AMD Polaris architecture – GCN 4.0

Mauller · 24 Mar 2016 at 17:43

Mtom said:
Well either the specs are bullcr*p, or AMD has some magic in the new architecture if a card with those specs ran Hitman 1440p 60FPS.

Apparently that is the Pro card though, also it must have been some engineering test sample as the sisoft clock for it was low.

muziqaz · 24 Mar 2016 at 17:44

That gcn core number seems nonsense. 256bits are correct for Polaris 10. Polaris 11 is 128bit

Rroff · 24 Mar 2016 at 17:59

muziqaz said:
Please tell me who was so smart to say that?

Before Fiji was launched there was a certain amount of HBM hype and GDDR5 is dead type comments so I'd find it mildly amusing if GDDR5 lived on for another generation yet.

Mtom · 24 Mar 2016 at 17:59

Mauller said:
Apparently that is the Pro card though, also it must have been some engineering test sample as the sisoft clock for it was low.

My statement still stands. If thats the Pro, then the full chip has like 3-400SP more...still would be strange to be as fast as Polaris 10 was hinted.
If it is really that fast with this low specs then the new GCN is tge real thing.

SimBy · 24 Mar 2016 at 18:01

Rroff said:
Before Fiji was launched there was a certain amount of HBM hype and GDDR5 is dead type comments so I'd find it mildly amusing if GDDR5 lived on for another generation yet.

It most likely is dead for enthusiast segment. I mean both Nvidia and AMD kinda agree on that.

Rroff · 24 Mar 2016 at 18:05

Oh I agree its days are numbered but some people were talking like HBM would immediately blow it out the water which hasn't been the case and looking like won't be the case for awhile longer yet.

Mtom said:
My statement still stands. If thats the Pro, then the full chip has like 3-400SP more...still would be strange to be as fast as Polaris 10 was hinted.
If it is really that fast with this low secs then the new GCN is tge real thing.

I think Fiji had to make some compromises to get a balance of DX11 and older and DX12 performance on 28nm that won't be the case on sub 20nm planar and may make quite a significant difference in some cases.

Mauller · 24 Mar 2016 at 18:07

Mtom said:
My statement still stands. If thats the Pro, then the full chip has like 3-400SP more...still would be strange to be as fast as Polaris 10 was hinted.
If it is really that fast with this low secs then the new GCN is tge real thing.

Well look at Fiji's theoretical performance compared to how it performs. if they have fixed shader utilisation bottlenecks and improved each shaders performance then it is perfectly feasible.

D.P. · 24 Mar 2016 at 18:18

Rroff said:
Oh I agree its days are numbered but some people were talking like HBM would immediately blow it out the water which hasn't been the case and looking like won't be the case for awhile longer yet.

I think Fiji had to make some compromises to get a balance of DX11 and older and DX12 performance on 28nm that won't be the case on sub 20nm planar and may make quite a significant difference in some cases.

From recent rumours, HBM also might have a short life.

Mauller · 24 Mar 2016 at 18:24

D.P. said:
From recent rumours, HBM also might have a short life.

Which rumours? HBM 2 will supplant it, then the AMD chart just showed Nextgen memory for Navi, that could mean anything such as HBM3 etc.

Kaapstad · 24 Mar 2016 at 18:54

muziqaz said:
Please tell me who was so smart to say that?

Never been a fan of HBM1

muziqaz · 24 Mar 2016 at 19:01

Kaapstad said:
Never been a fan of HBM1

You never been a fan of everything which starts with 4gb

Either way it is stupid to think that either hbm or gddr5 take over or die. Don't think anyone ever claimed that if they did they never meant it that way.

Orangey · 24 Mar 2016 at 20:01

Anyone care to speculate on "next-gen memory"? Asked before and got no answers.

Something fast as opposed to just wide like HBM?

Gerard · 24 Mar 2016 at 20:07

Kaapstad said:
Never been a fan of HBM1

You didn't have a crystal ball either so lets not go down the route of rewriting history

drunkenmaster · 24 Mar 2016 at 20:16

I would say memory with significantly more complex logic. HBM is essentially the tick, where the interface was the new bit and they went with relatively simple memory/logic design. The tock step is to use the same interface but focus on a significantly more complex chip, lower power usage, lower latency, better/more parallel access, increased density. Will it be called HBM3 or a completely new memory type, doesn't really matter. The next memory change will be more like gddr3 to gddr5 most likely. The interesting parts of HBM are less the chips and more how it's all put together. TSVs, die stacking, interposers, very wide bus. the next gen memory will also be tsvs, die stacking, interposer and very wide bus, just a different type of chip stacked together.

Rroff · 24 Mar 2016 at 20:39

Orangey said:
Anyone care to speculate on "next-gen memory"? Asked before and got no answers.

Something fast as opposed to just wide like HBM?

This I suspect:

drunkenmaster said:
lower latency, better/more parallel access

Certain aspect seem to be bottlenecked by queue latency and/or lacking parallel access.

jimhaumman · 24 Mar 2016 at 20:46

Orangey said:
Anyone care to speculate on "next-gen memory"? Asked before and got no answers.

Something fast as opposed to just wide like HBM?

Maybe GDDR5X in some gpus.
Micron begins sampling GDDR5X memory chips.
http://www.overclock3d.net/articles/gpu_displays/micron_begins_sampling_gddr5x_memory_chips/1

CAT-THE-FIFTH · 24 Mar 2016 at 21:08

D.P. said:
Back to school in the US is August to early September, June is still at school in the previous school year!I know, because I happen to live there.

First Day of School last year was September 9th, last day is June 28th. Back to school is the last couple of weeks in august first couple in September.

Dave2150 said:
Nice one, July is indeed within the US back to school period.

Agreed,Dave I like how D.P. is now saying a whole load of US sales and market research sites are now wrong. He first said I was wrong about July being back to school in marketing terms and once he was proven wrong tried to say I have "issues" and "junk",ie, trying silling name calling which sadly is quite common on US forums.

He desperately tries to deflect that he was proven wrong - one third of back to school sales happen in July with nearly 2/3 of buying decisions for August made in July. If AMD does not even have a soft release with reviews in July they have essentially missed the back to school period.

Google and Ipsos research said:
Parents research and make back-to-school purchases at different times – July is all but the unofficial month for researching back-to-school products. A recent study by Google and Ipsos Research concluded that 65% of U.S. internet users started their back-to-school research in the month of July. Many of these researchers, however, waited until August to make purchases – 48% of U.S. internet users made back-to-school purchases in August, only 35% made purchases in July.

Google and Ipsos MORI are not amateurs in the market research field - July and August make up 83% of back to school sales according to them. The middle or end of August would mean AMD has missed most of the critical sales period and also missed most of the research period too.

If AMD has no reviews out in July,they might as well admit Polaris has been delayed and if Nvidia get's Pascal out earlier,then AMD is in BIG trouble IMHO.

Also,D.P., don't bother replying as I will make sure I will continue repeating post 2061 til it is blue in your face since you really have issues yourself(using your own language),and seem to know better than many US sites which I quoted. Rage at them if you feel they are giving faulty data. They are US sites after all.

We will need to agree to disagree and leave it at that. If you want the last say it is your perogative,but I will ignore it.

Mauller · 24 Mar 2016 at 21:12

drunkenmaster said:
I would say memory with significantly more complex logic. HBM is essentially the tick, where the interface was the new bit and they went with relatively simple memory/logic design. The tock step is to use the same interface but focus on a significantly more complex chip, lower power usage, lower latency, better/more parallel access, increased density. Will it be called HBM3 or a completely new memory type, doesn't really matter. The next memory change will be more like gddr3 to gddr5 most likely. The interesting parts of HBM are less the chips and more how it's all put together. TSVs, die stacking, interposers, very wide bus. the next gen memory will also be tsvs, die stacking, interposer and very wide bus, just a different type of chip stacked together.

More than likely trying to improve the Latency and power use side of things, HBM already allows a vast amount of parallel access compared to what GDDR5 does.

Kaapstad · 24 Mar 2016 at 22:35

Gerard said:
You didn't have a crystal ball either so lets not go down the route of rewriting history

You are absolutely right I did not have a crystal ball.

I do not know much.

There are people more qualified than me on these forums.

But at least I admit it, unlike some of the people who try and waffle their way out of admitting they have got something wrong.

Kaapstad · 24 Mar 2016 at 22:37

drunkenmaster said:
I would say memory with significantly more complex logic. HBM is essentially the tick, where the interface was the new bit and they went with relatively simple memory/logic design. The tock step is to use the same interface but focus on a significantly more complex chip, lower power usage, lower latency, better/more parallel access, increased density. Will it be called HBM3 or a completely new memory type, doesn't really matter. The next memory change will be more like gddr3 to gddr5 most likely. The interesting parts of HBM are less the chips and more how it's all put together. TSVs, die stacking, interposers, very wide bus. the next gen memory will also be tsvs, die stacking, interposer and very wide bus, just a different type of chip stacked together.

Is this your way of admitting HBM1 was not up to the job ?