AMD: R8xx Speculation

[LordEC911]

To cut costs, it'd make sense, to first market the fully functional and salvage parts (i.e. xx50 and xx30) and later add the speed binned xx70 model - given their supposed speed increase over the competition, DX11 not to mention. Those two factors should provide enough momentum for great sales.

Correct, that would be the best way but hard to do.

Cutdown parts obviously have a cheaper MSRP and therefore need a higher volume to substain supply. If you are hitting the yields that you are planning for, no idea what that target is 80-90%(?), you have a low number of defects. Ontop of that how many of those defects can be binned into a salvage part, again throwing numbers out but maybe 40-70%(?). I think that is the main reason that on most midrange/performance chips we don't see salvage parts until the base chip is EOL or near EOL.

 

[brain_stew]

USD 349 ---> Rip off

"BOM" of every add-in card(5800 series) may be within the region of 120~150 USD.

Honestly, I don't think us as consumers can really complain with pricing in that region. Before the 4800 series hit the scene, Nvidia were set to launch the GTX 280 at $650, flash forward 1 year and you can double that performance for less than half the cost. Hardly, a bum deal. really. This is without direct competition for several months, ATI could have quite easily got away with skimming the market at $500+, with them having the only DX11 parts on the market.

 

[swaaye]

$349 high for a new high-end card!? Vincent must be new to the world.

 

[brain_stew]

$349 high for a new high-end card!? Vincent must be new to the world.

Seriously, we've never had it this good, some people have got a very short memory.

We've got a 2 teraflop card launching at $200!! 2 years ago that would buy you a ~115 Gigaflop 8600gt. People are seriosuly complaining that you can get more than 17x the compute power for your dollar compared to just 2 years ago? Can anyone point to a point in history where we've had as large a collapse in pricing as what we've seen over this 2 year span?

 

[Jawed]

I was referring to saving costs on the initial offerings.

Earlier you said "To cut costs, it'd make sense, to first market the fully functional and salvage parts (i.e. xx50 and xx30)" so I'm asking you, where's the HD4730?

Jawed

 

[neliz]

so I'm asking you, where's the HD4730?

Jawed

It's out there.. it's a salvaged RV770

 

[Jawed]

Precisely my point.

ATI's design doesn't generate reams of salvage parts.

Jawed

 

[Bouncing Zabaglione Bros.]

Precisely my point.

ATI's design doesn't generate reams of salvage parts.

Jawed

Obviously it's much better to have too many good parts than too many bad ones!

 

[turtle]

Precisely my point.

ATI's design doesn't generate reams of salvage parts.

Jawed

And now something to wake everyone up...

[turtle=major digression AND friendly jest]

Actually, I didn't want to go backwards when you first mentioned it, but you keep bringing it up...there IS a 4750 based on the rv740. You could call it a salvage part...I call it a "fixed" part, and 4770 was the salvage part, and I'll tell ya why.

4750 didn't come along that long ago, and really is just a fixed 4770, not requiring a power connector and using the same clocks (including gddr5). It even overclocks the same (havn't seen it at higher voltages though) and AFAIK uses the same stock voltage as 4770. I imagine after TSMC got their leakage under control, AMD was happy to release a leakage-stable part with "4770's" true TDP of ~50W (per Xbitlabs' review of 4770) and get rid of a part from the BOM.

Granted, it's only available in China...but it could just as well exist elsewhere, and probably would've had 40nm been fixed earlier.

It doesn't because you know as well as I, trying to wrangle perception of rv740 back from the weirdness of it's initial (sometimes?) leaky part, with oddly-needed power connector would be counter-intuitive so close to the launch of it's replacement that will fetch a slightly better margin.

The point is, Rv740 didn't have a salvage part because it initially WAS a salvage part. It's rarity because of yield, corresponding price, and leakage requiring the power connector made it a redundant, if not a worse option than the 4850, not to mention less-than-ideal TDP for notebooks, hence why we never really saw a mobile version. Now that it's fixed, it's irrelevant because of the impending Juniper (also on the fixed 40nm process) and clearing of Rv770 stock.

I STILL stand by the notion that the initial parts were to be what the 4750 is, with the 4770 having higher clockspeeds and 5ghz GDDR5, but the time for that argument is long passed. I believe Rv740 became a (much-needed) experiment on 40nm rather than the truly competitive (and replacement) part it was meant to be.

At any rate...yeah. 4770, in short, is a salvage part because it needed >75W to guarantee 750mhz on a die that is only 137mm2. Hardly outstanding for what it was, looking in the rear-view mirror and ahead at Juniper.

Juniper is supposedly what, 60W TDP and 181mm2. That alone should tell you something. While that is for the mobile parts, and may be for the 'pro' version, I'll be surprised if the "xt" version is much above 4770's listed TDP of ~80W.

Granted, your argument of 'salvage' might be completely defective units...Mine includes units that may require a (comparatively) shit ton of power to work reliably. While you could say the TDP that came about was impressive for the transistors, it was not impressive for the 40nm process, and I believe it was a failed product for AMD, especially when you think about how it's now-completed 4750 ~50W brethren would've competed in the mobile and performance marketplaces...and especially how a higher-clocked version could've effectively replaced rv770 and made oodles for the company.

[/major digression and bitching]

On a related notion, I still can't help but wonder if those rumors from Nappy @ Chiphell about the buses had a hint of truth. If you think that Juniper and Cypress were originally meant to be 960sp and 1920sp with 192 and 384-bit buses, it's not unreasonable to think that those parts would be replaced on 32nm (GloFo) with 128-bit and 256-bit parts using 7gbps GDDR5. They would seem to fall into place together.

This makes sense because not only is (256*7000)/8 ~ (384*5000)/8, but because it would allow ATi to essentially do an optimized version of R600->R670 or G80->G92. Start off not quite so big (205mm for Juniper, as was rumored, which became ~180mm, which fits), and end up with a superior product not quite so small, probably around the size of Rv740, on 32nm, with higher core clocks. This would've meant an ~400mm2+ Cypress that could've met/exceeded GTX295/G300, with a refresh part around the size of rv670-rv770. That'd make sense, no?

I'm starting to think this was AMD's game plan, and it could've been a good one had TSMC fixed 40nm earlier.

 

[Jawed]

The point is, Rv740 didn't have a salvage part because it initially WAS a salvage part.

No. (I know you admit your deviant definition later...)

RV740's better power/temps now seem to be in the same sense that RV770 also has better power/temps now than the initial chips did 14 months ago. That stuff happens simply because of node maturity.

Jawed

 

[CarstenS]

Earlier you said "To cut costs, it'd make sense, to first market the fully functional and salvage parts (i.e. xx50 and xx30)" so I'm asking you, where's the HD4730?
Jawed

Precisely my point.
ATI's design doesn't generate reams of salvage parts.
Jawed

And what exactly is your point? Does AMD have a 100% yield rate on RV740? Though I admit I cannot be sure about it, I honestly don't think so. And where do I state, that there'd be "reams of salvage parts"? Nowhere, right, kthxbye.

 

[Jawed]

And what exactly is your point? Does AMD have a 100% yield rate on RV740? Though I admit I cannot be sure about it, I honestly don't think so. And where do I state, that there'd be "reams of salvage parts"? Nowhere, right, kthxbye.

Think: why did it take so long for HD4830 to appear?

Your argument is that a salvage part should launch simultaneously with the new chip's un-salvaged variant(s). When has an ATI card launched in this manner? X1950Pro/X1650XT was the last time.

RV740, with all the troubles at 40nm, "should be a prime candidate for salvage parts". Yet there aren't any.

So your entire argument has zero basis. Just because <100% of chips arrive defect free doesn't mean that salvage parts can launch simultaneously. It's NVidia's strategy to launch with salvage parts, not AMD's.

Jawed

 

[CarstenS]

My argument is, this time around it'd make more sense than ever to launch with the standard edition (not the top-of-the-line-modell for this particular single-chip) and a salvage-part.

Reasons being as follows:
#1 AMD needs money and they are in a good position to make it, because...
#2 Nvidia (quite possibly) has zero to counter this for at least x-mas-season (in OEM terms) - so basically not competition for "DX11-stickered" boxes from that point of view
#3 Their DX11-chip "Cypress" quite possibly is a little bit bigger than RV740, increasing the chance of defects not catchable with their/any chosen redundancy scheme, thus reducing yields.
#4 With RV740, a salvage part would probably have crippled sales of RV730 - and it would have to be in the same price range. This is not the case with their DX11 chips.
#5 They can keep the competition under pressure for a few weeks/months longer, maybe even improve on the final speed of the top single-gpu offering through improvements in manufacturing or avaiability of faster GDDR5-RAM.
#6 By fusing salvage parts, they can give their now-fastest product an advantage, that cannot be compensated as easily for with overclocking, thus increasing the value and the urge with early adopters to actually buy the fastest model - maybe even sparing the salvage parts for chosen OEMs for the time being, if those salvaged parts should really be in short supply.


p.s.: additionally: RV740s were in short supply for quite some time now at all. Personally, I don't know the reason and issues caused by TSMCs 40nm-problems. Could the chips not hit the required frequencies or were they not usable at all? That's IMHO a major point in evaluating why there are no salvage RV740s.

 

[Jawed]

You continue to miss the key fact, volume. If you don't have enough chips you can't create a SKU - salvage parts need to be available in higher quantity than the normal chips to be worth selling (because demand will be higher for the cheaper SKUs). That's why HD4830 took a while to arrive, despite the fact that HD4850 had higher core clock than was originally planned for ("Dave saved us").

I certainly won't deny the possibility that AMD will change strategy. Plenty of prior times they've had a chance to do something similar, but haven't. HD3830 was rumoured in-coming before HD3870 launched:

http://prohardver.hu/tema/ati_rv670/hsz_1-50.html

Didn't arrive for ages. AMD even thought it had the jump on NVidia (9600GT), i.e. thought it had a Christmas cracker - and therefore would have had even more reason to saturate the market with HD3830.

The chips simply aren't designed to produce high quantities of salvage parts. There's rumoured to be 4 different chips in AMD's initial D3D11 line-up, though it appears only 2 of them will launch initially. That line-up of 4 chips, compressed under $300 or so, is clear indication that the architecture isn't salvage-centric.

Jawed

 

[CarstenS]

You continue to miss the key fact, volume.

(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.

 

[Jawed]

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.

The disagreement we're having is over timing. AMD eventually releases salvage parts for some chips. I'm not proposing that AMD doesn't sell salvage SKUs, I'm pointing out the history of salvage SKUs, which is tied very directly to the architecture.

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.

Of course it would be better selling salvage parts than trashing them - the key question is timing.

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."

OK, so let's say HD5830 is a salvage part. AMD has deals with OEMs for HD5670s and HD5650s based upon Juniper and deals for HD5850 and HD5870 based upon Cypress.

Intially AMD can only deliver large quantities of HD5850 and HD5650, with some HD5870s and some HD5670s. HD5830 performs ~same as HD5670.

This is sort of like HD4830 and HD4770 performing the same. HD4770 was, in theory, supposed to arrive some time around November-January, in theory. HD4830 took its place and launched 1 month earlier (at least).

Why didn't AMD sell HD4830 from June, with HD4770 scheduled to take its place in November? It's all very well saying don't trash them, sell them, but the fact is AMD's had ample windows to launch a salvage part simultaneously in the past but simply hasn't done so.

With OEM deals the landscape might be radically better this year than one year ago. Maybe that puts the pressure on AMD to have something ready, e.g. HD5830, because HD5670 is binning badly. All I'm saying is that due to relative volumes (full:salvage) this hasn't been feasible before. Maybe volumes overall will have increased to such an extent, with AMD in ascendency, that a month or two of salvage GPUs is feasible from launch.

Regardless AMD still has an architecture that is not designed for volume of salvage parts - it uses closely-spaced chips to provide the variety of SKUs, designed to all launch within a short window. So anything AMD does with HD5830 would be bandaging a serious failure in Juniper (I certainly won't argue that can't happen).

You can argue the same with HD5630, instead of HD5570. There could be a couple of months' lag before HD5570 arrives (I think that's what we're expecting, actually), so salvage Juniper can take the load.

What's the nature of an OEM's contract? Is substitution of 70s with 30s from the next chip up acceptable? Is one to two months of erratic supply and not knowing which board you're getting acceptable under the contract?

There could be salvage RV740s out there, but there aren't. Why, because there's not enough of them, for long enough, to cover the costs (engineering validation, contract agreement, marketing, packaging) involved in this variant SKU. Has AMD (not ATI) ever sold a 128-bit chip salvaged as 64-bit? Is it technically reasonable/feasible?

It's quite likely that RV740 is no longer in manufacturing, since 40nm production capacity is limited and Juniper/Cypress should have been in mass production since July, if not earlier.

The overwhelming factor is that AMD's had plenty of opportunities to practice what you preach, in varying degrees of intensity, but simply hasn't done so.

---

Some guesstimates, per wafer, 40nm still immature:

            Juniper    Cypress
area          181        320
chips         330        180
full yield    198         90
HDxx50        149         68
HDxx70         49         22
HDxx30         33         36

and bear in mind that Juniper will have 3x or 4x more wafers in production.

Jawed

 

[Vincent]

$349 high for a new high-end card!? Vincent must be new to the world.

I would rather spend £ 200 in my "Ultra fast" SAS Hard Disk.( 15000 RPM/Guarantee speed and quality.)

I just spent £90K in my new car. ( Porsche Panamera )

 

[digitalwanderer]

You should have gotten a BMW.

 

[Gubbi]

You continue to miss the key fact, volume. If you don't have enough chips you can't create a SKU - salvage parts need to be available in higher quantity than the normal chips to be worth selling (because demand will be higher for the cheaper SKUs).

If demand is that high, supply and demand mechanics could increase the price of salvaged parts, - to the point where AMD might purposely cripple good parts to sell in the salvaged product bracket.

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.

Cheers

 

[Bouncing Zabaglione Bros.]

If demand is that high, supply and demand mechanics could increase the price of salvaged parts, - to the point where AMD might purposely cripple good parts to sell in the salvaged product bracket.

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.

Cheers

Yeah, the SKU is about addressing a market segment. As has been shown in the past, it's better to use your good chips to fill that SKU and address that market segment rather than wait for the few salvaged/binned chips that can never fill the demand.

If you don't address that market segment because you'd rather wait to sell all your good chips at top dollar, you concede all those sales to your competitors. Better to make lower profit than no profit.