AMD Polaris architecture – GCN 4.0

[Seanspeed]

Amd made a cracking card called the 290 that offers similar performance and released a year before. Sure it will cost a few more pounds per month due to using more energy but this is PC gaming. My card is stock 290x performance and was bought at £200 with 4 games. These cards were a steal and still people go on about the gtx970. It's pure and utter madness.

To be fair, while 290X's that were selling for £200 were a steal, and still are(if you can find them), this was not some normal, official price drop from AMD or anything. It required people to get in on fantastic retailer discounts while supplies lasted, or as often was the case - buy used.

These cards were also often reference cards which were notorious for their painful loudness and poor cooling and had been put through torture mining bitcoins. I'm sure you're going to counter about how you got some great card, but dont think that being an exception proves the rule.

It's a very disingenuous argument in the end. I would say that right now, the 390 is a better grab than a 970, but to say that the hype over the 970, which came out nearly 2 years ago, was unjustified, is just revisionist nonsense.

 

[TheRealDeal]

To be fair, while 290X's that were selling for £200 were a steal, and still are(if you can find them), this was not some normal, official price drop from AMD or anything. It required people to get in on fantastic retailer discounts while supplies lasted. Check out the other tiers as well. Nvidia only have a win at the higher end.

These cards were also often reference cards which were notorious for their painful loudness and poor cooling and had been put through torture mining bitcoins. I'm sure you're going to counter about how you got some great card, but dont think that being an exception proves the rule.

It's a very disingenuous argument in the end. I would say that right now, the 390 is a better grab than a 970, but to say that the hype over the 970, which came out nearly 2 years ago, was unjustified, is just revisionist nonsense.

It's only unjustified if you thought a high end design would not overcome a plainly lower end design. It's been happening since the 7970/gtx680, 290x/780ti, 390x v gtx980 is starting to go the same way. To me it's quite obvious.

 

[Seanspeed]

It's only unjustified if you thought a high end design would not overcome a plainly lower end design.

They are very much in the same ballpark and have their own pros/cons. Trying to say that one is 'high end' and one is 'low end' says far more about your biases than any conjecture I could come up with. Thanks for saving me the trouble.

 

[D.P.]

By all means explain. From my point I have seen plenty of things done through hardware and software. You seem to have the knowledge though so give us the details needed to understand. Calling people hilarious without an explanation is hilarious in itself. Is it a full hardware implementation like GCN.

You are the one claiming it is done in software despite Nvidia and developers describing how it is done in hardware. So I assume you have some kind of proof of your assertions?

 

[TheRealDeal]

They are very much in the same ballpark and have their own pros/cons. Trying to say that one is 'high end' and one is 'low end' says far more about your biases than any conjecture I could come up with. Thanks for saving me the trouble.

Read the specs and come back.

 

[TheRealDeal]

You are the one claiming it is done in software despite Nvidia and developers describing how it is done in hardware. So I assume you have some kind of proof of your assertions?

My thoughts are it's much like Maxwell if not the same. I think all they have done is improve it through software. My overall thoughts on Pascal is its pretty much Maxwell shrunk. Again I am not an expert it's just my overall impression.

 

[Seanspeed]

Read the specs and come back.

I read the specs a long time ago. :/

Have they changed and I didn't realize?

 

[TheRealDeal]

I read the specs a long time ago. :/

Have they changed and I didn't realize?

Which card strikes you as a higher spec. I know these things are not always black and white but they are all 28nm so it's a little easier. The 290/x keeps getting stronger while what I consider weaker designs from Nvidia are falling back.

 

[D.P.]

My thoughts are it's much like Maxwell if not the same. I think all they have done is improve it through software. My overall thoughts on Pascal is its pretty much Maxwell shrunk. Again I am not an expert it's just my overall impression.

Well' you got the wrong impression. The facts are out there if you want to learn, I posted a very detailed link detailing async on AMD and nvidia hardware.


There are big differences between AMD's and Nvdia's DX12multi-engine support, pros and cons between them. AMD definitely have a superior solution overall but to say things like pascal is just maxwell, async is done in software on pascal etc. is just plain wrong.

 

[Seanspeed]

Which card strikes you as a higher spec. I know these things are not always black and white but they are all 28nm so it's a little easier. The 290/x keeps getting stronger while what I consider weaker designs from Nvidia are falling back.

I have no problem admitting that I think the Hawaii architecture was a bit more forward-thinking in terms of where low level API optimizations were heading.

But there was also no denying that whatever memory bandwidth and ACE advantages it had, actual performance results were entirely mixed. And still are. Hawaii has only recently started to edge ahead, which is hardly a major victory against a card that came out nearly 2 years ago. Go back to when the 970 released and for a good while, it was ahead of the 290/290X overall fairly comprehensively.

So for you to classify Hawaii as 'high end' and GP204 as 'low end' is not only a gross exaggeration, it is complete disinformation. Untrue and intended to try and push a reality that didn't exist.

 

[TheRealDeal]

Well' you got the wrong impression. The facts are out there if you want to learn, I posted a very detailed link detailing async on AMD and nvidia hardware.


There are big differences between AMD's and Nvdia's DX12multi-engine support, pros and cons between them. AMD definitely have a superior solution overall but to say things like pascal is just maxwell, async is done in software on pascal etc. is just plain wrong.

Just post up the evidence then. I am willing to learn. All I want to know is what's different between Maxwell and Pascal where it matters. I see a huge clock boost which seems to account for the performance increase between what's on offer.

 

[TheRealDeal]

I have no problem admitting that I think the Hawaii architecture was a bit more forward-thinking in terms of where low level API optimizations were heading.

But there was also no denying that whatever memory bandwidth and ACE advantages it had, actual performance results were entirely mixed. And still are. Hawaii has only recently started to edge ahead, which is hardly a major victory against a card that came out nearly 2 years ago. Go back to when the 970 released and for a good while, it was ahead of the 290/290X overall fairly comprehensively.

So for you to classify Hawaii as 'high end' and GP204 as 'low end' is not only a gross exaggeration, it is complete disinformation. Untrue and intended to try and push a reality that didn't exist.

It's not low and high. I said higher spec. Any how you basically agreed with me in your first point so in this threads interest let's agree to disagree. Close enough lol.

 

[D.P.]

Just post up the evidence then. I am willing to learn. All I want to know is what's different between Maxwell and Pascal where it matters. I see a huge clock boost which seems to account for the performance increase between what's on offer.

https://www.pcper.com/reviews/Graph...ition-Review-GP104-Brings-Pascal-Gamers/Async

 

[Seanspeed]

It's not low and high. I said higher spec.

No, you specifically said that the 290 was a "high end" design and the 970 a "lower end" design.

Feel free to backtrack if you need to, though.

 

[drunkenmaster]

DP, Pre-emption was ALWAYS possible, it's never not been possible, the point behind async computer is to allow compute and graphics loads to flow through in the most efficient way possible. Pre-emption is NOT the most efficient way, allowing each SM to share the power between different loads is more about what async is. The ability to stop something processing and process something else instead isn't the goal behind async. Async is supposed to be the better alternative to pre-emption.


http://i.imgur.com/GgSZy0K.png

http://i.imgur.com/0ghidyN.png

EDIT:- those images are way too big so I'll just link them.

Look at the latter slide in particular, still supports pre-emption.... async is NOT pre-emption, async is the thing that dramatically improves upon pre-emption which has always been done. Yes reducing the overhead of pre-emption helps, but not much. The main point really is an entire architecture where the hardware scheduler and the shaders(in whatever number of groups they are in) can work together to split up a list of both graphics and compute commands together in the single most efficient way possible without constantly pre-empting workloads.

That Nvidia is focusing on reducing the overhead of pre-emption directly speaks to their lack of true async compute support.

 

[Rroff]

DP, Pre-emption was ALWAYS possible, it's never not been possible, the point behind async computer is to allow compute and graphics loads to flow through in the most efficient way possible. Pre-emption is NOT the most efficient way, allowing each SM to share the power between different loads is more about what async is. The ability to stop something processing and process something else instead isn't the goal behind async. Async is supposed to be the better alternative to pre-emption.


http://i.imgur.com/GgSZy0K.png

http://i.imgur.com/0ghidyN.png

EDIT:- those images are way too big so I'll just link them.

Look at the latter slide in particular, still supports pre-emption.... async is NOT pre-emption, async is the thing that dramatically improves upon pre-emption which has always been done. Yes reducing the overhead of pre-emption helps, but not much. The main point really is an entire architecture where the hardware scheduler and the shaders(in whatever number of groups they are in) can work together to split up a list of both graphics and compute commands together in the single most efficient way possible without constantly pre-empting workloads.

That Nvidia is focusing on reducing the overhead of pre-emption directly speaks to their lack of true async compute support.

The dynamic load balancing image from DP's link pretty much says it all when it comes to Pascal - quite limited in terms of concurrent compute capabilities with nothing like the broadness of AMD's approach - but Pascal better able to bring GPU resources to the task at hand when the graphics queue has cleared compared to Maxwell.

Judging from the apparent changes AMD are making with Polaris compared to Fiji, etc. it looks like they've realised that real world optimal use of async lies somewhere between the 2 different approaches employed by nVidia and AMD.

 

[AlamoX]

 

[flopper]

Well' you got the wrong impression. The facts are out there if you want to learn, I posted a very detailed link detailing async on AMD and nvidia hardware.


There are big differences between AMD's and Nvdia's DX12multi-engine support, pros and cons between them. AMD definitely have a superior solution overall but to say things like pascal is just maxwell, async is done in software on pascal etc. is just plain wrong.

Async is disabled on ashes game test on all Nvidia hardware including the 1080 atm. If you find a 1080 to send them they can test it and find out if they can enable async for nvidia.

 

[Mtom]

As i understand, the only thing Pascal does better than Maxwell in terms of async, is that it can cancel tasks to swiych between compute or graphics workload, and not get stuck in one of those until it is done.
In my book its still not real asynchronous workload, just fadter and more flexible shifting between the two.

 

[andybird123]

Async is disabled on ashes game test on all Nvidia hardware including the 1080 atm. If you find a 1080 to send them they can test it and find out if they can enable async for nvidia.

Nope, there are already reviews up showing async on on the 1080 giving an improvement over async off.

But dont let that stop your fud campaign