AMD: R8xx Speculation

[hkultala]

Interesting numbers even if they're fishy. Since Cypress is "two shader blocks", each of 10 SIMDs, what we might be seeing there is a GPU with four SIMDs turned off in one of the shader blocks. Such an unbalanced configuration would be problematic...

Jawed

Where do you base the claim of "two shader blocks" ?

AFAIK it's just 20 shader processors instead of the 10 of 4800-series, without any "blocks".

Some of the figures just put them in two columns because it makes the figures look nicer (having all 20 vertically would make the figures too tall)

 

[Kaotik]

Where do you base the claim of "two shader blocks" ?

AFAIK it's just 20 shader processors instead of the 10 of 4800-series, without any "blocks".

Some of the figures just put them in two columns because it makes the figures look nicer (having all 20 vertically would make the figures too tall)

It has 2 blocks, or at least that's how Dave Baumann explained it here, so it's not just due size on graphs
http://forum.beyond3d.com/showpost.php?p=1383812&postcount=41

Mint, take a look at the patent diagram that Jawed singled out on the previous page.

The SC's are deteriming the tiles; they are both being fed triangles from the the Primitive Assembler; those triangles are duplicated from the PA, but, at any point in time, will the SC's be working on different triangles? Yes, because they are there to ensure that there is sufficient workloads to be fed into both shader engines and have them well balanced and busy (and I've yet to see much reason to suggest that Fermi isn't operating on a similar system / structure in terms of tiling the pixel load over the 4 shader clusters).

The discussion was about the dual rasterizers in Cypress

 

[Arnold Beckenbauer]

Aren't there two Dispatch Processors (RV870)?

 

[Mindfury]

 

[Arty]

1120 alus and more importantly only 16 rops .. underwhelmed.

 

[GZ007]

With 16 Rop-s and the clocks no wonder it had those results. Now thats a waste of silicon if they are from the cypress cores.
I don know who will buy that thing for 230-240 EUR. 5770 is 150-160 EUR and even 5850
which will be much faster for 260 EUR.
Shouldnt that thing fill the gap betwen 5770 and 5850

 

[smw]

And that's not the strangest thing - it says that the maximum board power is 175W so 25W more than the 5850.

 

[Squilliam]

It looks more like an HD 5790 than an HD 5830! I wonder why that is? Do they really have a bunch of cores lying around that that strange combination of features is the only way to put them to good use?

High maximum board power implies high leakage, the loss of 16 ROPs/480SP indicates quite a major defect. That means that most of the spare dies aren't in particularly good condition and suffer from a either a major defect or a major problem with leakage or both. I wonder how many fall inside this category and outside the bounds of the 5850/5870 and how many will be otherwise good dies going to the bad pile?

 

[GZ007]

This was posted on bsn a while as nvidia response to the tough AMD comunity questions

AMD: Is it true that yields are currently just 2% due to poor engineering and lack of ‘dual redundant vias’ connecting the chip’s metal layers, and that most GF100 chips won’t even be able to hit 650MHz.

BSN*: This is quite an interesting question, given that that story was coming from satire website. In any case, Cypress has the same issue, given the mountain of Cypress wafers at TSMC - those chips can't even hit 5850 speeds. But hey, don't shoot the messenger.

Maybe these are those mountains of wafers in tsmc that will turn into something. For them it will be still more than 0$ for a wasted waffer.

 

[Shtal]

I was thinking lately about Cypress refresh chip, If ATI do same - as they did HD4870 ----> HD4890.

How much performance will be increase? GPU + Memory; based on the speculation in the image below.

Edit: I was thinking about an average 15% percent improvements, maybe with 8x AA enable.

 

[itsmydamnation]

It looks more like an HD 5790 than an HD 5830! I wonder why that is? Do they really have a bunch of cores lying around that that strange combination of features is the only way to put them to good use?

High maximum board power implies high leakage, the loss of 16 ROPs/480SP indicates quite a major defect. That means that most of the spare dies aren't in particularly good condition and suffer from a either a major defect or a major problem with leakage or both. I wonder how many fall inside this category and outside the bounds of the 5850/5870 and how many will be otherwise good dies going to the bad pile?

i dont think its leakage i think its clock, seeing it has already been explained that the disabled units still get power, so its clocks are in the middle of a 5850 and 5870 and so is the power requirements.

 

[eastmen]

I was thinking lately about Cypress refresh chip, If ATI do same - as they did HD4870 ----> HD4890.

How much performance will be increase? GPU + Memory; based on the speculation in the image below.

Edit: I was thinking about an average 15% percent improvements.

http://forum.beyond3d.com/showpost.php?p=1397342&postcount=57

lighthouse has his 5870 running at 1ghz/1.2ghz which should help clue you in on what such a chip can do.

Personaly I on't think 1ghz is fast enough.

 

[Lightman]

http://forum.beyond3d.com/showpost.php?p=1397342&postcount=57

lighthouse has his 5870 running at 1ghz/1.2ghz which should help clue you in on what such a chip can do.

Personaly I on't think 1ghz is fast enough.

WOW!
I've grown from a man to a house

I must go on drastic diet soon then

 

[neliz]

With 16 Rop-s and the clocks no wonder it had those results. Now thats a waste of silicon if they are from the cypress cores.
I don know who will buy that thing for 230-240 EUR. 5770 is 150-160 EUR and even 5850
which will be much faster for 260 EUR.
Shouldnt that thing fill the gap betwen 5770 and 5850

Where do you get your prices from? I bought an OC 5850 for €239. This 5830 should retail for €180 since 5770 OC models can be found for €145.

 

[CarstenS]

Where do you get your prices from? I bought an OC 5850 for €239. This 5830 should retail for €180 since 5770 OC models can be found for €145.

Maybe some more of the PDF already leaked somewhere? And good luck to try and get a 5850 reliably for under 260 Euro (here in germany). Right now it's more like 280...:
http://geizhals.at/deutschland/?cat=gra16_512&asuch=HD 5850&v=l&plz=&dist=&xf=132_1024&sort=p

 

[neliz]

Maybe some more of the PDF already leaked somewhere? And good luck to try and get a 5850 reliably for under 260 Euro (here in germany). Right now it's more like 280...:
http://geizhals.at/deutschland/?cat=gra16_512&asuch=HD 5850&v=l&plz=&dist=&xf=132_1024&sort=p

MSRP for the 5830 is $219 - €199

I'f I were you, I'd just order the card from holland and pay €6 more on the shipping fee.

 

[Mindfury]

 

[GZ007]

It seems AMD had a good reason to leave out DP from the 57xx cards.

 

[mczak]

Those specs indeed look disappointing. It's strange to see 16 rops but still 256bit memory interface. Its power efficiency looks downright terrible.
With 6 simds out of 20 and half the rops disabled, that really looks like a garbage bin to be able to use any die (which works at all). The loss of so many simds wouldn't be that bad imho at that clock (still quite a bit more flops than HD5770) but half the rops... There can't even be cost savings with the pcb since power draw is so high (and memory interface apparently stays at 256bit). Interesting that clock is still relatively high, I'd guess though they might use about the same voltage as HD5870 (or even higher) so possibly any chip can reach this.

 

[Alexko]

Those specs indeed look disappointing. It's strange to see 16 rops but still 256bit memory interface. Its power efficiency looks downright terrible.
With 6 simds out of 20 and half the rops disabled, that really looks like a garbage bin to be able to use any die (which works at all). The loss of so many simds wouldn't be that bad imho at that clock (still quite a bit more flops than HD5770) but half the rops... There can't even be cost savings with the pcb since power draw is so high (and memory interface apparently stays at 256bit). Interesting that clock is still relatively high, I'd guess though they might use about the same voltage as HD5870 (or even higher) so possibly any chip can reach this.

I'm guessing the TDP is set so high because there'll be a lot of variability between products in terms of leakage—this is, after all, the ultimate salvage part. Plus the clock is relatively high, close to that of the 5870: without power gating and with TSMC's 40nm process's legendary leakiness, it's not unexpected. I'm a bit more surprised and disappointed by the disabled ROPs. I wonder if this is due to actually high defect rates or just a commercial decision. Cypress has 4 memory controllers with 8 ROPs per controller, so I guess they just disabled 4 ROPs per controller, hence the identical bus width.

In the end it has a bit over 30% more processing and filtering power than the 4890. It has half the ROPs but if I'm not mistaken they're twice as good, and with the clocks being almost equal, it shouldn't be too much of an issue. Memory bandwidth is about the same too...

So this looks like an HD 4890 with a bit more punch, DX11 and a slightly lower power draw, perhaps much lower depending on how lucky you are. Provided it's not too expensive, it should be a decent deal.