GCN vs VLIW5 architecture performance comparison

[Leadbox]

While trying to explain it all to my thick self, somebody said to look at it like so:

40*40=1600
28*28=784

784*100/1600=49% = full node
A significantly higher % would mean half node

Something like that....I think

 

[AtenRa]

Transistor density is design dependent.

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

You are talking about architectural/design transistor density. Im talking about manufacturing/process transistor density.

From 40nm to 28nm, TSMC says that we get x2 the transistor density and we can "confirm" this from the transistor density the 7700/7900 has over 40nm chips.
GPUs have small cache sizes compered to CPUs. That makes it easier to compare transistor density even against different architectural designs.

Now, because of this x2 density, HD5770 could be close to half the die size if designed and manufactured at 28nm.
It would certainly be much smaller than 123mm2 even if it would not be close to half the die size.

 

[Arkadrel]

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?

Good question...

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.

1040M transistors * X = 12.19 transistors pr mm^2
X = 85.4 mm^2

*IF* they could get same amount of transistoers pr mm^2, which they might not be able to because as others have said its a matter of design.
The 5770 design varies alot from that of a 7770.

Lets just go with a safe ~90mm^2 for a hypothetical 28nm 5770/6770.


Then you just have to look at how much performance a 7770 offers over a 5770.... at the same power usage
(we dont know a 28nm hypothetical 5770's power draw), and compaire that to the size differnces.

Theres no answear to the question because its based on assumptions and guesses.

The 7770 does have alot of features and stuff a 5770 would be lacking, so even if it boiled down to the hypothetical 5770
being more die-size effecient, the question is if that would even make it a better product?

The only thing to take away from this thread at the end of the day is IPC went up with the GNC design.



***edit:

Also holy crap at the 7770 being 123mm^2 thats tiny.

My gut tells me the 5770 might actually be a tiny bit more die-size effcient than the 7770.
But again that probably comes at the cost of features and stuff.


Size (guestimation):
123mm^2 vs 90mm^2 = 36% differnce.

performance (Techpowerup chart based off avg in games)
7770 vs 5770 = ~30% differnce (@1920x1200)


I guess it depends on what game/application you run though... because:

Here a 5770 score 17.2 fps vs the 7770 factory oc version fro XFX scoreing 29.4 fps, is a pretty big step.
About a ~71% differnce in performance here in Heaven Benchmark.

IF a game uses as much tessellation as Heaven Benchmark does, then the 7770 might actually beat a 5870 in it

 

[Tuna-Fish]

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

I remember reading a quote that said that Redwood was strictly pad-limited, or put in other words, it's simply not possible to make a chip with a 128-bit GDDR5 interface that's smaller than 104mm².

Of course, they could make a 28nm Juniper with more shaders.

 

[MrK6]

Excellent article and analysis. I wouldn't have expected the IPC/performance efficiency to be that great.

 

[AtenRa]

I remember reading a quote that said that Redwood was strictly pad-limited, or put in other words, it's simply not possible to make a chip with a 128-bit GDDR5 interface that's smaller than 104mm².

Pad limited means you can install more (cores) until you fill the pad. If it was Core limited at 40nm then yes, they couldn't make a smaller Redwood with the same SP count and 128-bit memory controllers smaller than 104mm2.

But even if they were core limited at 40nm they are not at 28nm

Edit: slide added

Edit 2: If it was pad-limited at 40nm it will also be at 28nm.

 

[maddie]

Why is it so critical to use a square die? If you're pad limited, can't the designers correct this with a rectangular form?

I'm sure there are other factors involved placing practical limits to how far this can be exploited, but why not a more rectangular shape than a square?

Perimeter to surface area is lowest for rectangular quadrilaterals.

 

[Phynaz]

Why is it so critical to use a square die?

It's not. Don't confuse the die with the package.

 

[Olikan]

GCN vs Vliw-4

http://www.inpai.com.cn/doc/hard/168475_30.htm

 

[biostud]

the 5770 has ~1B transistors while the 7770 has ~1.5B and running at same speed we get from ~0-35% increase in speed, depending on the game.

 

[biostud]

GCN vs Vliw-4

http://www.inpai.com.cn/doc/hard/168475_30.htm

why didn't they use the 6970 vs 7870?

 

[Olikan]

why didn't they use the 6970 vs 7870?

6930 and 7870 have same shader count :thumbsup:

 

[biostud]

6930 and 7870 have same shader count :thumbsup:

But the overall chip design would be a better comparison, since they use around the same amount of transistors (2.64B vs. 2.8B), so maybe a die shrunk VLIW-5 chip would be as efficient as a GCN based chip.