AMD To Launch Radeon HD4890 "RV790" In April

[LordEC911]

If it doesn't have a separate shaderclock domain the shaders won't scale.
Maybe thats the reason why 13% higher clocks means 6% performance boost.


HD4890 is listed in Holland
http://tweakers.net/pricewatch/23669...dr5-pci-e.html

Huh? It will still scale linearly since the core speed would increase the ROPs, TMUs and shaders.
That 6% increase is supposedly clock vs clock.

Curious maybe someone can explain but why does the ATI card only have 256-bit memory interface, It has a bit more ram than the GTX yet the GTX has 448-bit memory interface?

GDDR5. They can have a 256bit interface and still have similar bandwidth vs the wider buses on the G200 and GDDR3.

[flippin_waffles]

My wild guess here is that AMD knew there was fair, but not an overwhelming amount of improvements capable out of the RV770 core. So they poured the majority of engineering resources into RV870, which there are now rumours of an earlier than expected release. Still though, I bet the HD4890 will be a worthy successor to the HD4870 in a couple ways. However, there is a good chance that i'm lost out in left field here too.

[Salsoolo]

My wild guess here is that AMD knew there was fair, but not an overwhelming amount of improvements capable out of the RV770 core. So they poured the majority of engineering resources into RV870, which there are now rumours of an earlier than expected release. Still though, I bet the HD4890 will be a worthy successor to the HD4870 in a couple ways. However, there is a good chance that i'm lost out in left field here too.

maybe GT300 and RV870 would go out around windows 7 final.
i could be wrong but thats what i noticed when vista came and new cards was out, i think theire connected somehow.

[LordEC911]

My wild guess here is that AMD knew there was fair, but not an overwhelming amount of improvements capable out of the RV770 core. So they poured the majority of engineering resources into RV870, which there are now rumours of an earlier than expected release. Still though, I bet the HD4890 will be a worthy successor to the HD4870 in a couple ways. However, there is a good chance that i'm lost out in left field here too.

More like AMD realized how close RV770 was competing already with G200 and with the delays on 40nm, decided a RV790 would be a good filler until they could get RV870 out.

40nm was suppose to be ready Q4 '08, was delayed a quarter, RV870 was a late Q2 target. Then was pushed back to Q3, with only a rumor of another delay pushig it back into Q4, even though most rumors said RV870 is still on track and looking like a Q3 launch.

[kemo]

Were do you guys read that 6% faster clock per clock ,,

Internal benchmarks claim 6% faster than RV770

that means that RV790 is 6% faster than RV770 ,, 850Mhz vs 750Mhz , it could even be 6% from 4870 to 4890

[LordEC911]

Were do you guys read that 6% faster clock per clock ,,

that means that RV790 is 6% faster than RV770 ,, 850Mhz vs 750Mhz , it could even be 6% from 4870 to 4890

Ummm... because 6% doesn't make sense...
The only way it makes sense is if it is 6% increase clock per clock.

[largon]

Curious maybe someone can explain but why does the ATI card only have 256-bit memory interface, It has a bit more ram than the GTX yet the GTX has 448-bit memory interface?

To be precise;
HD4870 has eight GDDR5 chips each 1Gbits (128MB) in capacity yielding total of 8×128MB=1GB and since each chip has a bus width of 32bits the whole bus combined is 8×32bit=256bit.
GTX260 has 14 GDDR3 chips, each 512Mbits (64MB) totalling 14×64MB=896MB, total bus is 14×32bit=448bit. GTX280 has 16 chips and as such for the calculations: 14->16.

Were do you guys read that 6% faster clock per clock
that means that RV790 is 6% faster than RV770 ,, 850Mhz vs 750Mhz , it could even be 6% from 4870 to 4890

It's not wise to believe everything you read. Fact is the 13% frequency increase alone will gain 13% more in performance.

And come on, would it make sense producing a different silicon for mere 6% total performance increase. They could have cherrypicked age-old RV770s to get even more than 6%, cheaper and with less of a hassle.

[ColonelCain]

It's not wise to believe everything you read. Fact is the 13% frequency increase alone will gain 13% more in performance.

And come on, would it make sense producing a different silicon for mere 6% total performance increase. They could have cherrypicked age-old RV770s to get even more than 6%, cheaper and with less of a hassle.

True, it does seem like a bit of a hassle, but it is a "new" chip after all.


Also, do we yet have confirmation of it being on the 55GT process?

[Muunsyr]

Curious maybe someone can explain but why does the ATI card only have 256-bit memory interface, It has a bit more ram than the GTX yet the GTX has 448-bit memory interface?

To be precise;
HD4870 has eight GDDR5 chips each 1Gbits (128MB) in capacity yielding total of 8×128MB=1GB and since each chip has a bus width of 32bits the whole bus combined is 8×32bit=256bit.
GTX260 has 14 GDDR3 chips, each 512Mbits (64MB) totalling 14×64MB=896MB, total bus is 14×32bit=448bit. GTX280 has 16 chips and as such for the calculations: 14->16.

-Snip-

To add to what largon has said, that the 4870/4890's are using GDDR5, where as the GTX's are using GDDR3.

The GDDR5 is quad pumped, so 900Mhz has four bits per clock cycle = 3600Mhz effective. 3600Mhz x 256 bit = ~921600Mbps = 900Gbps.

The GDDR3 RAM is only two bits per clock cycle. So 999MHz x 2 = 1998MHz effective clock speed and 1998MHz x 448bit = 895104Mbps = ~874Gbps

So you see, the total actual memory bandwidth is much more similar between the cards, its just that they use different methods to get there.

Just for the hell of it I will also mention that from a card manufacturing point of view, the lower bus width (eg, 4870) is easier to implement - there is half the number of wire/traces to cram onto the circuit board with 256bit bus vs. a 512bit bus. This leads to a cheaper to produce card and is possibly part of the reason why the radeons are cheaper. It is also the technique that has been successfully used to make many a mid-range card. Create high end card, let manufacturing processes mature and cheapen, chop memory bandwidth in half and use a cheap but matured memory technology running at full steam and you have a decent performing mid range part. Anyway, enough lecturing, someone will correct me soon.

-Muunsyr

[Shadowmage]

Beyond3D thinks that ATI upped the die size by a tiny amount, based on pixel counting from the leaked pictures. A small IPC increase looks very plausible.

[LordEC911]

To add to what largon has said, that the 4870/4890's are using GDDR5, where as the GTX's are using GDDR3.

The GDDR5 is quad pumped, so 900Mhz has four bits per clock cycle = 3600Mhz effective. 3600Mhz x 256 bit = ~921600Mbps = 900Gbps.

The GDDR3 RAM is only two bits per clock cycle. So 999MHz x 2 = 1998MHz effective clock speed and 1998MHz x 448bit = 895104Mbps = ~874Gbps

So you see, the total actual memory bandwidth is much more similar between the cards, its just that they use different methods to get there.

Just for the hell of it I will also mention that from a card manufacturing point of view, the lower bus width (eg, 4870) is easier to implement - there is half the number of wire/traces to cram onto the circuit board with 256bit bus vs. a 512bit bus. This leads to a cheaper to produce card and is possibly part of the reason why the radeons are cheaper. It is also the technique that has been successfully used to make many a mid-range card. Create high end card, let manufacturing processes mature and cheapen, chop memory bandwidth in half and use a cheap but matured memory technology running at full steam and you have a decent performing mid range part. Anyway, enough lecturing, someone will correct me soon.

-Muunsyr

Yep, you forgot to divide by 8.
4870 has 115.2Gbps of bandwidth and GTX260 has 111.9Gbps.

[Muunsyr]

Yep, you forgot to divide by 8.
4870 has 115.2Gbps of bandwidth and GTX260 has 111.9Gbps.

Not quite, dividing by 8 gives us GBps (gigabytes per second) as apposed to Gbps (Gigabits per second)

(It must be late over there, because I also make that out to be 112.5GBps for the Radeon, not 115.2...)

[Odyssey96]

I hope the 4890 will be able to put pressure on the GTX285 and 275

[LordEC911]

Not quite, dividing by 8 gives us GBps (gigabytes per second) as apposed to Gbps (Gigabits per second)

(It must be late over there, because I also make that out to be 112.5GBps for the Radeon, not 115.2...)

Well, I have never seen anyone post those numbers you did...

[ownage]

http://www.pcgameshardware.de/aid,67...fikkarte/News/

[Calmatory]

Higher clocks and lower FPS? Seems like a possible driver problem for me. Then again, there should be no major changes in the driver area if only the clocks are changed. ...completely different story if the whole shader structure/amount is changed...

Something is wrong there.

[Lightman]

Promising gains in FC2

[Muunsyr]

Something does smell off.

Given that one of these tests is clearly not indicative of the performance difference, this leaves us with one tests results which is hardly enough to start judging the card on. Especially so that different drivers are used for each card.

Edit: Also, is that 850/2.200 and 750/1.800 the core/memory frequency of each card respectfully? If so, it would suggest that the 4890 card tested has 1100Mhz (4400Mhz effective) memory. That or the reviewer can't type properly.

[perkam]

PCGH's early numbers are never spot on, but you can see a nice delta developing on the higher resolutions.

Plus, I want to see this thing clocked.

Perkam

[Smartidiot89]

Higher clocks and lower FPS? Seems like a possible driver problem for me. Then again, there should be no major changes in the driver area if only the clocks are changed. ...completely different story if the whole shader structure/amount is changed...

Something is wrong there.

There has to be, or else I will be the first one saying "ATI did an Nvidia..."

[Jamesrt2004]

shouldnt they have tested with the same beta driver for both cards + what does it mean x64/x86 catalyst driver :/ like did they test one with 64 bit and the other card with 32bit

[Eastcoasthandle]

Those results from the 4890 do not come from PCGH and they say so in their thread. Take a look at the crysis results. There you will find a watermark from someone else.

[ILikeCosmosS]

i am having high hopes for the 4890 to compete with 285 and the 4890x2 compete with 295

[perkam]

i am having high hopes for the 4890 to compete with 285

Unlikely.

and the 4890x2 compete with 295

Very likely.

Perkam

[eric66]

lol 4890 has standard fps in crysis there a problem there but far cry test is incredible nearly 25 percent