ATI RV740 review/preview

[Tchock]

@ Jawed, do you know how OEMs tend to approach these cards between the single/dual slot issue?

Do they prefer single slot for being cheaper, or do they prefer dual slot for the better thermals without having to use additional cooling?

I'm not sure OEMs even deviate from the reference design (at all). Unless ATI has another version with a cheaper cooler that's more justified for the 80W TDP this card has for OEMs exclusively.


I wonder if the G/DDR3 versions would be enough to satisfy the OEMs by costing a whole load less.

 

[Squilliam]

I'm not sure OEMs even deviate from the reference design (at all). Unless ATI has another version with a cheaper cooler that's more justified for the 80W TDP this card has for OEMs exclusively.


I wonder if the G/DDR3 versions would be enough to satisfy the OEMs by costing a whole load less.

Cheap is better for OEMs. I was amazed sometimes at how the OEMs never included better than a low end card for up to the highest end computers. You could pay $1000 more, but barely have any of that money spent on the GPU, just a few more Htz.

 

[Tchock]

Cheap is better for OEMs. I was amazed sometimes at how the OEMs never included better than a low end card for up to the highest end computers. You could pay $1000 more, but barely have any of that money spent on the GPU, just a few more Htz.

It's hard to quantify GPUs. GigaHURTs are easy to *ahem*.

Yeah, either prebuilt or OEM, the configs are mostly terrible. And the component selection is sub-par at best (no solid caps, thin layers, passable cooling etc)


Unless it's from Falcon Northwest.

 

[Jawed]

826M transistors compared to 770's 956M. (Chiphell)

There's quite some trannies allocated to the extra width of the MC, the majority I assume are for the cluster?

Only 40% of RV770 is clusters, about 390M transistors. Of that, I estimate 64% is ALUs, excluding the redundant ALU lanes, which I like to lump in with the TU when looking at a die shot and which makes scaling estimates a bit simpler.

So RV770 has about 250M transistors for its 800 ALU lanes and 140M for the rest. RV740's 640 lanes would be ~200M transistors, subject to new functionality and lack of double precision. The clusters, as a whole, would be about 312M, leaving, as you observe, a hell of a lot of transistors, 514M, for MCs, RBEs, the hub, PCI Express etc.

This compares with ~566M in RV770

Analogue stuff isn't supposed to shrink particularly well - dunno how to account for that.

Jawed

 

[CarstenS]

Does anyone have confirmation whether or not the HD 4770 has a PCI-E connector? I didn't see it on the above link. If it does, should I expect a lesser SKU without said connector?

When viewing the larger version pics in the link you can make out them connectors.

Dual-slot cooler for HD4770 - that's going to divide opinion!

Jawed

Me, i am all for it as it seems to be used to vent the card's heat out of the chassis.

 

[mczak]

HD4770 details:
http://publish.it168.com/2009/0331/20090331020601.shtml

Probably not really news by now, but I consider the lineup a bit disappointing. Not to say these are bad cards (for the price point), but a bit a pity considering the chip could do so much more. With some faster gddr5 ram (might not even need faster core clock) it could potentially beat or at least equal a HD4850 in performance from what we know. And don't even think about the HD4750, this one will be so limited by memory bandwidth it probably will only be slightly faster than HD4670 (that link didn't mention memory clock, but I doubt it's going to be faster than those magic 1Ghz). I suspect it will ship with a fairly low core clock so should be an amazing overclocker, too bad you won't gain any performance at all by doing so...

 

[no-X]

With some faster gddr5 ram (might not even need faster core clock) it could potentially beat or at least equal a HD4850 in performance from what we know.

The 650Mhz prototype at Guru3D was only 8-10% slower than HD4850, so how could be the final product with more than 15% faster core (750MHz) end slower than HD4850?

I think this ~135mm² 128bit GPU will be only slightly slower than ~235mm² 256bit G92b (GTS250 / GF9800GTX). That's not really bad, is it?

 

[Kaotik]

The 650Mhz prototype at Guru3D was only 8-10% slower than HD4850, so how could be the final product with more than 15% faster core (750MHz) end slower than HD4850?

I think this ~135mm² 128bit GPU will be only slightly slower than ~235mm² 256bit G92b (GTS250 / GF9800GTX). That's not really bad, is it?

If it's beating HD4850, it's beating 9800GTX, and should beat GTX+/GTS250 which are about on par with HD4850

 

[Lukfi]

How could it possibly beat the HD 4850 if it has 960 paper GFLOPS opposed to 1 TFLOPS and less memory bandwidth? :|

 

[mczak]

The 650Mhz prototype at Guru3D was only 8-10% slower than HD4850, so how could be the final product with more than 15% faster core (750MHz) end slower than HD4850?

The prototype already had the 3200Mhz gddr5 ram. HD4850 is already somewhat memory bandwidth limited, so will be this card (even has 20% less memory bandwidth than HD4850 at this memory clock, and that's not accounting for possibly slightly less efficiency of gddr5), hence a 15% core increase is probably only going to yield like 5% more performance.
Also from that new leak, AMD has this positioned between 4830 and 4850 (I'm not going to argue it will probably be closer to 4850 rather than 4830).

I think this ~135mm² 128bit GPU will be only slightly slower than ~235mm² 256bit G92b (GTS250 / GF9800GTX). That's not really bad, is it?

No it's not, I was merely arguing with some faster ram it could probably beat it (800Mhz gddr5 is really the slowest even available - I'm not saying AMD should use the fastest parts available for this budget part but is 1000Mhz gddr5 as used on the 4850 really that much more expensive?). Well, at least those 800Mhz gddr5 parts should probably overclock quite well.

 

[Blazkowicz]

and how do they keep on selling the 4850 if it's more expensive and slower? the slides say a $99 4770 replaces the $99 4830, while the RV770/790 cards go on in the same price brackets.

 

[mczak]

How could it possibly beat the HD 4850 if it has 960 paper GFLOPS opposed to 1 TFLOPS and less memory bandwidth? :|

That's a tiny difference in TFLOPs. Also, don't forget that core clock is higher so some things which don't scale with more shader units (like setup rate) should be faster (rop units should also be faster though I doubt it makes a difference as those are likely bandwidth limited anyway). But yes, the memory bandwidth difference will likely play the decisive factor here which one is going to be faster IMHO...

 

[Kaotik]

How could it possibly beat the HD 4850 if it has 960 paper GFLOPS opposed to 1 TFLOPS and less memory bandwidth? :|

Because paper GFLOPS != real life GFLOPS, RV770 could be more efficient in real life situations.
Also, if they use faster than 933MHz GDDR5, it will have more memory bandwidth

and how do they keep on selling the 4850 if it's more expensive and slower? the slides say a $99 4770 replaces the $99 4830, while the RV770/790 cards go on in the same price brackets.

Well, there has been rumors about phasing out RV770 products and bringing more RV790 variants to fill the gap (there was even some news about new RV790-products supported in Cat 9.5)

So it's not impossible that while 4770 replaces 4830, it would still perform better than HD4850, and phasing out RV770's would give room for a RV790 performing somewhere between HD4850 and 4890 at old 4850's price point

 

[Blazkowicz]

Probably the RV790 replaces RV770 on 4850 and 4870 cards (no need to do new commercial launches for using a minor revision of the same GPU)

 

[Kaotik]

Probably the RV790 replaces RV770 on 4850 and 4870 cards (no need to do new commercial launches for using a minor revision of the same GPU)

More likely 4850 at least will be discountinued due RV770 performance.
Also, there's more to RV790 than just Burst Memory Reads support it seems, check the RV790 thread for that (though I don't know yet what it is)

 

[Tchock]

Are they kinda stuck between a rock and a hard place here?


IF 650/3200 - under 75W TDP, but close enough to be risky. 2600XT GPU recover issues, do not want.

IF 750 (maybe even 800) /3600-4000 - what's the real point of it then? Early revs would probably use power extremely close to a 4850, cannibalizing 4850 sales, and make 40nm rather moot (cost savings negated by 40nm frontier properties for yield I guess- for now at least)

Self-overclocking to reach those clocks, however is a different domain, and much more welcome from a psychological perspective (reviewers and customers are human no?), not that hard hitting on yields either.

I guess for both 740 and 790 there is more headroom for clocks to be pushed. Influenced by partner pressure? Could be.

 

[Kaotik]

Are they kinda stuck between a rock and a hard place here?


IF 650/3200 - under 75W TDP, but close enough to be risky. 2600XT GPU recover issues, do not want.

IF 750 (maybe even 800) /3600-4000 - what's the real point of it then? Early revs would probably use power extremely close to a 4850, cannibalizing 4850 sales, and make 40nm rather moot (cost savings negated by 40nm frontier properties for yield I guess- for now at least)

Self-overclocking to reach those clocks, however is a different domain, and much more welcome from a psychological perspective (reviewers and customers are human no?), not that hard hitting on yields either.

I guess for both 740 and 790 there is more headroom for clocks to be pushed. Influenced by partner pressure? Could be.

If they discontinue 4850 there's no sales to cannibalize, this chip is after all a lot cheaper to manufacture so it might be worth it even if these are sold cheaper too.

 

[AnarchX]

If they discontinue 4850 there's no sales to cannibalize, this chip is after all a lot cheaper to manufacture so it might be worth it even if these are sold cheaper too.

According to the AMD slides from IT168 HD4850 will be continued and only 4830 will be replaced.
Maybe the told partners to focus on 4850 OC-SKUs and 1GiB cards.

 

[Kaotik]

Maybe, though despite the slide it is possible that the HD4850 (and HD4870) will be discontinued a bit later when there's more RV790-SKUs to replace the gap left by RV770's (rumors of more RV790 based solutions in Cat 9.5)

 

[no-X]

The prototype already had the 3200Mhz gddr5 ram. HD4850 is already somewhat memory bandwidth limited, so will be this card (even has 20% less memory bandwidth than HD4850 at this memory clock, and that's not accounting for possibly slightly less efficiency of gddr5), hence a 15% core increase is probably only going to yield like 5% more performance.
Also from that new leak, AMD has this positioned between 4830 and 4850 (I'm not going to argue it will probably be closer to 4850 rather than 4830).

Even GTS250 isn't as much BW limited to gain only 5% of performance by 15% OC. HD4850 is BW limited only when using MSAA 8x. And nobody cares about MSAA 8x performance of $99 product.

Anyway, according to a person who spoke with someone from Sapphire, the cards are ready, but manufacturers wants to wait with the lauch, because HD4830s aren't sold-out yet. I believe, that HD4850 is in similar situation and it stays in the roadmap only beacause the manufacturers needs to clear off.

It doesn't make sens to keep HD4850 alive even if the HD4770 will be 5% slower.