HD 4770 Previews/Reviews

[SNiiPE_DoGG]

Finally got a card on Friday from an anonymous source. Then spent most of the day sending emails to my rep at ATI for drivers without a response, spent the better part of the weekend calling in favors to get the most current drivers sent to me off the grid.

No thanks to ATI, a review will be up one way or another.

@ Rasamaha: would you mind telling us which program you used to bypass ATI's CCC overclocking limitations?

Can you do some idle power testing in the review to show what the consumption would be -IF- the powerplay was working? I just wanna know how low it can go

[blaH]

would be interesting if there will be non reference card with just 4 RAM chips, maybe power consumption would be better and there would be less of those repeating post about it

[Mr. K6]

Can you do some idle power testing in the review to show what the consumption would be -IF- the powerplay was working? I just wanna know how low it can go

Me too, very interested for HTPC use.

[Mats]

My guess is that the 4750 is the low power consumption alternative since it got GDDR3.

[Boissez]

No need to test it, it's 72GB/s ~ GTS 250.

Not quite... GDDR5 has much higher latency than GDDR3 but a 40% overclock is mighty impressive nonetheless

[Macadamia]

What if the 4750 matched the 9600GT in cases where bandwidth limitations weren't that severe?


I heard that was their target previously, 9800 for the highend version and 9600 for the GDDR3. Granted, the lake of salt is right... there.

[trinibwoy]

Not quite... GDDR5 has much higher latency than GDDR3 but a 40% overclock is mighty impressive nonetheless

Latency isn't particularly relevant for GPUs. They are built to hide it. In any case you can never get any reliable measure of actual bandwidth. It varies according to workload. Best metric for comparison that we've got is theoretical maximums.

[Boissez]

Latency isn't particularly relevant for GPUs. They are built to hide it. In any case you can never get any reliable measure of actual bandwidth. It varies according to workload. Best metric for comparison that we've got is theoretical maximums.

The HD4850 would be about 10-15% slower if it had GDDR5 instead of GDDR3 - i'd say it's relevant

[SNiiPE_DoGG]

The HD4850 would be about 10-15% slower if it had GDDR5 instead of GDDR3 - i'd say it's relevant

...what are you talking about? ggdr5 vs gddr3 is the only difference between the 4850 and the 4870... I'd say its a pretty big difference.

[Boissez]

...what are you talking about? ggdr5 vs gddr3 is the only difference between the 4850 and the 4870... I'd say its a pretty big difference.

I don't quite get your point. I'm just arguing that GDDR5 is about 10-15% slower than GDDR3 given the same bandwidth.

[ToTTenTranz]

The HD4850 would be about 10-15% slower if it had GDDR5 instead of GDDR3 - i'd say it's relevant

erm... no.

Those graphs are comparing GDDR3 with GDDR5 at half the speed of GDDR3. For example, in that case it's 993Mhz GDDR3 versus 497Mhz GDDR5. However, it would be pretty stupid for the HD4850 to have 500Mhz GDDR5. If the HD4850 had GDDR5, it would be 750Mhz minimum.

I don't quite get your point. I'm just arguing that GDDR5 is about 10-15% slower than GDDR3 given the same bandwidth.

No, it's given the same.. let's call it the same "single data-rate clockspeed".

It's always been like that. DDR was already slower than SDR if DDR was underclocked to the half of the speed of SDR.

What if the 4750 matched the 9600GT in cases where bandwidth limitations weren't that severe?


I heard that was their target previously, 9800 for the highend version and 9600 for the GDDR3. Granted, the lake of salt is right... there.

Well.. ATI has launched their share of highly bottlenecked cards before, like the HD4650 DDR2 (phail...).
Or even worse: the FirePro V3750, which is a 320sp RV730 with a 64bit memory controller (yeah, the original 320sp R600 had 8x more bandwidth).

[hurleybird]

erm... no.

Those graphs are comparing GDDR3 with GDDR5 at half the speed of GDDR3. For example, in that case it's 993Mhz GDDR3 versus 497Mhz GDDR5. However, it would be pretty stupid for the HD4850 to have 500Mhz GDDR5. If the HD4850 had GDDR5, it would be 750Mhz minimum.



No, it's given the same.. let's call it the same "single data-rate clockspeed".

It's always been like that. DDR was already slower than SDR if DDR was underclocked to the half of the speed of SDR.

That's exactly what the dude is saying. GDDR5 will be slower than GDDR3 at the same *effective* clockspeed. Ergo, while the 4770 can get in the same ballpark as GTS 250 in terms of bandwidth, it's memory subsystem will still probably be slower on the whole. Of course, take all this speculation with a grain of salt as the GTS 250 is a lot older, the memory controller in the HD 4770 is tweaked vs. the 4800 series, and the less complex 128-bit interface might result in some kind of lower latency (whether or not it would be significant, I have no idea)

[Andi64]

I don't quite get your point. I'm just arguing that GDDR5 is about 10-15% slower than GDDR3 given the same bandwidth.

Not exactly. That is true given the same bandwidth but maintaining the timings required for 1000Mhz (GDDR5-4000) operation. Running at 500Mhz will tolerate better timings thus increasing the performance.

Anyway, at the same BW GDDR5 will be slower, but it's far from 10% with adequate timings.

[SKYMTL]

Can you do some idle power testing in the review to show what the consumption would be -IF- the powerplay was working? I just wanna know how low it can go

In an upcoming article, yes. That will be coupled with some additional cooling tests since right now there is no way to accurately test the heat output of the 40nm core against any other card due to the oddball offset of the heatsink mount. After some preliminary tests with some modded heatsinks, I am certain the results will shock many people who are defending the move to 40nm.

[Wadkiller]

Did you get 2 cards to do some crossfire testing SKYMTL?

[SKYMTL]

I heard that was their target previously, 9800 for the highend version and 9600 for the GDDR3. Granted, the lake of salt is right... there.

Great for them but the 8800 GT / 9800 GT has been their target for the LAST TWO YEARS. Even the HD 4830 didn't win convincingly against the 9800 GT.

[SKYMTL]

Did you get 2 cards to do some crossfire testing SKYMTL?

As I stated above, getting even one card was a stretch this time. Without ATI's support, websites have to go either to board partners (who didn't actually get the cards until last week) or to our contacts in Asia who may or may not get it to us in time. As it stands, I have one and will probably receive another a bit later in the week.

[BrowncoatGR]

In an upcoming article, yes. That will be coupled with some additional cooling tests since right now there is no way to accurately test the heat output of the 40nm core against any other card due to the oddball offset of the heatsink mount. After some preliminary tests with some modded heatsinks, I am certain the results will shock many people who are defending the move to 40nm.

X1650 compatible heatsinks dont fit?

[Wadkiller]

Sorry asking something you already posted

I'm really curious as to how the crossfire will perform. If a single card perform almost on par with a 4850, in crossfire they surely be close aswell, right? I'm either getting two of these or two 4850's. Not sure which will be better. I't s pity the stuff is so hellishly expensive in South Africa compared to the USA/UK

[SKYMTL]

X1650 compatible heatsinks dont fit?

Those I don't have so I don't know if they would fit. If you can find a heatsink with a 4.5cm mounting hole offset then let me know. For the time being I have drilled some extra holes in a HR-03 GT bracket I had lying around.

[ToTTenTranz]

In an upcoming article, yes. That will be coupled with some additional cooling tests since right now there is no way to accurately test the heat output of the 40nm core against any other card due to the oddball offset of the heatsink mount. After some preliminary tests with some modded heatsinks, I am certain the results will shock many people who are defending the move to 40nm.

The only people who should be defending 40nm at this point is ATI, because they can put more cores into each waffer and start testing and tweaking this new process.
No "new" process has ever been instantaneously beneficial for power consumption and overclocking capacity, as far as I know.

Great for them but the 8800 GT / 9800 GT has been their target for the LAST TWO YEARS. Even the HD 4830 didn't win convincingly against the 9800 GT.

I think you got things backwards.
AMD doesn't target GPUs, it targets price-points. What chip is in what graphics card and how old the architecture is, is completely irrelevant for the end user. What matters is what performance+features you get for how much money, period.
And the HD4830 is quite successfull at its price point. Where I live, the HD4830 is priced at the level of a 9600GT.

That's exactly what the dude is saying. GDDR5 will be slower than GDDR3 at the same *effective* clockspeed.(...)

He did say bandwidth, not clockspeed. If he "meant" clockspeed, all is well .

[Wadkiller]

X1650 compatible heatsinks dont fit?

Yes they will. According to the Ecpreview review they use the same 43mm spacing as on the x1650/8600gt/8500gt

I was kinda worried about that, cause i have two Dangerden Maze 4's with both 43mm & 53mm (as the 8800gt) holes.

[trinibwoy]

I don't quite get your point. I'm just arguing that GDDR5 is about 10-15% slower than GDDR3 given the same bandwidth.

You can't really draw that conclusion from the data we have. First of all, Damien's test underclocked the GDDR5 way below its normal operating parameters. Secondly there are a host of other variables that affect the efficiency including the sizing of buffers and command protocols used by the memory controllers.

So simply underclocking GDDR5 to half its speed and saying "see it's slower than GDDR3 at the same speed" isn't a very scientific test. Or in other words there's no way we can say that RV740's 72GB/s isn't as good as G92's 72GB/s.

[SKYMTL]

The only people who should be defending 40nm at this point is ATI, because they can put more cores into each waffer and start testing and tweaking this new process.

I agree but they are marketing it as efficiency personafied.

AMD doesn't target GPUs, it targets price-points. What chip is in what graphics card and how old the architecture is, is completely irrelevant for the end user. What matters is what performance+features you get for how much money, period.
And the HD4830 is quite successfull at its price point. Where I live, the HD4830 is priced at the level of a 9600GT.

Upon its release the HD 4830 was in the same price point as the 9800 GT. To the dollar here in Canada. It was only through the latest round of price reductions that the HD 4830 is now able to compete price-point wise with the 9600 GT.

Pricing based on a feature list is completely understandable but there comes a price point where people start looking less and less at features and more about gaming capacity. To me, that is right around the $100 - $120 price points and above since below that you can get the same features (albeit less gaming potential) for much less money.

[ToTTenTranz]

I agree but they are marketing it as efficiency personafied.

You mean the marketing guys are trying to deceive people?!?? Now that's a first!

Upon its release the HD 4830 was in the same price point as the 9800 GT. To the dollar here in Canada. It was only through the latest round of price reductions that the HD 4830 is now able to compete price-point wise with the 9600 GT.

So you agree with me when I say the HD4830 is, right now, a successfull product in its price range? Of course, a RV770, even a salvaged one, should be more expensive to make than the G94b, but that's one of the main reasons for the RV740 appearance anyways.