AMD: R8xx Speculation

[OpenGL guy]

So Early-Z is querying RBE-Z before the shader starts processing the pixel? Yuck, long long long latency.

No, it doesn't work that way. Early-Z means do the full RBE Z before the pixel shader. This way you reduce shader workload. Late-Z means do the full RBE Z after the shader. These are independent of HiZ and have been around since at least R300.

Agreed, I'm struggling to discern when hierarchical-Z would be best turned off, unless the current render state has no Z at all.

The only time you'd need to disable HiZ is if the pixel shader were outputting Z.

 

[elsence]

You do realize that the 5870 lead increase isn't so much from the fact that it isn't bandwidth limited

I said the opposite, i said that probably the 5870 is bandwidth limited...

but rather it simply isn't nearly as limited as its competition is.

So let's take the 5870 vs GTX285 data.
Is the 5870 less bandwidth limited than GTX285?
I don't know.

For example:
Let's compare GTX285 with GTS250 (much more similar than your comparison ATI vs NV architecture...)

GTX285 has +75% the pixel fillrate and only +10% the texel fillrate of GTS250 and +125% the memory bandwidth.

What i can tell for sure is that NV made a terrible design with GTX285, lol

But let's say that your assumption is correct, i don't want to argue about this at all.

This isn't the major role imo, the major role is that the 850MHz 32ROPs/1600SP ATI design has better specs than the GTX285.

For example:

Take a GTS250 (738MHz core / 1100MHz mem) and underclock the mem at 900MHz.
Take a low power 55nm 9600GT SKU (600MHz core / 900MHz mem)
You can clearly see that the underclocked GTS250 is more bandwidth limited than the 9600GT.
Despite that it is increasing its lead...

That and the FACT that ATI has some nice 4x and 8x AA algerithims in place right now to so reduce the hit for those settings.

Certainly ATI's antialiasing algoriths brought so far less hit than NV's, especially in the 8X case.
That's why in my analysis of the AMD data, i suggested not to include the 8X data, because it can lead to wrong conclusions.
I don't care if the 5870 is more bandwidth limited in relation with GTX285.
I care if the 5870 is bandwidth limited by itself.
In the sense: if we increase the mem speed from 1,2GHz to 1,5GHz, if we gonna see a significant perf. increase...

 

[elsence]

At only 4.8Gbps.

Yep, it is natural at this early stage of GDDR5.
This happened in the past also, for example 4870 had 4Gbps Qimonda ICs and ATI clocked them at 900MHz (instead of 1GHz)
I was talking about what ICs is logical for ATI to buy from companies like Samsung and Hynix at this stage.

I'd like to see a comparison of 1GB HD4850 and 1GB HD4870 to be sure about that, now...

Jawed

I search the web, i found some reviews:

http://www.bit-tech.net/hardware/graphics/2008/11/28/gigabyte-gv-r485mc-1gh-radeon-hd-4850-1gb/1

http://www.hardwarecanucks.com/foru...gabyte-radeon-hd-4850-1gb-passive-review.html

http://www.bjorn3d.com/read.php?cID=1527&pageID=6566

It seems to me that +33% (4/3) is a fair average at 1920X1200 4X AA 16X AF.

 

[Silent_Buddha]

Well according to Davros 5750 costs ~150, same as 4870. 4850 is only ~100. That's 50% more - sure it will be faster but it doesn't sound like a simple replacement to me (of course, launch price of 4850 was higher too but it means the 4850 could still be a viable option and AMD doesn't have any replacement part at 100 yet.).

Well, there is still that rumor that ATI will be launching 4 chips. So far we've only gotten confirmation (somewhat) of two of them. Rv870 and Rv840(?). So there's still possibly two more that haven't been leaked.

Regards,
SB

 

[Alexko]

Well, there is still that rumor that ATI will be launching 4 chips. So far we've only gotten confirmation (somewhat) of two of them. Rv870 and Rv840(?). So there's still possibly two more that haven't been leaked.

Regards,
SB

Sure they have:

 

[Silent_Buddha]

Ah, right-o, so the other two will be released in Q1 2010.

Regards,
SB

 

[mczak]

Sure they have:

Well so redwood/cedar only launch next year. So no 4850 replacement for now neither (well maybe by that time 5750 will be used at the 100$ spot...).
Plus I sort of doubt redwood will achieve HD 4850 performance even in a high-end configuration.

 

[FrameBuffer]

Well so redwood/cedar only launch next year. So no 4850 replacement for now neither (well maybe by that time 5750 will be used at the 100$ spot...)

iirc the 5850 is the 4850 segment replacement. Or are you referring to the current pricing ? That will for the most part rely upon the competition. 48XX series last I heard will go into EOL status after launch. Rv840 series could see dayligt earlier however there is little incentive until the channel is cleared some.

 

[fellix]

By the way, looking at the test drivers' CCC for 5870, I can't find any different settings regarding changes or additions to the available AA modes.

 

[Acert93]

I have followed this thread off and on but have not been able to keep up. Any word on multi-chip designs on a single package, Crossfire enchancements, etc? Memory footprint utilization, latency, geometry, etc are all interesting issues for SLI like approaches and I am curious to know if AMD did much to make Crossfire more efficient/better results in these areas. Getting multiple chips working as a simple chip is a big task, curious to see if any unfounded rumors had any merit. Thanks!

 

[3dilettante]

I have followed this thread off and on but have not been able to keep up. Any word on multi-chip designs on a single package, Crossfire enchancements, etc?

No word on any of that.
There's no shot of what's under the cooler for the alleged X2 board, it just appears to be longer.

The diagrams make no mention of a crossfire sideport, though there is a crossfireX compositor in a similar area of the diagram.
The compositor still hangs off of the hub, and barring an improvement there, there would be no point for a super high-bandwidth link between the chips.
There isn't much else given other than the box on the diagram.

 

[Alexko]

Well so redwood/cedar only launch next year. So no 4850 replacement for now neither (well maybe by that time 5750 will be used at the 100$ spot...).
Plus I sort of doubt redwood will achieve HD 4850 performance even in a high-end configuration.

Well there are rumors saying that Redwood has 640 SPs, but I doubt that: it's too close to Juniper (800). My bet's on 480 SPs in which case, indeed, Redwood can't achieve 4850-level performance.

But yeah, from what we're hearing, Juniper-based cards sure seem a little too expensive. They have low power consumption, Eyefinity and DX11 on their side, but I don't think that will be enough to woo people away from the cheaper 4800s.

 

[LordEC911]

Well there are rumors saying that Redwood has 640 SPs, but I doubt that: it's too close to Juniper (800). My bet's on 480 SPs in which case, indeed, Redwood can't achieve 4850-level performance.

But yeah, from what we're hearing, Juniper-based cards sure seem a little too expensive. They have low power consumption, Eyefinity and DX11 on their side, but I don't think that will be enough to woo people away from the cheaper 4800s.

Have we gotten any confirmation on that SP count?
I'm still leaning towards 960SPs, which was part of the reason why some thought RV790 was going to have the same setup, plus it will decrease the gap between Juniper XT and Cypress Pro.

 

[Silent_Buddha]

Hmm, that new cooler actually doesn't look half bad once you get the AIB logo's and graphics on them...

http://gathering.tweakers.net/forum/view_message/32598832

They need to switch the girls though, the one on the 5850 is much tougher looking than the 5870 girl. Dual Swords FTW!@#!@#

And pretty amazed at how cool it stays even at load at 30% fanspeed with that tiny exhaust opening.

Regards,
SB

 

[trinibwoy]

Yeah you're right, that's a big improvement. They could still do with some help from Apple's design team though.

 

[Chris123234]

Here:
http://www.brightsideofnews.com/new...ing-revealed2c-end-for-hd4800-draws-near.aspx
Salt truck backing to unload: meeep, meeep, meeep

I spazzed out for a second at this typo:

with 5870 being priced at 199.

______________________________________________

Thats what i found on xtremesystems

The 5870 Six is the most beautiful(ly clean faceplate) card I've ever seen. What is the premium on this one? And does the X2 come with a Six (or Twelve) version?

 

[turtle]

Have we gotten any confirmation on that SP count?
I'm still leaning towards 960SPs, which was part of the reason why some thought RV790 was going to have the same setup, plus it will decrease the gap between Juniper XT and Cypress Pro.

Really curious about this myself as of late.

960 really does seem to make sense, business-wise, even if it is against the conventional [strike]wisdom[/strike] rumors.

I could write my typical page of bullshit why either would make sense (regarding timing of release, die size, price, competition, etc) but I'll refrain...just this once.

 

[3dcgi]

Wouldn't allocating triangles to a bin require the rasterization portion of the workload as well?

You only need to setup the verts to get their screen space position. You don't need to finish rasterization by determining pixel coverage.

 

[3dilettante]

You only need to setup the verts to get their screen space position. You don't need to finish rasterization by determining pixel coverage.

My interpretation of Larrabee's software rasterizer is that it does get coverage during bin setup.
The bins will contain coordinates and coverage info for the triangles.
Perhaps I'm not interpreting this correctly:

There are two phases to the processing. In the front-end, each
PrimSet is given a sequence ID to identify where in the rendering
stream it was submitted. This is used by the back-end to ensure
correct ordering, as discussed below. The PrimSet is then assigned
to a single core, which performs vertex shading, tessellation,
geometry shading, culling and clipping to produce triangles (or
other primitives). The core then rasterizes each triangle to
determine which tiles it touches and which samples it covers
within each of those tiles. The result is a series of X,Y coordinates
and sample coverage masks for each triangle. This data is stored
in the bins along with indices that reference the vertex data.

 

[3dcgi]

My interpretation of Larrabee's software rasterizer is that it does get coverage during bin setup.
The bins will contain coordinates and coverage info for the triangles.
Perhaps I'm not interpreting this correctly:

The quote does seem to indicate that though it seems wasteful for large triangles. I'd think you want to bin before amplification (scan conversion).