AMD RV770 refresh -> RV790

JoshMST said:
As for how double precision is handled... my understanding is that in the group of 5 stream processors, one was double precision and the other 4 were single precision (plus that group of 5 has the general registers assigned to them and the branch prediction unit). Is my understanding of this incorrect?
Click to expand...

Double precision is handled by ganging together pairs of slim ALUs, but without MADD capability.
From my fuzzy recollection, there was mention that this was made easier because there were already other operations that linked separate ALUs together, so it made DP easier to implement.
 
JoshMST said:
Oh man, I haven't been here in ages. Posted the article, then thought... wonder what the B3D fellows are having to say about it all. Then of course you all have 25 pages worth of a topic on RV790...

The PCPer article I wrote was merely speculative, and I of course could be so wrong as to be laughable (when am I not?). Just seems to me that it would be very strange if AMD were to design a new chip that just improved clockspeed and power consumption, especially when looking at their past design history.

As for how double precision is handled... my understanding is that in the group of 5 stream processors, one was double precision and the other 4 were single precision (plus that group of 5 has the general registers assigned to them and the branch prediction unit). Is my understanding of this incorrect?
Click to expand...

Hi mate, really enjoyed reading your article, especially the beginning history part. You should come around more often! Although I don't share you opinion that RV790 will be anything more than RV770 + 150-200MHz and faster GDDR5. :smile:
 
v_rr stop quoting me on XS. ;)


I stand corrected on Problem 1: 220W is doable with 6 + 8. Case point: GTX 280 (236W TDP?)

But problem 2 is still there. 100Mhz =/=/=/=/= 50W Delta.

I do not buy the no extra functional units argument. It could be 800SPs, ALU power is sufficient (for now- p/s @Jawed, did efficiency drop by a lot for the ALUs going from 64 to 160 4+1s?) What about TUs?
 
Thanks Homer. You are most probably right about the chip (I haven't heard anything official or unofficial one way or the other), but it seems the likeliest scenario. Then again, how much fun is it to sit down for a while and speculate on a clockspeed jump?
 
JoshMST said:
Thanks Homer. You are most probably right about the chip (I haven't heard anything official or unofficial one way or the other), but it seems the likeliest scenario. Then again, how much fun is it to sit down for a while and speculate on a clockspeed jump?
Click to expand...

Not much fun I know, but I hate getting all hyped up and being let down when the real specs come out.

The only architectural changes I might expect are the removal of the SidePort and something hoom proposed earlier in this thread, a rebalance to bring the rest of the chip up to spec with the extra couple of SIMDs they (supposedly) added when they discovered they were pad limited. Though TBH I don't expect even that. Not to say I'm not excited about RV790, especially with its potential for AIB overclocking.
 
no-X said:
The same trick as for NV47? Reportedly canceled, but in fact renamed and launched as new generation...
Click to expand...
I refuse to believe that. Even nVidia is not so stupid as to release a *fourth* generation based on the same architecture (more or less). DX11 is around the corner and nV wants to claim the technology leader title.
 
OBR is just an attention whore. A POS attention whore. He newer had the card, not this time, neither the last 2 times (that I know of) he crapped all over the good folks at XS. I got screwed over before RV770 launch along with tons other people, never again for me, thanks. If you want to still take his posts without a huge mountain of salt, go ahead, it's your time.
 
Considering how much time thay had with rv790 and how long it is expected to be on the market as a leader, I'd say sideport will be there.
 
JoshMST said:
As for how double precision is handled... my understanding is that in the group of 5 stream processors, one was double precision and the other 4 were single precision (plus that group of 5 has the general registers assigned to them and the branch prediction unit). Is my understanding of this incorrect?
Click to expand...
The four lanes, x, y, z and w, are used to compute double-precision, while the transcendental lane, t, doesn't take part. t can be used at the same time as the double-precision computation for anything else, subject to operand bandwidth restrictions.

For MUL and MAD x,y,z,w produce a single result each clock.

For ADD a pair of lanes is used to produce a single result. This means two results can be produced per clock, A+B using lanes x,y and C+D using lanes z,w.

There aren't any double-precision transcendental operations. This means a double-precision divide is, for example, a rather drawn out process of 12 instruction clocks.

Jawed
 
INKster said:
Or..., one or more of the GT21x GPU's... ;)
Click to expand...

lol... you made my day. Do you believe in what you say?

Let´s see:
- Cebit = massive desktop and mobile G9x renamings -> Nvidia thinking-"We have here many GT200 40nm chips to launch in 1 month on April 6, but lets not show nothing and rename the entire lineup and put these G92 on the table glass"
- Cebit = supposed to have GT218 ES on the show but that didn't happened
- Cebit = Daylitech: NVIDIA's attempts to produce a die-shrunk 40nm GT200 chip were "disastrous at best"
GT200 was designed for 65nm. Nvidia took tons of time and respins to make one half node shrink to 55nm.
Now they show a entire new GT200 on a entire new half node (40nm) of a new full-node (45nm) in 1 month to a chip initially designed for a 65nm node. There is still missing some days to the April's fool's day ;)

AnarchX said:
Whats about a GTX 260-216 with higher clocks or a half GTX 295 with 240SPs?
Click to expand...
half GTX 295 with 240SPs? You are better then Nvidia on renaming :LOL:
PS: Search for GeForce GTX280, GeForce GTX285

In case you didn´t know there are plently of GTX260-216 OC,XXX,PSC,watever in the market. HD4870 1GB OC,xxx,PCS,watever is more then enought to them.
So another renamed GT200? I thought we where all tired of that.
 
Last edited by a moderator:
A few days ago on chiphell:
1) RV790XTX version < 1GHz, probably 950Mhz
2) RV740Pro something about 6 layer DDR3 version
3) RV740 without external power cannot wake up unless lower frequency
4) 40nm performance chip clients have been "badly beaten"
and later in thread RV790 has:
a) Die area larger than RV770
b) More pins than RV770

Apparently there is no plan for a RV790x2 product. There is also something about chip being 800 shaders and the memory speed that cant quite make out.

Finally NDA expires on April 8 USA time.
 
Lukfi said:
What the (chip)hell does that wanna say? :unsure:
Click to expand...

That 40nm chips obviously are based on the so called octopus(tm) technology. Unless you throw each chip at least 50 times against the wall it's not going to be soft enough to be usable ROFL :LOL:
 
Lukfi said:
What the (chip)hell does that wanna say? :unsure:
Click to expand...



I'm pretty sure the original expression had something to do with 40nm and minefields... Chiphell users love that. Site's down for Discuz upgrade now so can't crosscheck.

Landmine = bad revision (similar to that), needs another metal spin to reach clocks or expected power draw levels.

R600 on 80nm would be akin to a "landmine"s, RV6XXs at 65nm the same too, but mostly not due to the process itself.
670 avoided that despite being the first on 55nm.

740 mined at least once, hence the longer than expected delay since tapeout.
 
larger die area with a 6 and 8 pin connector. It's going to be interesting finding out what they tweaked and by how much since the shaders are still at 800. Perhaps a much tighter ALU to TEX ratio and finally something more than 16 pixels per clock.
 
You must log in or register to reply here.
Back
Top