GCN vs VLIW5 architecture performance comparison

[thilanliyan]

http://ht4u.net/reviews/2012/amd_radeon_hd_7700_test/index4.php
Translated:
http://translate.google.com/transla...views/2012/amd_radeon_hd_7700_test/index4.php

Originally posted in XS:
http://www.xtremesystems.org/forums/showthread.php?279169-GCN-vs.-VLIW5-performance-improvements

Please no arguments about cost k?
Everyone knows the 7 series cards are not priced where most of us want them to be.

 

[BD231]

Ich kann nicht lesen

 

[Idontcare]

http://ht4u.net/reviews/2012/amd_radeon_hd_7700_test/index4.php

Originally posted in XS:
http://www.xtremesystems.org/forums/showthread.php?279169-GCN-vs.-VLIW5-performance-improvements

Interesting analysis :thumbsup:

We see the new microarchitecture has higher IPC. Do we have any idea on how it fairs in terms of die-size efficiency?

 

[thilanliyan]

Ich kann nicht lesen

Warum?

EDIT:
Scheiße, es ist in deutscher Sprache!

http://translate.google.com/transla...views/2012/amd_radeon_hd_7700_test/index4.php

 

[thilanliyan]

Interesting analysis :thumbsup:

We see the new microarchitecture has higher IPC. Do we have any idea on how it fairs in terms of die-size efficiency?

Efficiency as in performance per die size? SPs per die size?

 

[Concillian]

This is def interesting since ROP and TMU performance is equal (minus GCN efficiency improvement, a pure shader comparison if TMU and ROP are "equal".

28% better is pretty good, and bodes well for 7850 if it does end up at the rumored 1280 / 80 / 24. My only fear is that it ends up at like 896 / 56 / 24.

Maybe they'll try to straddle current GPUs with the 7770 between 5770 and 6850 and put the 7850 at 896 or 1024 to land between 6870 and 6950. That would suck.

 

[Concillian]

Interesting analysis :thumbsup:

We see the new microarchitecture has higher IPC. Do we have any idea on how it fairs in terms of die-size efficiency?

AT:

Cape Verde, the GPU at the heart of the 7700 series, is AMD’s smallest 28nm GPU. With a die size of 123mm2 it’s only a hair bigger than the 118mm2 Turks GPU that powers the 6600 series

Juniper / 5770 is 166mm2 according to ATs 5770 article.

 

[tviceman]

Those are some pretty nice architectural improvements!

 

[BD231]

Warum?

EDIT:
Scheiße, es ist in deutscher Sprache!

http://translate.google.com/transla...views/2012/amd_radeon_hd_7700_test/index4.php

magisch! I thought this switch was meant more for GP-GPU stuff, how are they certain these gains come purely from GCN exactly?

 

[notty22]

You can find a few gem benchmarks where we see the improvements.
Two games that favored Nvidia .
The 7750 performed faster than the 6850/5850 in Crysis 2.

Batman AC
7750 Faster than the 5850 / 7770 hold up against the 6850

 

[busydude]

Juniper / 5770 is 166mm2 according to ATs 5770 article.

Does it mean that a 5770 would be ~half that size when die shrunk on 28nm?

 

[Skurge]

Does it mean that a 5770 would be ~half that size when die shrunk on 28nm?

I remember IDC saying that TSMC's 28nm process is just their 32nm process rebadged. So I don't think so.

 

[busydude]

I think this time it is a bit different. 40nm->32nm is half node.. whereas 40nm->28nm is a full node transition. I think it was Benskywalker who made that observation.

 

[WMD]

You can find a few gem benchmarks where we see the improvements.
Two games that favored Nvidia .
The 7750 performed faster than the 6850/5850 in Crysis 2.

Batman AC
7750 Faster than the 5850 / 7770 hold up against the 6850

Crysis 2 is just an extreme tessellation benchmark just like GT4 in 3dmark 11. The bottleneck is tessellation rather than compute perdormance. It goes without saying the more advanced tessellator and the higher clockspeed on the new cards will perform better than the 6850. But at 28fps average the results are more academic than any practical use. You will be playing without tessellation on this class of graphics card and the 6850 will still perform better.

 

[formulav8]

I remember IDC saying that TSMC's 28nm process is just their 32nm process rebadged. So I don't think so.

Well I believe each node means overall the components are around 28nm. Some are more and some less. Its more of an average I think??? Can't remember anymore... Getting to old these days

 

[Lonbjerg]

I think this time it is a bit different. 40nm->32nm is half node.. whereas 40nm->28nm is a full node transition. I think it was Benskywalker who made that observation.

Wrong.
Full nodes:
90 nm
65 nm
45 nm
32 nm
22 nm
Half nodes:
80 nm
55 nm
40 nm
28 nm
20 nm

 

[busydude]

What term do you suppose is the right one to use?

A full half node transition?

 

[Idontcare]

I remember IDC saying that TSMC's 28nm process is just their 32nm process rebadged. So I don't think so.

The shift happened at 45nm. Everything post 40nm is back to regular math.

Wrong.
Full nodes:
90 nm
65 nm
45 nm
32 nm
22 nm
Half nodes:
80 nm
55 nm
40 nm
28 nm
20 nm

You count the spans between nodes, not whether the nodes were full or half.


40nm->32nm is a half-node transition.

32nm->28nm is a half-node transition.

40nm->28nm is a full-node transition, it just so happens to transpire between two half-nodes.

45nm->32nm is also a full-node transition which also happens to transpire between two full-nodes.

But it should be noted that neither full-nodes or half-nodes have a formal definition in the industry. They are labels of convenience.

Their numerical differentiators does correlate with chronology of successive release, but the numerical differentiators itself has no bearing on the feature sizes of the components on the chip. (they used to, but that convention broke down decades ago, literally, back in 0.35um)

 

[Idontcare]

What term do you suppose is the right one to use?

A full half node transition?

You just call it a full node transition. If you want to expand on it any further then you specify the nodes themselves.

"A full 40nm to 28nm node transition"

 

[AtenRa]

Full node = ~2x more transistors at the same die size

Full node needs a complete redesign of the chip, half node doesn't need a redesign of the chip.

Intel nodes

90nm to 65nm = full node

65nm to 45nm = full node

45nm to 32nm = full node

32nm to 22nm = full node

TSMC nodes

90nm to 80nm = Half node

90nm to 65nm = full node

65nm to 55nm = Half node

65nm to 40nm = full node (more like Full + half node because we had up to x2.33 transistor density)

40nm to 28nm = full node

28nm to 20nm = full node

Edit: Just to add, AMD's 28nm is a half node over 32nm with ~10% shrinkage

 

[Lonbjerg]

AFAIK Intels dosn't bother with halfnodes.
But then again they are special case, considering all their FAB's and R&D bugdet...

 

[AtenRa]

Does it mean that a 5770 would be ~half that size when die shrunk on 28nm?

Lets see,

HD5770 = 1040M transistors at 166mm2 = 6,26 transistors per mm2

HD7770 = 1500M transistors at 123mm2 = 12,19 transistors per mm2

Almost 2x transistor density, i will say that HD5770 could be close to half the die size at 28nm.

 

[Phynaz]

Lets see,

HD5770 = 1040M transistors at 166mm2 = 6,26 transistors per mm2

HD7770 = 1500M transistors at 123mm2 = 12,19 transistors per mm2

Almost 2x transistor density, i will say that HD5770 could be close to half the die size at 28nm.

It's not that simple. You're comparing apples and oranges because you are comparing two different architectures.

Basically you don't have enough information to determine this, because 5770 was never designed to 28nm rules.

 

[AtenRa]

Im comparing transistor density, not architectures.

1040M transistors at 166mm2 with 40nm process
1500M transistors at 123mm2 with 28nm process

Transistor density is almost x2, so HD5770 could be close to half the die size at 28nm process

TSMC's 28nm HKMG

Edit: added photo

 

[Phynaz]

Transistor density is design dependent.

For example, cache is typically very dense. On the other hand, speed paths aren't.