AMD: R8xx Speculation

[Ailuros]

Assuming that theres a higher end SKU, then assuming that it won't be the one used for their high end model, isn't that an assumption upon an assumption?

Even knowing that a dual chip version is forthcoming, we still don't know how that would fit into their naming scheme. Is a 1600SP dual chip the 70 model or is it the 70 X2 model? Even if its the X2 model, does that mean that there may also be an X4 model? Its difficult to piece together what will actually happen here.

It's an assumption based on the following:

RV830 800 SP
RV840 1200 SP
RV870 RV830X2 native dual-core 1600SP
R800 RV870X2 3200SP

My best guess is this, they will try to get the most mileage out of a single chip as they can. So RV840 means single chip, RV870 is double RV840 and R700 is quadrupel or something to that effect. RV840 is the 170-80mm^2 version. So the 5850 is single chip, 5870 is double and what we would call the 5870 X2 is quad RV840. RV 830 is probably going to be a mainstream 56xx part IMO.

What do you get if you multiply 170-180mm2 by 4? Your scenario would mean that ATI's perf/mm2 advantage is completely gone which I severely doubt at this point is the case.

However I remember a tweet from that Catalyst guy and he implied that the current naming scheme is going to be changed, I guess if they are going to go multigpu/same die/module they really do need to distinguish the difference.

I don't see how that would change the above speculative number crunching.

 

[seahawk]

The most scary thing I see is the conflciting interests that must exist between the "single core" version of RV830 and the "dual core" version. The single core version obviously is a a mainstrem chip, so you would want to make it as small as possible, which means less texturing power, less cahce, less die size. However als this would badly hurt the dual version.

 

[Squilliam]

It's an assumption based on the following:


RV830 800 SP
RV840 1200 SP
RV870 RV830X2 native dual-core 1600SP
R800 RV870X2 3200SP


What do you get if you multiply 170-180mm2 by 4? Your scenario would mean that ATI's perf/mm2 advantage is completely gone which I severely doubt at this point is the case.

I don't see how that would change the above speculative number crunching.

The problem I see with that 1200 SP part is that it looks like an orphan. Why use one chip for 3 SKUs essentially and then develop a second and only use it for the one? Infact by creating so many SKUs from the one chip it would give room for a fourth salvage SKU in creating selling so many identical chips. That SKU could easily be 800SP.

The economies of scale in producing fewer models of chips more than mitigate for the perf mm^2 losses and the high end of the market really doesn't care how many chips the card has so long as it performs well and even more that it performs the absolute best. The perf per mm^2 is something that only chip makers ought to concern themselves with. This time around the GDDR5 memory they will be using will be a far lower cost component of the overall price of the cards so they can afford to use a larger chip at the same pricing levels. Furthermore they may not release an X4 version until next year after they shrink it down once again to 28/32nm.

 

[Vincent]

People should not expect the unrealistic performance gain on relatively low transistor budget to High end counterpart from NVIDIA.

 

[3dilettante]

ATI's current high end solutions are dual-GPU, MCPs. An MCP could in theory host 3 or 4 GPUs (just like what IBM did with the POWER4).

MCP would be one step further than AMD has gone so far.
Unless I missed something on the latest boards, it's two separate packages on the PCB.

I just don't see how you'd call it a 'dual core'. That's about the most retarded thing in the world and even more confusing.

Many of those writers who have used the "dual core" or "native dual core" moniker haven't been known to have the most sophisticated usage of terminology.


In my eyes, the only real way that such a term would make sense would be if one made a similar distinction to the one AMD made between Athlon64 X2 and Pentium D.
A GPU would be "native dual core" if the design included hardware with the dedicated purpose of on-die communication between two otherwise complete computing engines, not through a repurposed external bus.

Not that the Nth iteration of the dual-core rumor is enough to indicate that it's finally come true.

 

[gongo]

If a game is written for RV800 DX11 hardware, will it run better on RV700 DX10.1 than plain DX10 path?

 

[Humus]

Even to this very day, it's often to have encounters with people that still confuse the term "dual-core" as a particular product brand name (mostly Intel's)!
A silly everyday questions like: "What's the difference between Intel Dual Core and Core2 Duo processors?"... OMG!

Which of course is the reason Intel went with this stupid name of the product in the first place.

 

[Novum]

If the game implements a D3D10.1 fall back, that features some of the performance improvements from the D3D11 engine and not only a D3D10.0 fall back, then yes, it would run faster.

 

[I.S.T.]

Even to this very day, it's often to have encounters with people that still confuse the term "dual-core" as a particular product brand name (mostly Intel's)!
A silly everyday questions like: "What's the difference between Intel Dual Core and Core2 Duo processors?"... OMG!

Actually, Intel has a line of CPUs(Based off of the Core 2 Duos. I have one, they're nice little budget CPUs) out that have the words Dual-Core in their name. That is probably a big source of the confusion...

 

[Ailuros]

MCP would be one step further than AMD has gone so far.
Unless I missed something on the latest boards, it's two separate packages on the PCB.

Correct.

Many of those writers who have used the "dual core" or "native dual core" moniker haven't been known to have the most sophisticated usage of terminology.

I'm personally keen to use any reasonable alternative terminology (presupposition that it actually exists in the Evergreen family).

In my eyes, the only real way that such a term would make sense would be if one made a similar distinction to the one AMD made between Athlon64 X2 and Pentium D.
A GPU would be "native dual core" if the design included hardware with the dedicated purpose of on-die communication between two otherwise complete computing engines, not through a repurposed external bus.

Not that the Nth iteration of the dual-core rumor is enough to indicate that it's finally come true.

Then PowerVR's SGX543MP would be by that definition a native multi-core (as cores can be scaled in odd or even amounts from 2 to 16).

Squilliam,

The problem I see with that 1200 SP part is that it looks like an orphan. Why use one chip for 3 SKUs essentially and then develop a second and only use it for the one? Infact by creating so many SKUs from the one chip it would give room for a fourth salvage SKU in creating selling so many identical chips. That SKU could easily be 800SP.

Exactly; as I said already it's one part of the story that simply doesn't fit into the entire puzzle. Once more when you spent resources to go from a 800 to a 1200SP chip why not go one step further to 1600 instead of theoretically using 2*"RV830". It doesn't sound to me as a layman as it would take less overall R&D or be cheaper to manufacture.

The economies of scale in producing fewer models of chips more than mitigate for the perf mm^2 losses and the high end of the market really doesn't care how many chips the card has so long as it performs well and even more that it performs the absolute best.
The perf per mm^2 is something that only chip makers ought to concern themselves with.

Of course is it a primary concern for chip makers but if it should take over 700mm2 to beat a <500mm2 single chip high end solution, whereby in the most recent past something similar could be achieved with 2*263 vs. 583mm2 then any former advantage for AMD seems to vanish. Further to that power consumption could be an additional headache too, unless of course they've used this time a much lower transistor density per sqmm.

This time around the GDDR5 memory they will be using will be a far lower cost component of the overall price of the cards so they can afford to use a larger chip at the same pricing levels. Furthermore they may not release an X4 version until next year after they shrink it down once again to 28/32nm.

Ram on GPUs might get cheaper over time but the amount of onboard ram also increases consistently over generations.

Now if they won't release a X2 and if the rumored 1600SPs are true for "RV870" then they might not have a high end solution at all then. Smaller processes can be used by the competition too and if NV managed to get 2*GT200b chips on one SKU it doesn't sound impossible to go that route beyond 40nm too.

 

[dkanter]

MCP would be one step further than AMD has gone so far.
Unless I missed something on the latest boards, it's two separate packages on the PCB.

Many of those writers who have used the "dual core" or "native dual core" moniker haven't been known to have the most sophisticated usage of terminology.

In my eyes, the only real way that such a term would make sense would be if one made a similar distinction to the one AMD made between Athlon64 X2 and Pentium D.
A GPU would be "native dual core" if the design included hardware with the dedicated purpose of on-die communication between two otherwise complete computing engines, not through a repurposed external bus.

Not that the Nth iteration of the dual-core rumor is enough to indicate that it's finally come true.

I think the term 'native' is idiotic (no offense). I think for ICs the real distinction is monolithic integration (or not, with alternatives being integration at the package level, system level, etc. etc.)...and GPUs are a type of IC.

It really matters whether it's a single piece of silicon and it's certainly one reason the earlier Pentium D's sucked.

You're right that AMD is really doing two GPUs on a PCB, not an MCP.

DK

 

[3dilettante]

I think the term 'native' is idiotic (no offense). I think for ICs the real distinction is monolithic integration (or not, with alternatives being integration at the package level, system level, etc. etc.)...and GPUs are a type of IC.

It really matters whether it's a single piece of silicon and it's certainly one reason the earlier Pentium D's sucked.

"Native" is a distinction harped on predominantly by (AMD for the most part?) marketing, but there's a small distinction at the root of it that is worth noting.

The Smithfield Pentium D was one piece of silicon. Each core hung off of the SMP-capable bus through package-level wiring, but the die itself was monolithic.
I see that there was some mention of a possible additional connection between the two cores, but I haven't seen any elaboration on that account.

 

[dkanter]

"Native" is a distinction harped on predominantly by (AMD for the most part?) marketing, but there's a small distinction at the root of it that is worth noting.

The Smithfield Pentium D was one piece of silicon. Each core hung off of the SMP-capable bus through package-level wiring, but the die itself was monolithic.
I see that there was some mention of a possible additional connection between the two cores, but I haven't seen any elaboration on that account.

That's definitely worth noting. Smithfield was definitely ugly and something I'm sure everyone would rather forget. I suspect it is generally an exception and a situation that won't occur again.

I don't like the term native because it is a marketing one that isn't widely used outside AMD (or consistently within AMD - let's see what they call Magny-Cours!) and it also doesn't convey what we would agree is the real distinction (data sharing and communication latency).

Clovertown/Harpertown is as much a quad-core as Barcelona or Shanghai or Nehalem. However, it is less integrated (2 dice) and the communication is much higher latency, so the cost of sharing data is higher.

There are other differences - it's harder to drive a multi-drop bus, etc., but they aren't quite as profound.

Anyway, this is really all a tangent. Popping up one level, GPUs are multi-core processors with 10-30 cores. Multiple GPUs can work together in separate PCI-E slots on the same motherboard, or on the same PCB, etc. etc.

The different levels of integration are tricky to make an analogy against a CPU, since the needs are different. GPUs don't have coherency, so their data sharing is of a different variety, and latency is far less essential.

As you pointed out, it's not MCP, primarily because what multiple GPUs really need is more space to cool better and deliver power cleanly.

Anyway, a perhaps separate question is whether having two dice in a single board is really advantageous in any fashion (programming, performance, etc.) over using two different cards.

 

[Squilliam]

Squilliam, Exactly; as I said already it's one part of the story that simply doesn't fit into the entire puzzle. Once more when you spent resources to go from a 800 to a 1200SP chip why not go one step further to 1600 instead of theoretically using 2*"RV830". It doesn't sound to me as a layman as it would take less overall R&D or be cheaper to manufacture.

The thing is im not entirely convinced that the 180mm^2 chip will have 1200SP. Doing the calculations based off the 256mm^2 RV770 given perfect scaling we see ~130mmm^2, but as scaling is never perfect and we have to consider the extra transistors for DX11 it seems like a high end SKU is more likely to have roughly 1000 SP along with a few extra tweaks for interdie communication. This certainly does yield something which makes their comments about 'faster than a GTX 285' the truth.

As for the lower end SKU, I can see them sticking with ~640 SP for their RV830 line of chips which makes sense for improved transistor density and the transition to DX11 for their lower power mobile and 56xx range of chips. Power efficiency is more important here than absolute performance anyway.

So RV840 = performance line, RV830 = power efficient, mid range and laptop parts which is the way I see it.

Of course is it a primary concern for chip makers but if it should take over 700mm2 to beat a <500mm2 single chip high end solution, whereby in the most recent past something similar could be achieved with 2*263 vs. 583mm2 then any former advantage for AMD seems to vanish. Further to that power consumption could be an additional headache too, unless of course they've used this time a much lower transistor density per sqmm.

The advantage for AMD is the better yields and economies in producing a smaller range of chips. They can lose in the mm^2 war but make it up in volume and yields. Power consumption is always a headache, but until we see the actual parts we cannot comment on that really. Performance per watt is something thats already pretty good for AMD at this point, and these chips are limited by thermals and power use more than they are by die size.

Ram on GPUs might get cheaper over time but the amount of onboard ram also increases consistently over generations.

But the difference is this. The premium with using GDDR5 is vanishing quickly, that means that it will soon cost little more to use GDDR5 than it would to use GDDR3. Note how the difference in price between the 4870 and 4850 has shrunk quickly in both relative and absolute terms.

Now if they won't release a X2 and if the rumored 1600SPs are true for "RV870" then they might not have a high end solution at all then. Smaller processes can be used by the competition too and if NV managed to get 2*GT200b chips on one SKU it doesn't sound impossible to go that route beyond 40nm too.

That is true, Nvidia could go for a dual chip SKU, but it doesn't really make sense unless their new top end GPU is delayed until next year. Theres no point in using two chips when one will do it better.

 

[LunchBox]

Well the 1200 SP is not that weird if they were only going to use it for mainstream IMHO, I mean just take a look at RV740. I'm pretty sure they'll have other options for the Graphic card manufacturers if they decide to use it in a different configuration... like to what they are doing to their current products.

 

[aaronspink]

Which of course is the reason Intel went with this stupid name of the product in the first place.

Actually Core2 Duo/Core Duo/etc I thought was the most succinct product naming by AMD or Intel in a long long time. OMG a name that actually says exactly what you are getting!

 

[aaronspink]

Exactly; as I said already it's one part of the story that simply doesn't fit into the entire puzzle. Once more when you spent resources to go from a 800 to a 1200SP chip why not go one step further to 1600 instead of theoretically using 2*"RV830". It doesn't sound to me as a layman as it would take less overall R&D or be cheaper to manufacture.

Probably because it would require another backend flow. Another set of masks. Another package design. another bring up. Another set of DFM tests. Its not like, OOh we have this, stamp stamp. Ohh now we have something that is 2x this.

Then there is the additional inventory, demand forecasting, etc.

Of course is it a primary concern for chip makers but if it should take over 700mm2 to beat a <500mm2 single chip high end solution, whereby in the most recent past something similar could be achieved with 2*263 vs. 583mm2 then any former advantage for AMD seems to vanish. Further to that power consumption could be an additional headache too, unless of course they've used this time a much lower transistor density per sqmm.

Comparing additive Sq mm isn't accurate. 4x 180 is likely cheaper to produce and test than 1x600. If they can be performance competitive with 4x180 then they likely come out ahead. The can also likely get better power performance out of 4x180 than 1x600 due to inter-die variation.

 

[seahawk]

800SP for a RV740 replacement does seem possible. 800ALUs 32 TMUs 16 ROPs - why not?

The other could be 1200 SP 48 TMUs 16 ROPs with 256Bit memory interface.

 

[no-X]

It could also be based on different configuration (16 ROPs vs. 32 etc.). Something like RV530: 12 ALUs, R520 16 ALUs...

 

[CarstenS]

800SP for a RV740 replacement does seem possible. 800ALUs 32 TMUs 16 ROPs - why not?

Because you'd make it even more bandwidth limited than RV740 with it's near linear scaling with added bandwidth already is.