R590 - Will we see this beast?

[trumphsiao]

Makes little sense to call it X1800GTO rather than X1900GTO if it's based on the R580 core. And where does Kristopher get X1800 Pro?

If R590 is indeed a mid-range part (R570 code name would have made more sense in this case) and R580 has to hold the high-end fort until October as the source suggests, ATI will be in for a long summer.

If my theory is right, I think there are plenty of unqualified R580 croes in the stockpile which are suited for 400~250 price range. What I dont quite follow is why ATI name unqualified R580(with a few quad disabled depends on each chip) as X1800GTO/XL . R580 has 4 dispatch processors and each have 4 quad PS . Disable one Dispatch Processor you loss 4 Quad and still you have 12 Quad . I dont think X1800GTO/Pro is suitable unless ATi doesnt have enough 12 Quad~8 Quad PS R580 .

 

[BrynS]

Hmm... $279 isn't that encroaching into RV560's likely price range? I wonder what kind of volume of X1800GTO inventory is available/being ramped, as a 3-quad, 36-shader 256MB part for $279 would surely fly off the shelves?

Also, that DailyTech piece is the first time I can recall a specific launch timeframe (October, 2006) for R600 @80nm being mentioned.

 

[Pete]

I thought R600 has been rumored to arrive with Vista, ~2H06, for some time now? As for a X1800GTO encroaching on RV560 territory, is GTO a temporary SKU reserved for leftover parts (R420s and now R520s), or is it a stable SKU?

 

[Unknown Soldier]

Yes but Vista has been delayed as reported by microsoft to Q406

 

[neliz]

I think the disabled feature will be Fetch$

but seriously.. what happens to RV560 then?

 

[Dave Baumann]

I think there is a clue in RV560's codename.

(Edit: Now, here's a thought - does V denote chip cost/expected pricing, or is it a bit clearer and actually denotes bus width?)

 

[Geo]

I think there is a clue in RV560's codename.

(Edit: Now, here's a thought - does V denote chip cost/expected pricing, or is it a bit clearer and actually denotes bus width?)

What? They've been using the V for Value for uhh, quite some time. You're saying it's actually a "5", for 512-bit? Or. . .or. . .it takes two fingers to make a V, and since 1/2 a V is the symbol for division, we have to divide by 2? Or. . or. . .ah hell, I have no idea what you're pointing at!

 

[Dave Baumann]

The explaination that we've have is that the "V" in the codename denotes "value", which correlates to price by some method or another. However, is that "value" term actually dicated by the bus width? Since ATI started using it has there been a 128-bit chip without the "V"? Has there been a 256-bit chip wih the "V"? Off the top of my head, I say no to both those cases.

 

[Geo]

http://www.beyond3d.com/misc/chipcomp/?view=chipdetails&id=68&orderby=release_date&order=Order&cname=

?

I am of course trying furiously to disprove this theory as I want RV560 to be 256-bit. Tho, frankly, looking at die-size, I don't think it can be. While I haven't seen a die size for Parhelia, R300 was 218mm2, which is 45% bigger than current RV530. Now, there is an assumption there that "not much smaller than that rules out 256-bit". . . This would be exacerbated if we assume RV560 is 80nm. . .

Edit: Actually, it'd be nice to get a die size for Parhelia added to the B3D tables if some kind soul can provide it. As the first consumer 256-bit, it'd be nice to have there for comparison to R300 --particularly if it turned out to be signficantly smaller. . . Actually, NV35 (NV's first 256-bit) as well doesn't have a die size in the table.

 

[Jawed]

Eric indicates that per-chip 32-bit addressing requires extra pins beyond those required by 64-bit addressing:

http://www.beyond3d.com/forum/showpost.php?p=694356&postcount=59

He guesstimates 30% extra pins - but that needs checking.

If so the "minimum die size" to support a 256-bit total bus width would be larger in R5xx.

If you view pre-R5xx as:

• X300
• X600
• X700
• X800

then it seems that RV560 is destined to take on the role of X700... (and match-up against 7600).

It seems that ATI is taking a big hit on die size (or margin, if you prefer) at each product category.

I wonder if that hit was predicated on GDDR4 being ready (hence significantly higher performance), because it's arguable that the GDDR4 compatibility built-in to R5xx is currently going to waste.

Jawed

 

[Geo]

I *think* what you just said is "sort of geo, but that 218mm2 is probably too small these days for ATI to do 256-bit too".

 

[Dave Baumann]

It seems that ATI is taking a big hit on die size (or margin, if you prefer) at each product category.

Flip-side of the same coin, but I would suggest they are leaning on process more.

And, yes, from what I've heard RV560 is being veiwed by ATI as the competition to the high end G73. I'm not sure if there isn't any reason why they couldn't have each MC address 64-bits though, so I'm not sure that need preclude 256-bit though (RV560 also has some changes from the rest of the R5xx line AFAIK).

Geo, yes, good point - that was pre-R300/Silicon Valley designs though, so we'll see...

 

[Jawed]

Yes.

---

But 256-bit (in the midrange) and 512-bit (at the high end) look irrelevant in the face of GDDR4.

Or at least they would do if R5xx was using GDDR4.

We could easily see the best midrange cards with ~37GB/s by the end of the year. And high end cards with a minimum of ~74GB/s reaching upto ~100GB/s.

RV560 with ~37GB/s (~X800XTPE) as compared with RV530 with ~22GB/s isn't spectacular (since RV560 should be 2x RV530) but it's still pretty tasty.

Jawed

 

[Geo]

We could easily see the best midrange cards with ~37GB/s by the end of the year. And high end cards with a minimum of ~74GB/s reaching upto ~100GB/s.

RV560 with ~37GB/s (~X800XTPE) as compared with RV530 with ~22GB/s isn't spectacular (since RV560 should be 2x RV530) but it's still pretty tasty.

That much, eh? I think I hadn't really internalized GDDR4 until right there.

 

[Jawed]

I wonder how much die-level redundancy there is in R5xx?

Since R5xx is seemingly dependent on much more memory (due to the increased number of fragments in flight) it could be argued that memory-block redundancy could help R5xx yields. Particularly in RV560 and R580.

This leads me on to thinking about general redundancy within R5xx. And perhaps relating it to the apparent redundancy in Xenos:

which seems to feature four shader arrays, with one dropped for redundancy to deliver three functioning arrays.

Perhaps in RV530 and R580, for example, there are four fragment shader quads per shader unit, with a commensurate set of four blocks of registers in the register file. But for the sake of redundancy, only three fragment shader quads are activated per shader unit.

So a fault in any single quad's fragment pipeline or register file would lead to that quad being switched off. Leaving the shader unit as still capable of 3:1 ALU:TEX.

Texture unit redundancy, per shader unit, would have to be handled in a more fine-grained fashion though...

Jawed

 

[Fodder]

Since ATI started using it has there been a 128-bit chip without the "V"? Has there been a 256-bit chip wih the "V"? Off the top of my head, I say no to both those cases.

Not quite what you're asking, but R200/8500 and RV200/7500 launched simultaneously and both had a 128-bit memory bus. Of course, there were other features missing, but it doesn't fit your hypothesis.

Edit: Ah I see your next post, yes pre-R300 again.

 

[caboosemoose]

The explaination that we've have is that the "V" in the codename denotes "value", which correlates to price by some method or another. However, is that "value" term actually dicated by the bus width? Since ATI started using it has there been a 128-bit chip without the "V"? Has there been a 256-bit chip wih the "V"? Off the top of my head, I say no to both those cases.

Personally I think the latter is a function of the former. In other words, all the chips which ATI considers to be value have been given the "V" nomenclature, and it just so happens that all the value chips have been limited to 128-bit on the memory bus side. Of course, recently, ATI hasn't made a GPU at the high end that wsn't entirely feature complete, which makes hair splitting over the meaning of "V" a little pointless. Unlike NVIDIA which produced the three quad NV42 with a 256-bit bus, ATI's one-rung-down-rung-down-from-the-high-end offerings, such as R430 on 110nm have been fully feature complete.

Granted, a 256-bit bus currently goes hand in hand with the high end, but I certainly don't think the "V" in RV560 procludes that chip from having a 256-bit.

 

[AlphaWolf]

I'm not sure you can use the past as an indicator with regards to the Value segment. The high end has expanded well beyond the $399 msrp, surely the value segment has expanded some also.

 

[Fodder]

surely the value segment has expanded some also.

Indeed it has, eg. we now have RV530 and RV560, where a couple of years ago we had only RV350.

 

[CJ]

This was just posted by the same source as always.

X1900XTX 11045 3dm (baseline performance) / X1900XT 10858 3dm
X1900XL ~9000 3dm (positioned against 7800GTX)
X1800GTO ~6600 3dm (positioned against G73 = 7600GS / GT)

RV570XT will be 10 / 12% faster than a X1800XT / 7800GTX
RV570XL will be 20% faster than a X1800XL / 7800GT
RV560 will be 35% faster than a 6800GS and 10% faster than a X1600XT
They will be released in August/September and all will be 80nm with onchip CF support and Gemini GPU support.

Ok, discuss.