R600 has 512-bit external memory bus

[sam0t]

Nice to get some confirmation in the middle of the rumour mill. Check it out:

Linky

 

[tisb0b]

sweeet I can't wait to see how it performs

 

[Wolfkin]

Hate to burst your bubble but it is still just a rumor, although very likely to be true.
Still hasn't been confirmed by AMD-ATI yet but I hope it will come with 512 bit.

 

[Anarchist4000]

With beyond3d saying it's true it's likely true. I'm sure they have inside sources that actually confirmed it before they posted anything on it.

 

[SX2233]

i hope the video card still has that lil red Ati sticker on it and comes in red colored pcb!

 

[Stereophile]

So what does this crazy memory bandwidth actually mean for gaming ?
Free or almost free anti aliasing ?

 

[kleox64]

So what does this crazy memory bandwidth actually mean for gaming ?
Free or almost free anti aliasing ?

free AA is only attainable through embeded memory, IMO however 2/4X AA will take very very little performance hit depending on what resolution.

 

[InorganicMatter]

Still not 100% confirmed. All AMD said was that R600 would have "new levels of memory bandwidth to the graphics subsystem, and bandwidth is critical to graphics performance." "New levels" could easily mean "new levels compared to R580," meaning 384-bit.

I stand by my claims that a 512-bit bus would be too expensive and have far too low of yields for AMD to waste their time with at this time.

 

[Stereophile]

I just reread that thread at B3D. While ATi might not have come right out and said we've got a 512bit memory bus, there seems to be little doubt among the writers at B3D it's got 512bit. This isn't coming from the inquirer. They quote 153 GB/sec compared to 86 GB/sec for a 8800GTX. That's nearly double !

 

[arthur_tuxedo]

Will all that bandwidth even be usable? It seems to me that if your bandwidth gets too far ahead of your core, then it's just going to waste.

 

[gostriker]

Will all that bandwidth even be usable? It seems to me that if your bandwidth gets too far ahead of your core, then it's just going to waste.

With my system and the current games - this new generation of gpu's is basically a waste. I assume Dx10 games will change this.

 

[Anarchist4000]

Keep in mind the transistor counts from last gen to this gen basically doubled(if not more) on both G80 and likely R600. So double the bandwidth and double the core as well. Plus i'm sure they will eat up some of the bandwidth with additional levels of AA.

I really wouldn't be surprised if G80 ended up about half the performance of R600. With the MUL not being effective for most shading, it only has the shading power of around a R580 to begin with. That's not to say G80 isn't still a good card if it's half the price.

 

[Stereophile]

I really wouldn't be surprised if G80 ended up about half the performance of R600. With the MUL not being effective for most shading, it only has the shading power of around a R580 to begin with. That's not to say G80 isn't still a good card if it's half the price.

Could you elaborate more ? What's MUL, a function ?

 

[Anarchist4000]

It's a math unit that does multiplication, hence MUL. It's sort of 1/3 of the shader power of the entire card. Could change with a driver update but at this point it's doubtful.

 

[Silverghost]

With a 384-bit bus, the supposed 1.2GHz GDDR4 will give the R600 115 GB/s of bandwidth. That's a lot! It fits well with the "New levels of memory bandwidth to the graphics subsystem" quote. 512-bit would be excessive and expensive. Unless they will be using slower ram speeds.

 

[mrmodman]

what exactly makes it expencive explain?

 

[Marvelous]

I thought Dave from Beyond3D worked for ATI...

 

[Talonz]

what exactly makes it expencive explain?

512-bits = lots of traces, lots more silicon area, which means more defective chips, so each one is more expensive.

 

[lithium726]

512-bits = lots of traces, lots more silicon area, which means more defective chips, so each one is more expensive.

has more to do with PCB complexity than die area.

 

[Talonz]

has more to do with PCB complexity than die area.

PCB is cheap, I think the hard part is getting timing issues right given the number of traces that need to be routed to the memory.

Disclaimer: Not based on any reputable source.

 

[Russell]

512-bits = lots of traces, lots more silicon area, which means more defective chips, so each one is more expensive.

Not so much considering that the current chips have a 512bit internal bus. Making the chip 512bit externally is unlikely to change the complexity by very much. I doubt they'd go so far as to make the chip 1024 internally.

 

[Brent_Justice]

Obviously the secret sauce for the G80 is GDDR4. I unforutnately do not know off the top of my head how high GDDR4 can scale, I'll have to research that. GDDR3 scaled very well and I assume GDDR4 has a lot of leg room, I have a feeling that we will see beyond 3 GHz eventually? But that's just a guess. On a 384-bit bus that would be 144 GB/sec of bandwidth at 3 GHz. Of course, on a 512-bit bus that would be 192 GB/sec.

I would love 1 GB video cards with over 200 GB/sec of mem bandwidth please

 

[lithium726]

not the PCB itself, getting enough layers and getting the traces dense enough going to the memory in order for it to work.

Its not that PCB is expensive, is that every little bit of it matters so much on high end graphics cards, and space on the board is limited.

 

[skeeder]

sounds like this could be a really big flop if not done right. Hopefully it will be done right, 512bit is insane and I don't think the average computer on this forum wouldn't be bottlenecked.

But besides Crysis which has been delayed until Q3 2007...what will all this power go towards that a simple 7900 can't run?

 

[BinarySynapse]

sounds like this could be a really big flop if not done right. Hopefully it will be done right, 512bit is insane and I don't think the average computer on this forum wouldn't be bottlenecked.

But besides Crysis which has been delayed until Q3 2007...what will all this power go towards that a simple 7900 can't run?

folding???


Anyone think maybe multiplexed address/data lines to the ram and possibly a quad pumped setup like P4's FSB would work well?

 

[EOS30D]

8800GTX　575MHz/1800MHz 384bit 768MB GDDR3　memory banwidth 86.4GB/s

(MADD(2flops) + MUL(1flop)) × 1,350 MHz × 128 SPs ＝ 518.4G Flops


R600(X2800XT?) 800MHz/2500MHz 512bit 1024MB GDDR4　memory banwidth 160GB/s

4way× (MADD(2flops) + MUL(1flop))× 800 MHz × 64 SPs ＝ 614.4GFlops

 

[rahavsmt]

^^^^

More competition is better for us consumer. Nvidia can introduce a refresh and knock down the prices for 8800 and ATI will do the same too.

 

[^eMpTy^]

Not so much considering that the current chips have a 512bit internal bus. Making the chip 512bit externally is unlikely to change the complexity by very much. I doubt they'd go so far as to make the chip 1024 internally.

He's right about the traces though...512bit isn't too hard on the chip, but the number of traces that have to run across the pcb will make it more expensive to produce...whether or not it will be prohibitively expensive or cause any problems is anyones guess...

 

[Obi_Kwiet]

I duno. Maybe they have something worked out with their board partners that will make it easier. After all, they have been making PCBs for qutie awhile.

 

[ManicOne]

512 bit would also effect minimum die size. R600 is launching at 80nm, with leaked pictures showing the gpu package to have about 2140 pins. If AMD/ATI push to 65nm, then managing die contacts for 512 bit on such a small die would seem nightmarish.

I tend to agree with B3D though; I cannot conceive of any other reason for 2140 pins outside of 512 bit.

 

[phide]

I really wouldn't be surprised if G80 ended up about half the performance of R600. With the MUL not being effective for most shading, it only has the shading power of around a R580 to begin with.

I guess we're just making stuff up today. Well, as always, I'll choose reality:

G80 only falls in HDR shader performance, and I assume that's because the precision is twice that of R580/G7x. All shader precision is also up with G80, and G80's shader output more closely approximates the Microsoft DirectX software rasterizer. Not bad, huh?

 

[That's_Corporate]

sounds like this could be a really big flop if not done right. Hopefully it will be done right, 512bit is insane and I don't think the average computer on this forum wouldn't be bottlenecked.

But besides Crysis which has been delayed until Q3 2007...what will all this power go towards that a simple 7900 can't run?

So us 24" - 30" widescreen guys can run games at 60 FPS with the detail and eye candy pumped.

 

[mavalpha]

Even if true, nVidia has already shown us once before that bandwidth isn't everything.
6800 Ultra=1.1GHz, 256-bit, 35.2GB/s
7800GTX=1.2GHz, 256-bit, 38.4GB/s

The difference in bandwidth was negligible, yet I wouldn't trade even a pair of NV40's for a single G70.

Optimizations in core design make for a much bigger part of of the gains from one generation to the next, though the headroom makes revisions and refreshes easier.

 

[Obi_Kwiet]

Uh, 24 pipes vs. 16? What did you expect?

 

[Silus]

I really wouldn't be surprised if G80 ended up about half the performance of R600. With the MUL not being effective for most shading, it only has the shading power of around a R580 to begin with. That's not to say G80 isn't still a good card if it's half the price.

I find that very unlikely. With these rumored specs, R600 really looks good and I'm sure it will beat this first version of G80, but not by much. Unless AMD/ATI has something up its sleeve, I think the performance difference won't go over 5%-10% in most cases.
Shading power of around R580 ? What reviews have you been reading about G80 ? Shader performance with G80, is beyond anything R580 or G70 ever offered. There's just no comparison really.

 

[ManicOne]

Of course G80 destroys R580. However if you look at theoretical peak shading power, G80 doesn't seem that much more powerful. The trick is in its efficiency though. R580 being non unified virtually never reaches its theoretical maximum shading power. G80 on the other hand pretty much sits at peak power all of the time. Efficiency is one of the major pluses of unified architectures; unified gpu's have virtually none of their transistors (especially wrt ALUs) sitting idle.

And as phide showed above, the benefits are pretty bloody obvious.

 

[phide]

I'm still of the opinion that G80 isn't terribly efficient. It seems so from one perspective (in terms of power consumption and transistor count), but not so from another (doesn't seem to reach theoretical performance). Of course, you still have to factor in the major changes in the severity of optimizations, and G80 looks good in this respect. NVIDIA over-optimized for the G7x series for a reason.

I did some math a little while ago, and G80 is still transistor-for-transistor more efficient than G71, but not tremendously so (~5%). I assume that later drivers may bring performance up higher and increase efficiency by some moderate amount.

If these R600 rumours are true (and they're shaping up to be), and if ATi can grasp the capability of the hardware through their drivers, the X2800s are likely going to pass the 8800s in terms of performance, and at least match it in IQ. With the excessive memory bandwidth, the performance penalties of multisampling are reduced, and generic multisampling will probably be close to free. But can ATi counter coverage sampling anti-aliasing?

 

[Endurancevm]

Was hoping to have "free" AA...that would be nice.

 

[That's_Corporate]

I'm still of the opinion that G80 isn't terribly efficient. It seems so from one perspective (in terms of power consumption and transistor count), but not so from another (doesn't seem to reach theoretical performance). Of course, you still have to factor in the major changes in the severity of optimizations, and G80 looks good in this respect. NVIDIA over-optimized for the G7x series for a reason.

I did some math a little while ago, and G80 is still transistor-for-transistor more efficient than G71, but not tremendously so (~5%). I assume that later drivers may bring performance up higher and increase efficiency by some moderate amount.

If these R600 rumours are true (and they're shaping up to be), and if ATi can grasp the capability of the hardware through their drivers, the X2800s are likely going to pass the 8800s in terms of performance, and at least match it in IQ. With the excessive memory bandwidth, the performance penalties of multisampling are reduced, and generic multisampling will probably be close to free. But can ATi counter coverage sampling anti-aliasing?

...or even more importantly, which card will play Privateer better...

 

[TheRapture]

While it may indeed prove to be faster....the R600 will NOT be "double" the performance of G80.....no way.