The Official RV630/RV610 Rumours & Speculation Thread

[Arun]

Rumoured Data Points
- Both RV630 and RV610 will be manufactured on TSMC's 65nm process.
- There is no really reliable information on die size and/or transistor counts.
- RV610 is expected to be an extremely small chip that doesn't dissipate much heat.
- Some reports claim these chips correspond to the Radeon X2300 and X2600 Series.
- Other reports (mostly one by Theo Valich) claim they are the Radeon X2400 and X2200 Series.
- HKEPC claims up to 128W for RV630; various other sources disagree with that information.
- VR-Zone claims RV630 has a 128-bit memory bus, and RV610 a 64-bit memory bus.

Extra Tidbits & Fun Facts
- Where's RV550, and does it have a Radeon X2xxx Series brand name?
- There have been some rumblings of 65nm yields not being too brilliant yet. It's hard to say how much this will affect chips the size of RV630 and RV610 though, which also have some possible redundancy.

Noteworthy Internet Rumours (from newest to oldest)
"ATI's RV610 goes AGP" [Fudzilla]
"AMD中階繪圖核心RV630 支援PCI-E 5GT/s及GDDR4" [HKEPC]
"AMD to overtake Nvidia in WHQL driver wars" [The Inquirer]
"DAAMIT's RV610 and RV630 are 65 nanometre parts" [The Inquirer]
"ATi 65nm RV610/RV630 Cards Info" [VR-Zone]
"ATI's RV610 and RV630 details pop into view" [The Inquirer]
"ATi RV610 & RV630 Info Emerged" [VR-Zone]
"ATI RV610 and R630 to launch in April" [The Inquirer]
"AMD GPUs on 65nm, and NVIDIA's on 80nm" [Beyond3D]

Thread Discussion Starting Points
- Does all of this rumoured information seem reliable to you? What, if anything, sounds fishy?
- Are the rumoured products looking good or bad from your point of view?
- What extra facts would you expect to be true for RV630 and/or RV610?
- What do you think this means for the rest of the AMD Radeon line-up?

- No reliable information has been leaked on the number of ALUs/TMUs/ROPs/etc.
- So, what do you expect there? Do you think some architectural tweaks have been done?
- Do you believe the ratios of the different units have been changed? How so? How would it affect things?

Last Updated on March 5th

 

[Sound_Card]

Funny thing is that I keep hearing RV630 and R630. So is their one of the two? Or do they both exist? Most likely just a typo and someone forgot to put in a v.

My speculation on RV630:

x2600xt and x2600pro variants
65nm
256mb and 512mb GDDR3 editions
Anywhere from a 600mhz clock to 700mhz
256bit interface
Anywhere from 64 shaders to 96 to 128.
XT = 16 ROP's Pro = 12

 

[turtle]

Funny thing is that I keep hearing RV630 and R630. So is their one of the two? Or do they both exist? Most likely just a typo and someone forgot to put in a v.

My speculation on RV630:

x2600xt and x2600pro variants
65nm
256mb and 512mb GDDR3 editions
Anywhere from a 600mhz clock to 700mhz
256bit interface
Anywhere from 64 shaders to 96 to 128.
XT = 16 ROP's Pro = 12

I think it is obviously a typo.

My specualtion on RV630:

x2600xt, x2600pro, and "X2" or "dual" variants
65nm
256mb/512mb GDDR3 and GDDR4 editions
Clocked to attempt to be competitive with G84 parts
128-bit interface
XT = 80 PRO = 60 ALUs
XT = 16 PRO = 12 shaders
4-8 ROPs (depending on R600's structure)

Like I've guessed before, approx 1/4 spec of a R600. If I'm wrong, and it's 1/2 spec, I shall cream my pants with happiness. I still don't expect that until the Q3 parts launch though (speculation: 1/2 and 3/4 spec 'rv660' and 'rv670' on 65nm), which I expect to have the 256-bit bus.

IOW, in Q3 I expect:

RV660 speculation:
x2650XT and perhaps pro and/or "x2" or "dual" variants.
65nm
256/512MB GDDR4
unknown clocks
256-bit interface
XT = 160 PRO = 120 ALUs
XT = 32 PRO = 24 shaders
8-16 ROPs

RV670 speculation:
X2950PRO and "x2" or "dual" variants
65nm
512MB GDDR4
unknown clocks
256-bit or 512-bit interface
240 ALUS
48 Shaders
12-24 ROPs

 

[Sound_Card]

Overall I think performance is around a x1950pro or a xt. Slightly faster.

 

[turtle]

Overall I think performance is around a x1950pro or a xt. Slightly faster.

I agree. Somewhere in that vicinity.

 

[Arty]

Like I've guessed before, approx 1/4 spec of a R600. If I'm wrong, and it's 1/2 spec, I shall cream my pants with happiness. I still don't expect that until the Q3 parts launch though (speculation: 1/2 and 3/4 spec 'rv660' and 'rv670' on 65nm), which I expect to have the 256-bit bus.

Does RV560 even exist? If it does, then does 8800GTS feature something other than the G80 core?

I think ATI aborted RV560 in favor of a borked RV570 quite late in the game. If they have planned their next generation learning from their past mistakes, we wont see a RV660/RV670.

 

[Dave Baumann]

Does RV560 even exist?

Yes. You can buy it in X1650 XT's.

 

[Anon Lamer]

Overall I think performance is around a x1950pro or a xt. Slightly faster.

In shading only though. Theres more to fast gfx cards than just shading. I think performance will be around x1650xt or a bit better but below x1950pro for the 2600XT.

 

[Anon Lamer]

Yes. You can buy it in X1650 XT's.

I thought RV560 was a RV570 with 1/3 disabled and 128 bit bus. So its a separate construction with its own layout?

 

[Arty]

Yes. You can buy it in X1650 XT's.

Ofcourse. But I meant that in the sense is it designed from the ground up like RV530, which its not.

 

[Arun]

Ofcourse. But I meant that in the sense is it designed from the ground up like RV530, which its not.

I think the initial rumors were that RV560 was a chip manufactured at TSMC to meet capacity constraints. If TSMC's supply problems for 80nm turned out not to happen, though, I'd imagine they might have decided to cancel it. And of course, I don't know if this rumor was true...

 

[Sound_Card]

In shading only though. Theres more to fast gfx cards than just shading. I think performance will be around x1650xt or a bit better but below x1950pro for the 2600XT.

Hmmm, I don't think performance will be around a x1600 series card. I think ATi will up the ROP and TMU count on these cards comapared to last. Shader power will be a great deal better than a x1950pro IMHO.

We will just have to wait and see, but my money is on x2600xt being around 5 to 15% faster than the x1950pro across the board.

 

[vertex_shader]

I think it is obviously a typo.

My specualtion on RV630:

4-8 ROPs (depending on R600's structure)

Another x1600 series?

 

[vertex_shader]

I think ATi will up the ROP and TMU count on these cards comapared to last.

I hope :smile:

 

[trinibwoy]

Another x1600 series?

G73 only has 8 rops and did just fine so I'm not sure why you're blaming ROP count for X1600's mediocrity.

 

[Jawed]

G73 only has 8 rops and did just fine so I'm not sure why you're blaming ROP count for X1600's mediocrity.

X1600XT has 4 ROPs + 4 TMUs.

Jawed

 

[turtle]

X1600XT has 4 ROPs + 4 TMUs.

Jawed

Right, but obviously x1600xt was a decision they had to make, and I don't believe they will have to make again. The reasoning being obviously that R520->R580 changed the shader structure. They surely could have gone straight to RV560 for x1600 (making x1600 effectively x1300) or made a 8/8/8 part (1/2 R520 with 128-bit bus) but perhaps they felt it was too expensive to produce on 90nm for it's price-point, had been planned on 80nm all-along and the delay screwed up their plans, or something else, I have no idea. Obviously it was the wrong decision, but X1600 was a freak of nature. The refresh's low-end (1/4 R580) at the earlier-gen's mid-range...It just didn't make sense, because the change in shader structure was just not a significant enough jump in performance.

I personally believe R680 will be just a shrink of R600 with faster clocks and GDDR4, therefore the whole family will be derived from the initial part, like last generation was from R580 (x1300 excluded) and pretty much all generations before it, and every gen on Nvidia's side. The difference being of course, I expect the true mid-range/upper mid-range parts to be on schedule and not delayed months on end like RV560 and RV570 were. In essence, X2600 is the spiritual successor of x1600, or x1300 had R580 not existed (as X1300 was essentially 1/4 R520). RV660 would be the successor to the X1650, and RV670 to the X1900GT/x1950pro.

Like I said before, if RV660 does not exist, and therefore RV630 is 1/2 the shaders out the gate, I shall cream my pants with happiness. In reality though, I do not expect this to happen. Not only because 1/2 would with a 128-bit bus would seem insane next to a full part with a 512-bit bus, but more-so I expect RV610/RV630 to be relatively lack-luster compared to R600, 1/8 and 1/4, geared towards lower power and laptops as well keeping the die size down to reap profits.. They are the first 65nm GFX chips after-all, I don't expect them to be overly complex comparatively. That being said, 1/4 of a R600 may be able to take on G84 (which looks to be 1/2 G80, with a 128-bit bus).

Another reason to believe these parts are to stay the long haul, and not be refreshed in the process: They are 65nm, like R600 will eventually become with R680. I believe they will keep their niche, rather than be replaced. With the launch of R680 in late Q3 or sometime Q4, you could see a nice spectrum of products ranging from 1/8 (x2300) to 1/4 (X2600) to 1/2 (RV660) to 3/4 (RV670) to 1/1 (R680) all on 65nm. A much better spread in a more timely fashion than last gen.

As for the ROP count...We hear 16 for R600 an awful lot...Much more than I'd like. If that is true, then I expect 4 ROPs...Yes, icky, I know. If it's 32 ROPs, then I expect 8 ROPs...Personally, I think that sounds rather exciting and is what I'm hoping for.

 

[INKster]

I don't believe ROP count is an accurate measure of expected (and compared) performance anymore.
DirectX 10 and the new hardware architectures certainly have de-emphasized the raw number, in favor of expanded per-ROP functionality.

 

[turtle]

I don't believe ROP count is an accurate measure of expected (and compared) performance anymore.
DirectX 10 and the new hardware architectures certainly have de-emphasized the raw number, in favor of expanded per-ROP functionality.

I certainly agree, and not only for DX10. Xenos is also a prime example of performance with limited ROPs but expanded functionality per rop. I expect the R600 generation to improve on this, but any way you cut it, 16 ROPs sounds awful low for the amount shaders/ALUs this beast is likely to have.

If it is 320 MADDs (64*5/64x4+1 or 320 scaler) 16 just seems wildly disproportionate.

 

[Jawed]

There's the age-old problem of what the ROPs (and TMUs) can do. One R6xx ROP might be worth two R5xx ROPs, for example.

Also, there is a distortion in RV530 associated, I believe, with it's superfluous bandwidth. Note RV560 has ~the same bandwidth as RV530...

That distortion clouds an analysis of scaling, I reckon. e.g. RV530 is essentially 1/4 of R580, yet it has 35-45% the bandwidth (depending on GDDR3 or 4). RV560 has ~45% of the bandwidth of R580+, with ~1/2 of the pipeline capability - this almost perfectly lines up.

The R6xx range looks like it's going to scale the other way, comparing the extreme high-end 512-bit bus scaled down to RV630. e.g. I'm thinking RV630 has 1/4 the pipeline capability, but 1/5 the bandwidth. Against that, R6xx should be, if the patent documents are to be believed, more bandwidth-efficient.

It's a shame we don't have any, rumoured, handy pipeline configuration nomenclature to argue over, e.g. 16-2-20-4 ...

I suppose if we're expecting R600 to be more than 2x faster than R580, then RV630 should be ~2x faster than RV530, i.e. in RV560 territory. Maybe with more bandwidth, say ~29GB/s.

Jawed