AMD: R8xx Speculation

[Vincent]

You should have gotten a BMW.

 

[Silent_Buddha]

(have read the rest too)

Let me get this straight: Instead of selling as many chips as you can, you're proposing to not sell salvage models at all if you can't deliver vast amounts of stock for this?
Doesn't make sense to me.

Of course, it would be better to have as many chips to saturate demand, but selling 100k instead of 500k to me makes more sense than to sell zero instead of 500k.

Besides, you chose to ignore this:
"#6 …- maybe even sparing the salvage parts for chosen OEMs for the time being, if those salvaged parts should really be in short supply."


Again, I am not denying that you could sell more of a cheaper SKU than of a higher-level one.

No, I'm pretty sure his point is...

AMD for Rv6xx and Rv7xx hasn't had a high incident of salvage parts.

Salvage parts are generally sold with lower margins and thus lower price.

Lower price generally generates more demand.

If you don't have a good quantity of salvage parts you end up being forced to sell full working parts as salvage parts or end up with shortages of a product that people are now demanding. To the point where they won't even consider the non-salvage parts because they are now set on the price point of the salvaged parts.

In other words, without a good quantity of salvage parts selling them is a lose-lose situation with no upside.

As with 4830, once they stock piled enough salvage parts it was eventually launched/released for purchase. And when stock quickly ran out it was just as quickly EOL'd... Or at least that was the plan until they hit snags with Rv740 and 40nm production not ramping up as desired.

However, launching 4830 simultaneously with 4870/50 would have been disasterous for the bottom line since there wasn't many salvage parts.

Regards,
SB

 

[rpg.314]

Well, I think their x870+x850 strategy is already a pretty good one. Give the cheaper card 1 less power connector so it can't be OC'ed into the x870 range. And from that, whatever is left, sell it as a x830.

SB does have a point that selling salvage parts early could addict people to cheaper parts. On AMD's CPU side, deneb launched with all kinds of salvage parts. But now many of them have been EOL'ed. That's because their process wasn't as mature.

With gtx280, 260, the yields were too low to sell only 280 and the price delta bw 280 and a 260 was enough to make launching them simultaneously feasible.

So all in all, it is not straightforward as launch-all-in-one-go always.

 

[swaaye]

Similar to when you had a one in three chance of getting a top binned Athlon XP when buying an ultra cheap bottom feeding slow one.

I think more of Conroe. Say the E6300 that while having half cache could do ~80% overclocks on stock voltage, making it faster than even the official extreme edition chips of the time. Intel's yields must be fairly amazing for how their chips clock. Hell my Q6600 does a 25% overclock while undervolted by 0.1v.

/end derailment

 

[hoom]

Here is a review of the 4750.
http://en.inpai.com.cn/doc/enshowcont.asp?id=6778&pageid=5400

Pretty weird that whole same clocks as 4770 bit.
Can't think of a good reason for it other than ATI making clear the difference between 'non 6 pin' & '6 pin' SKUs (though that review is indicating that they expect to see 6 pin 4750s too).

 

[Blazkowicz]

this is weird as shit. the 4730 performed the same too (it was a 4870 crippled to 4770 specs). that gives a whole line-up of identically performing 47x0 parts, on par with the 4850.

 

[no-X]

HD4730 was released as a replacement of unavailable HD4770. Btw. in some drivers HD4730 is detected as HD4810 - maybe the name was changed to HD4730 to make its purpose (HD4770 replacement) more obvious.

The list of similarly perfroming boards is a bit longer... HD4730, HD4750, HD4770, HD4830, HD4850 and HD4860...

 

[Kaotik]

Is there any numbers on how many transistors the ATI tesselator in R6xx/7xx takes, and does the R8xx/DX11 tesselator take more than that?

 

[Jawed]

Another fundamental problem with using Cypress in salvage form is the configuration of the memory system.

If RBEs/L2s/MCs are being de-activated for yield purposes, then to convert a 256-bit chip to 128-bit chip results in 6 different combinations of active MCs.

How are you going to take those 6 different combinations and make a card? Are you going to have 6 different substrates, one for each combination of active MCs? In this scenario each substrate connects a different combination of MCs to the same memory connections on the card's PCB, so that the PCB design doesn't change. This enables you to have a relatively cheap PCB, with minimal layers and minimal size.

Or, do you use a standard Cypress board (relatively expensive) and omit the memory chips corresponding with the unused MCs? (This is what NVidia did with 8800GTS when it launched with 8800GTX. NVidia also seems to have done the same thing with the older GTX260s, which shared the same PCB as GTX280).

Though it seems that nowadays NVidia doesn't even seem to be salvaging MC/ROP partitions on GT200b to make GTX260 and GTX275. This is because the PCB design no longer has space for 16 memory chips, with the option to remove those corresponding with a broken partition. Instead the board has space for just 14 chips. Is NVidia using one of 8 substrates to connect GT200b to an unchanging PCB? Or is NVidia just junking GT200b when any partition fails (or when any one of the 7 partitions it has chosen to use fails)?

So, how does HD4730 map 128-bits from 256-bits?:

http://www.elitebastards.com/cms/in...sk=view&id=734&Itemid=27&limit=1&limitstart=2

Looks like a full board with half the memory chips missing. Quite expensive.

Jawed

 

[Kef]

According to fuzilla RV870 (Evergreen if you prefer ...) would be 1.6 times faster than RV770. If true, that could imply 1280 shader processors (1280/800 = 1.6). Of course we don't know the clocks but they are going to be in the same ballpark anyway ... 750 - 850 Mhz ..

http://www.fudzilla.com/content/view/15287/1/

<edit>Is 1.6 * RV770 performance really going to be enough to compete with GT300?</edit>

 

[sc3252]

According to fuzilla RV870 (Evergreen if you prefer ...) would be 1.6 times faster than RV770. If true, that could imply 1280 shader processors (1280/800 = 1.6). Of course we don't know the clocks but they are going to be in the same ballpark anyway ... 750 - 850 Mhz ..

http://www.fudzilla.com/content/view/15287/1/

<edit>Is 1.6 * RV770 performance really going to be enough to compete with GT300?</edit>

Now you also have to consider their vagueness when they put down 1.6x faster. They put down that it would be 1.6 times faster than the 4870 but said compared to the rv870. This might be true, but when it is released it could be 2x faster at the top and 1.7x faster at the 5850. Wow they are within 10% now since they claimed it was 1.6 times faster.
Really I think 1.6 times faster is a good bet, but since they never put down which part was faster they could be just printing something as usual to look like they are in the know.

 

[Ailuros]

I'd very surprised if either way Cypress wouldn't reach up to twice 770 performance. That 1.6x sounds a tad too "averaged" to me.

<edit>Is 1.6 * RV770 performance really going to be enough to compete with GT300?</edit>

If G300 is only 1.6x times faster than GT200 most likely yes

 

[Kef]

I'd very surprised if either way Cypress wouldn't reach up to twice 770 performance. That 1.6x sounds a tad too "averaged" to me.

if G300 is only 1.6x times faster than GT200 most likely yes

I agree, I would expect something like twice the performance as well.

 

[fehu]

should be roughly 1.6 times faster than the RV770 chip

doesn't it means tha if RV770 is "100", then RV870 is "260"?
a lot more than "twice", and maybe somewhere near 1600 enhanced shader cores

 

[Ailuros]

I agree, I would expect something like twice the performance as well.

Many if not all of us have been many times wrong on what we "expected". You made a hypothetical question and I couldn't resist answering with a hypothetical answer that's all

 

[Ailuros]

doesn't it means tha if RV770 is "100", then RV870 is "260"?
a lot more than "twice", and maybe somewhere near 1600 enhanced shader cores

The first it should mean to anyone reading it, is that neither ALU unit counts nor sterile FLOP numbers define the final performance of a GPU.

Just for the record's sake if RV770 is 100 than in order to reach 1.6x you need to add 60%.

 

[Silent_Buddha]

It depends does he mean it's 160% the speed of Rv770 = 60% faster?

Or does he mean it's 160% faster than Rv770?

Personally, I'm expecting it to be around 50% faster, and anything more will be a pleasant surprise. Same goes for GT300. Then again I've been around for a while where new video card generations don't generally do more than 50% better compared to the last generation.

Think many have been spoiled by the past few years.

Regards,
SB

 

[leoneazzurro]

Oh, casually 2 Teraflops divided by 1.2 Teraflops is 66% more.

So it's purely a Teraflop-based rumor.

 

[neliz]

I wouldn't care about "1.6" since it's wrong anyway, let's start at 70% faster than RV770 and work from there.

 

[Silent_Buddha]

So after nearly 2 pages. We're back to...

He either means it's 60% faster or 160% faster.

And since it's Fuad. There's a huge chance it's neither of those.

Regards,
SB