• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

ATI Radeon HD 6000 Series ''Southern Islands'' Graphics Cards For Sale from November

Rroff said:
Tessellation is one of the areas they have talked about boosting performance in with these cards.

Don't expect a massive performance boost - tho you might see it in some areas/benchmarks - this round is more about making things work more efficently, better balancing the card so that some areas aren't held back by others, boosting performance in some DX11 and eyefinity scenarios and so on.
Click to expand...


There is that but there's not much point in bring out new cards if there’s no solid reason to upgrade. When AMD launched the 4000 series it represented a 70% increase in performance over the older 3000 line which was built on the same manufacturing process. Given the ATX wattage limit AMD will do well to give us the same sort of bump (but who knows) but I tend to think the performance increase will be greater then what others have speculated judging by recent history.
 
If ATI are over hauling the pipeline then tessellation performance will probably improve by default I would think. Not that it will matter much in anything other than a benchmark.

Anyone remember the chart from around this time last year ? It had a mysterious set of spec's for a planned card.

IIRC it was something like a 2000 SP core with a 96 - 48 configuration and a 320 bit bus.
 
straxusii said:
They have lots of room.

5870 die size is 334 (mm squared)
480 die size is 529

So even if they shoot for a die size of 450, to be safe on power and heat, that's still plenty of potential.

Add in the more mature process, and that this is a new architecture, and it's looking possible they can produce a 6870 that can get close to a 5970 and certainly beat the 480 hands down.

The only potential issue I see is if they have to increase the memory bus to 384 bit if 256 creates too much of a bottleneck.
Click to expand...

2900xt was 420mm, and 529mm is literally no profit, theres a reason Fermi launched and Nvidia went from 100mil profit, to 150mil loss, something was being produced and losing money, to the tune of a quarter of a billion turnaround.

AMD won't come close to a 529mm core, not even the same ballpark. I'd think 400mm would be about the limit, basically as small as possible.

As for 3870 to 4870, the 3870 was just, horrible, in every which way, remember it was slower than a 2900xt, and had LESS transistors than a 2900xt, despite dropping 2 process's, 2900xt, 80nm, 700mil trannies, 420mm die.

3870 192mm die, 55nm, 666 trannies, slower, extremely bandwidth limited, 256mbit with gddr4(which didn't offer quad pumped like gddr5, it was just uber fast gddr2 essentially). GDDR5 doubled the bandwidth on the 4870, thats honestly where quite a bit of the speed came from, with a pretty same small die size increase. But thats largely because the 3870 was a cut down 65nm design, only on 55nm because, well, it was out and convienient. The 4870 was the only chip ever really designed for 55nm.

It really has zero bearing on whats going to happen on a much worse process.

You have to remember what the 3870 was, it STILL had the 512bit internal ringbus memory, thats probably 150-200million transistors, its a completely massively transistor heavy but ridiculous bandwidth bus, its fantastic tech and it sounds like Intel is moving that way for their chip in the future, the "normal" way saves huge amounts of power and space. AMD well over doubled the shaders, with only a 35% die size bump, with double the TMU. They did that because they cut about 25% of the too early to be useful stuff(another less nice word would be waste :p ).

AMD pretty much did double the transistor count, and die size of the 3870 to the 4870, they just did so with waste.

The thing is, the 5870 doesn't have a massively transistor wasteful ringbus memory architecture, it really doesn't have much that can be cut easily. Because they can't bump the die size up 60-70%, they simply can't over double the shaders like they did from 3870-4870, even doing so they didn't double the performance. The 3870 to 4870 while artificial(the 3870 should have been 6 months earlier, on the 65nm process) is the single biggest leap forwards AMD have ever had.

Unless AMD finds a way to efficiently go from a 256bit crossbar memory interface, to a 64bit one to cut as much die space as they did from the 3870, it just isn't going to happen.
 
RavenXXX2 said:
Higher clock rates and a more efficient process says it is possible. I don't care what others would prefer or expect performance wise from new cards, it's my opinion only, from my experiences with current tech at stock and overclocked.
Click to expand...

They won't be using a more efficient process though. They are using the same process as the previous cards.
 
straxusii said:
They have lots of room.

5870 die size is 334 (mm squared)
480 die size is 529

So even if they shoot for a die size of 450, to be safe on power and heat, that's still plenty of potential.

Add in the more mature process, and that this is a new architecture, and it's looking possible they can produce a 6870 that can get close to a 5970 and certainly beat the 480 hands down.

The only potential issue I see is if they have to increase the memory bus to 384 bit if 256 creates too much of a bottleneck.
Click to expand...

Another quick point to make, remember yields will be on anot quite exponential curve, so you make something 10mm^2 and your yields will be silly high, 520mm^2 would literally be right on the border of manufacturable and unmanufacturable, considering at launch there weren't releasable yields of a 512sp part, unmanufacturable would be most fair, yields will come up quickly because its right at the end of a very steep line on that chart, but you'd want to be going from what is probably 2-3 512sp cores per wafer, to at least 30-40 before the high end part would be able to be released in large enough numbers and at low enough cost to be a highly profitable card.

Die size wise, I really can't see AMD wanting to produce something 2900xt sized again, but if it did that would be around the limit, anywhere around 430mm^2 up and I reckon they'll be unable to make a proper profit on a £300 card, and thats just bad business, AMD have made such serious inroads specifically by NOT doing that, and its working very well, and the guys still in charge firmly believe small core strategy is working.

If 32nm was about, and good(it wasn't scrapped because it was just late, but because, it sucked balls aswell) then a 5970 speed part would be completely expected.

People also shouldn't get too excited about 28nm, I'm sure AMD/Nvidia could jump right to 28nm and have parts 3-3.5x the speed of a 5870, but 28nm is set to be around for a while, 22nm is miles off to be honest, so its likely AMD/Nvidia will be very conservative and plan to allow 2 generations to be on 28nm.
 
Turambar said:
They won't be using a more efficient process though. They are using the same process as the previous cards.
Click to expand...

I'm gonna go ahead and not point out that the 2900xt was 743/825mhz 3870 was 775Mhz/900, the 3870x2 was 825/900, the 4870 was 4870x2 were 750/900 , the 4890 was 850/975 and then the 5870 is 850/1200.

Which logically means that on the same truly awful process we can clearly expect what 1Ghz stock speeds ;)

Thats since, May 2007, in over 3 years from 80, 65, 55 and 40nm we've gone up a tiny bit over 10% in clock speeds. One of Nvidia's biggest failings is design a card for high clock speed on a process that can't do it, we've seen rumours of how much juice and the yields available for 512sp cards at the designed for clock of 800Mhz, Nvidia also have been stuck around the 1500Mhz shader speed for the best part of 3 years aswell. Clock speeds are really not going anywhere.
 
Last edited:
So is the 6xxx series definitely staying on 40nm? I'm looking on the net and some are saying its going 28nm and some saying 40nm :confused:

If they're sticking with 40nm how much performance is to be had on top of the 5xxx series line?
 
You must log in or register to reply here.
Back
Top Bottom