Personally 4k cbr 30 fps is the best solution for a tangible next gen experience with how things are looking right now. No point go native 4k or 60fps if your game doesn't deliver your vision because of those constraints.

 

[anexanhume]

7nm+ is slated to ramp up in 2019. Zen3 and AMD’s next gen gpu are slated for 7nm+ and 2020 releases.

The incompatibility of 7nm and 7nm+ really harms the console makers’ flexibility here. Presumably you’d want to make the processor on the best process available provided the cost and yield are amenable to a console budget and supply curve through the first year, but you most likely have to bifurcate your efforts if you want to be able to react to the foundry realities. Does this mean you make a Zen 2, Navi design on 7nm and 7nm+? Do you backport Zen 3 to 7nm in case 7nm+ falls through? I think it’s fairly clear Navi is the architecture of choice here, but the CPU doesn’t seem a certainty yet.

 

[metacore]

Do you think 384-bit GDDR6 is that out of the question? I’d say MS opened that door when they did it with XB1X.

Out of question? no but this 384bit wide interface on 300-400mm2 chip will create problems for effective die shrinks and 12 ram chips will always be need. MS might done it because they knew the scorpio engine will not undergo die shrink. It remains to be seen if that door is really open. It depends how much they are willling to complicate life for themselves later on. They may as well use 384bit bus and hope there will be solution two years later to replace it for something cheaper.

Then we have capacity. All those objects spinning with zen cores and drawn using primitive shaders will be memory hungry. Its forgone conclusion we wont see 8-16 bump in memory capacty but will 24 be enough? or use 32GB using 256bit clamshell with clear path to cost reduction. I don't envy architects trying to reach next gen dream with this 4k ball and chain, at cost and path to cost reductions in this day and age. Maybe console family at different pricepoint is the answer. BTW nice tech summary on resetera, Tim posts would help temper the expectations (i don't have account).

 

[DuckThor Evil]

I think a box with 2x Xbox One X with stronger CPU would be a great thing to have, especially if that is the base line developers were targeting. That is a nice boost over the current base models and I'm sure devs would put that to good use. The power levels are already reaching the point where software, budget and developer skills are the main determining factor on visuals.

 

[Xbat]

I think 16GB of gddr6 is what we getting