HD4650 - poor man 256-bit board

so this thing is basically a 3850 right?

 

http://www.3dnews.ru/news/fotografiya_i_spetsifikatsii_radeon_hd_4650/
Their source is ChipHell and they say, RV730 will have a 128 bit wide MB.

 

Hmm, even if it were "just a 3850", wouldn't be amazingly good for a x6xx class card?

 

If its priced at 3650 range it'll be awesome. I didn't mean it as a negative. However, 128bit bus could be problematic for performance.

 

Someone go and pick up some HD3690/3830 benches, it has 128bit membus

 

RV730 is 128-Bit:


http://bbs.chiphell.com/viewthread.php?tid=26649&extra=page=1

And interesting is, that it will need with its 320SPs/16TMUs@710MHz no external power connector.

 

looks like the circut board is made of plywood

 

Hmm, what's that second 16 in the table? ROPs? In this generation ROPs are tied to memory channels, so either it's 8 for a 128bit config or the channels are only 32bit wide. Either way, could be a very decent mainstream card - with basically tying a 3850 in arithmetic and texturing throughput, at least with the 128bit memory at 1Ghz I don't think it would suffer too much due to bandwidth. Sadly, the configurations with fast memory (speaking of cards of that price segment) are rarely seen in the wild usually...
If these specs are true, should blow out any D9M/G96 configuration out of the water, if that's what the competition is. Seems a bit like a mismatch, maybe some cut-down G94 would be a better match (what happened to the 9600gs)?

 

hardware.fr and other site claim a 256bit bus, also there's a HD4670 which has 900MHz core clock (good! 3870 is 775) and 900MHz memory.
so they seem going against the G94, the 4650 below it and the 4670 above it, possibly a contender for the 9800GT?

 

Hardspell says it has 360 SPs and 12 ROPs.

http://en.hardspell.com/doc/showcont.asp?news_id=3833

 

I reckon 360 ALU lanes would indicate 6 clusters of 60 ALU lanes, i.e. 3x quad ALUs per cluster, ALU:TEX 3:1.

It seems a bit strange to have 12 RBEs though with a 256-bit bus - I'd expect 1 quad RBE per 64-bits of memory

Jawed

 

192bit membus?

 

If true, would be quite a fast card indeed presumably.

Yes, with the RBEs now tied to memory channels 12 RBEs with a 256bit bus seems impossible - not sure a 192bit memory bus would make sense though for this class of product?

 

9600GSO is near that range right now.

 

RBEs are not tied with MCs!

RV63x: 1 RBE, 128-Bit
RV610: 1 RBE, 64-Bit
R600Pro: 4 RBE, 256-Bit
R600GT: 3RBE, 256-Bit
...

 

This is supposed to be RV7xx though, an architecture that seemingly does tie RBEs to MCs. Certainly does so in RV770.

Jawed

 

Two notes:
- Ati did change mem-controllers for 'lesser' products in the past (only x30 and above had a ringbus)
- 256 bit / 16 ROPs might just be a coincidence.

 

Yes, but this product is a downcut version (for whatever reasons) of another one. The "main" products still have a 128bit (G96) or 256bit (G94) bus. Not that a 192bit bus wouldn't make sense, but the "odd" resulting video memory size might even be problematic. A 384MB 192bit card can't be as cheap as a 256MB 128bit card, and some people might prefer a 512MB 128bit card simply because it's got more ram even though it would be slower.

 

I'd like to see minimum video memory size move up to 384MB. There's placement issues with any new configuration, but If they are going for super low cost 128bit would make the most sense. I was only suggesting they wouldn't be pioneering a low price 192bit part.