So here is what I learned so far (from these guys - hehehe):
(This is just a collection of rumors circulating the web about rv870 - so for such long post - but i keep adding to it.)
Update: added some updated info from other reports in bold
Radeon "7" Family (Evergreen) is 6 parts:
Trillian = "R800" + either multiple montior or 3-way GPU (could be the 950mhz/4gb R800 or a mcm dual cypress r800 aka 8-way rv830/rv770)
Hemlock = "R800" Dual GPU
Cypress = "RV870" Single GPU
Juniper = "RV840"
Redwood = "RV830"
Cedar = "RV810"
Low end codenames:
Cedar = "RV840"
Juniper = "RV830"
Redwood = "RV810"
Another possibility:
Hemlock, aka r800, aka “5870×2″
Cypress, aka. rv870, aka “5870″
Juniper, aka rv 840, aka “5770″
Redwood, aka rv 830, aka “5670″
Cedar, aka rv 810, aka “5350″ (like how i initially wrote in the first place with "cedar" at the bottom -god this is annoying)
CONFUSION BETWEEN CEDAR/JUNIPER/REDWOOD CODENAMES: They're either going to release a part that scores 12-13k (3/4 cypress) or one that scores 2.2K (1/2 redwood ie 1/8 cypress)... I hope there isn't a part that scores 2200P vantage... well - what is the f*@king point of something so slow in this generation? If they do it like this, it's gonna be like "entry level x 2 = midrange" "midrange x 2 = performance" "performance x 2 = enthus.(rv870)" "enthusiast x 2 = r800" with vantage P scores like "P2200 - P4500 - P9500 - P17000 - P28000" If they do this, there is a huge hole between the 9500P part and the 17000P part! Hello? IMO this reason why Cedar CANT be the rv810 part.
(If cedar is rv840, and sits between cypress & juniper, then it should be score somewhere in the 12k range.
It's score is not given, but it makes sense because cedar would be the part equivalent to the 4850 in the HD4xxx series, and probably fall at a price point somewhere between $349 & $199. Whereas, Juniper and Redwood would be the equivalent of the 4830 and the 4650 respectively. I'm hoping that Redwood is the slowest part. This makes more sense - bah I'm tired...Cedar, call it HD5850, should have specs somewhere between Cypress & Juniper and have a die size of somewhere around ~220mm^2)
Manhattan (mobile version of Evergreen Family) line-up:
Broadway XT -> HD5870 GDDR5 45W-60W
Broadway Pro -> HD5850 GDDR5 30W-40W
Broadway LP -> HD5830 (G) DDR3 29W 128bit
Madison Pro (or XT .. not sure) -> HD5750 GDDR5 20-30W
Madison XT (or Pro .... not sure) -> HD5730 (G) DDR3/GDDR5 20-25W
Madison LP / Pro -> HD5650 (G) DDR3 15-20W 128-bit
Park XT -> HD5470 GDDR5 12-15W
Park Pro -> HD5450 (G) DDR3 10-12W
Park LP -> HD5430 (G) DDR3 <8W
-Maybe the desktop parts will be given a similar branding...
Performance:
trillian --~200%+
hemlock -180-200%
cypress - 100%
cedar --- 75%
juniper -- 50%
redwood -25%
Press date: Sept 10th, 2009 Event loc. San Fran, USA?
Event loc. USS Hornet ACC Alameda, CA USA
NDA expires: Sept 17th. 2009
Hard release: Sept 24th. --> mid Nov. 2009 availability
Price range: $500 (high-end Hemlock) --> $50 (low-end Redwood)
3dmark Vantage Performance score:
Rv870 Cypress ~P16xxx - P17xxx - P18xxx
Rv830 Juniper XT ~P95xx
Rv810 Redwood ~P46xx
http://gathering.tweakers.net/forum/...ages/1362510/0
Channel Wholesale Price est:
5870x2/5890x2 - 400.00 USD - 19 October
5870 - 249.00 USD - 24 Sept
5850 - 189.00 USD - 24 Sept
Other retail estimates for high-end:
830pro GDDR3 1GB 99
830XT GDDR5 1GB 149
870pro GDDR3 1GB 249
870Pro GDDR5 1GB 299
870XT GDDR5 2GB 349
800PRO GDDR5 2GB 499 -> $449
800XT GDDR5 2GB 599 -> $499
800XT GDDR5 4GB 699 -> $549
(not sure i believe gddr3 claims, i think lower clockspeeds)
http://74.125.157.132/translate_c?hl...RYlAbgMfpgTZWg
http://translate.google.com/translat...ews_47469.htmlSupposedly, the yield over 60 percent, nearly two-thirds of those in 40 nanometer technology made DirectX-11-chip meets the requirements of AMD and is fully operational. That was the first 40-nm graphics chip, the RV740 (Radeon HD 4770), yet very different. And Nvidia also have problems.
-Estimates put rv870 @ 150 pcs per wafer, which = ~350mm, and GT300 @ 100 pcs per wafer, which = ~540mm
http://74.125.159.132/translate_c?hl...31Nhvke2loHBdw
http://translate.google.com/translat...CGH-Heft/News/For Evergreen, we are required to enter a good deal of new territory, but it is derived in some material respects from our current DX10.1 architecture. The efficiency of the HD 4000 series makes it a perfect starting point, but DX11 requires a lot of detailed work on the specifications that had to be added in certain places. In addition, the Tessellator and hence Hull and domain shader fairly radical changes were required at the scheduler. Overall I would say that our main scheduler that distributes the work of the ALUs and allocates the resources of the individual steps, the component, which required the biggest overhaul. Overall, he is more than twice as complex as in previous generations. When you see the final specifications, you will find that we have other pieces of hardware significantly inflated, including, inter alia...
-All parts have native HD audio support
http://www.hardware.info/en-US/news/..._in_September/A new feature in this generation of GPUs will be the 'dual shader engine' though a lot of speculation remains regarding this technology. It's possible that the word defines a new method of sharing graphics memory between the physical GPUs, which would allow multi-GPU setups to be more efficient.
Architecture explained by a newb:
Dual core / MCM: This generation supports a new "hard" method of dual gpu rendering that differs from previous Crossfire implementation that we see R800 and in rv870 crossfire situations. The chinese translation is "dual core" or split frame hardware- Has something to do with the way the shaders/alus etc operates inside an individual gpu (a method of simultaneous operations in the hardware, similar to dual core, yet not actually two dies in one package ala MCM). Perhaps there is no more real-time compiler in the driver and its all handled on the hardware level by the scheduler. Because the core of the chip is so modular & scalable with certain areas sharing parts of the die (ROPS + Memory controller logic), you are able to divide the specs in half (1600/256/80/32 to 800/128/40/16) and have two parts (rv870/rv830) and rv870 appears as two rv830's, yet it is still only one die. Hence term "dual-core" - more like "modular". Some people think, yes, but all GPUs are multi-core because each of the shaders is like a core by itself. Well, here we are dealing with two large arrays of 800 shaders, along with other standalone logic that communicates with either one of the two identical arrays. More specifically, one Rv870 die is composed of two Rv830-like 160x5 clusters/shader arrays (like two current-gen revamped rv770's in one die) sharing certain features, but connected via "internal crossfire" working in unison and the entire design is a continuation of R600 architecture. It is load balanced, efficient, and requires no crossfire software optimization (because it is hardware level communication); it works via SFR 8*8-32*32, and is bandwidth intensive. The board is using next-gen 5ghz Gddr5 to provide the required bandwidth. So, apparently they've slapped together two 40nm rv770's... so it's easy to see where "dual-core" confusion comes from. Cypress is like a 40nm 4890x2 in one die! -So, it is like a "native dual core" CPU. Now that specs are known for Cypress & Juniper (rv870 & rv830) you can expect that the remaining parts Cedar & Redwood (rv840 & rv810) are 3/4 Cypress & 1/4 Cypress respectively, and that Hemlock (r800) is 2 x Cypress in the same fashion as the HD4870x2 on a single PCB. So, Cypress is like a larabee, except that it uses 2 rv770 cores, whereas larabee is using several P54C cores. (Even though this isn't an actual rv870 die shot, rather an artist's rendition posted earlier - when you do see an actual one, it will look more like the first two images ie. a symmetrical reflection of two identical core areas over a center axis, rather than the third picture - which is an actual die shot of rv770. Notice in the 3rd pic that the rv770 is asymmetrical by design, not resembling a dual core architecture.)
<--- FAKE (photoshopped concept)
<--- CONCEPT "dual core" appearance
<--- REAL (rv770 single core gpu actual die)
more info here:
http://forums.amd.com/forum/messagev...VIEWTMP=Linear
http://translate.google.com/translat....com/index.php
http://forum.beyond3d.com/showthread...=49120&page=72
http://74.125.65.132/translate_c?hl=...K8IDKqx6xzHRww
-Funny, the more i read about this, the more it sounds like dual-core cpus. This could be the first attempt at making a GPU that fits all market segments like modern CPUs. Like how we have Core2duo & PhenomII and 20 sku's of each product with different caches & speeds, but there is really only 1 penryn & 1 deneb, and likewise... 1 cypress.
reference:
Intel Banias single-core cpu -vs- Intel Penryn dual-core cpu -vs- Yet another Rv870 concept drawing
Hemlock (R800) dual gpu "HD5870x2/ HD5890x2" is:
<- Hd4850x2
<- HD4870x2 is overlayed for size comparison
larger than 4870x2 single PCB
4.8+ Tflops (can it play crysis? :P)
29.5cm (11.6in) slightly longer than hd4850x2
2 x 2.2bil transistors (wiki says 1.8b x 2)
40nm 353mm^2 die x 2
19mm x 19mm at cost of $34 per die
now saying: 17.8 * 17.2 = 306.2mm ^ 2 est. 306-320mm2
1 6pin + 1 8pin PCI_E power connectors
DX11
3200SP (1600 x 2)
256bit bus x 2
160 Tex map. (80TMU x 2)
64 ROPS 32ROP x 2
1024/2048mb x 2 mem configs 4gb ~5-6ghz GDDR5???
Core clock 100mhz higher than Cypress
~$500 MSRP
3dmark vantage performance:
Extreme mode ~X13000
(2x Cypress) theoretical Performance P34xxx ???
has to be less though - P28-P30k range
http://74.125.65.132/translate_c?hl=...w9Nft0vRSl16FA
Cypress (rv870) single gpu "HD5870XT/Pro/1g/2g" is:
<-- This card might be the Juniper Es shown down below,
it's too fuzzy, i can't tell if its cypress hsf, or juniper hsf
HD2900XT PCB size
10.4" PCB 26.6cm
40nm, ~2.4 Tflops
2.2-2.3bil transistors (wiki says 1.8b) (likely ~2.0b)
353mm^2 die size
19mm x 19mm at cost of $34 per die
now saying: 17.8 * 17.2 = 306.2mm ^ 2 est. 306-320mm2 (likely right between 320-350 ~335mm^2 18.25mmx18.25mm)
2x 6pin PCI_E power connectors (top of card, not side)
2x DVI-D, HDMI, Displayport out with other possible configs.
Digital PWM more advanced than current gen
DX11
ATi Eyefinity
Enhanced UVD2 (marketing for ability to hardware decode simul 1080P streams.)
2nd Generation TeraScale Engine (marketing for "it can do more than 2tflops")
Reported brand confirmed as "HD5000" series.
1600SP (original rumor is 1200SP - thought to be for mid-enthusiast part now)
256bit bus
80TMU
32ROP
1024/2048mb mem configs
SIMD : 20
Shader Clock : 850 ???
Now reports core lower @ 750mhz (probably 3 different speeds for 3 models 650/750/850)
Memory Clock : 1250Mhz or greater samsung GDDR5 high-performance clamshell mode double bandwidth???
Bandwidth : 153Gbps ???
~$350 MSRP - If you plan on buying one of these models, you're going to want the one with the fastest memory, since the 256-bit memory is considered to be the new chips biggest drawback 4x64bit controller theoretical starvation for bandwidth. 2GB @ 5.8ghz is likely the highest-end configuration.
rv870pro 1G: 249
rv870XT 1G: 299
RV870XT 2G: 399
3dmark vantage performance:
Cypress ~P16xxx - P17xxx - P18xxx
aka ~HD4870X2 performance level ~ closer to GTX295 level
Guy reports vantage Extreme X8500 says is more than his x2, however almost 1000 points less than his 295
Gaming performance estimated at 1.8x 4890
So, reports show HD5870 @ X8500 & P17500, which puts it just a hair under GTX295 X8800 & P18300 performance in vantage:
This is quite fast for a single gpu - I hope its performance is the same in actual games.
*people expect single gpu gt300 to be 20% faster than this, but that's even more unsubstantiated :p
potentially an ~30% performance improvement in the same power profile vs. rv770
Cedar (Rv840) "HD5850" (still lots of conflicting info on the differences between rv840 / rv830)
Not much info here, expected to be a neutered cypress
-perhaps cedar (5850) is a down-clocked cypress w/ ddr3 like a repeat of the 4800 series.
1200 shaders
192bit 128/256bit
60TMU?
24ROP?
$199-$249
Juniper/Redwood (rv830/rv810) HD5830, HD5750, HD5770, 5650, ES, XT, Pro?
<--Reports say this is the card shown at Quakecon
<- Hd4770 on top, Rv830 on bottom
<-- Rv830 15.5 x 11.5
http://forums.techpowerup.com/showpo...63&postcount=9
Partial shroud HSF design
and Dual Slot design
Dual DVI-D, Displayport, HDMI
8.5" - 9.15" PCB
40nm
1.15-1.2b transistors
181mm^2 die size (rv830)
14mm x 14mm die 15.38mm x 11.76mm
1 6pin PCI_E power connector(rv830) no external power on rv810
DX11 - Two Crossfire Goldfingers on Rv830
800SP on Redwood
640SP on Juniper
800-960SP on Cedar (possibly 1200 or 1280)
800SP on Juniper
320-480SP on Redwood (sorry I got codenames mixed up when i first wrote this)
128bit bus
40TMU
16ROP
512/1024mb mem samsung/hyundai/hynix, gddr5 configs (no quimonda)
~$149/$99 MSRP
3dmark vantage performance:
Juniper XT ~P95xx
Redwood ~P46xx
About the speed of a HD4870 & HD3870
Possible specs for Cedar/Juniper/Redwood:
Cedar (rv840):
220 mm^2
128bit/256
1200sp
60TMU
24ROP
Juniper (rv830):
181mm^2
128bit
800sp
40TMU
16ROP
Redwood (rv810):
120mm^2
64bit
400sp
20TMU
8ROP
-All lesser end models (Juniper, Redwood, Cedar) (or Cedar->Juniper->Redwood in that order) will be limited by loss of memory bus width, shader count, TMU & ROP count, memory size (possible gddr3 variants), HSF design, ports & features according to their market placement of either performance, mainstream, or entry-level. Not much solid info on these AFAIK, and many of the reports have Juniper/Redwood with mismatched info (like 640/800 SP, but p95xx/p46xx Vantage scores). Also no definite info found about MCM...I also read there may be 480SP and 320SP parts for low-end and that Cypress may have 384bit bus, but that would change up its SP/TMU/ROP counts. Cypress would have to have 4 more memory chips on the back of the pcb - yet unknown
btw- wouldn't the memory pads need to be layed out like this in order to have 384-bit? But who knows... -
Picture of Cypress HSF (Chiphell forum members only):
http://translate.google.com/translat...3D1%26page%3D1
Quick Comparison of Die sizes:
HD4870 cooler 256 mm^2 16*16 / cypress HD5870 cooler xxx mm^2 x*x / R600 die HD2900xt 420 mm^2 20.5*20.5
-Vantage performance Comparison: (The efficiency of Evergreens performance will depend applications coding, DX9/DX10/DX10.1/DX11, tesselation, use of compute shader, etc. - vantage scores just relative to older series cards)
Hemlock ------~P28k-P30k (4870 quadfire, gtx295 quad sli, or better)
Cypress XT2--- P18,xxx (close to gtx280 sli)
Cypress XT1--- P17,xxx (gtx295)
Cypress Pro---- P16,xxx (hd4870x2)
Cedar ------ ~P12k (GTX285 or better)
Juniper XT ---- P9,5xx (hd4870 1Gb)
Redwood ----- P4,6xx (between hd4670 and hd4830)
TDP estimates: (i think hemlock TDP more like 340w+)
http://www.gamespot.com/pages/unions...union_id=11552Hemlock is R800, possibly branded as 59x0, 300w TDP(6+8pin) uses two RV870s using SFR. Possibly 2 different cards, one PCB codenamed Trilleon(?) estimated MSRP $499-$599. Slightly longer PCB than 4870x2.
Cypress is RV870, 58x0, 225w TDP(dual 6pin), 256bit, 32ROPs, 1600SP, 80TMUs, ~280-340mm2, performance of around a 4870x2/4890CF. 2-3 different cards, 1&2GB versions, MSRP $249, $299 and $349. PCB slightly longer than 4870(9.5in) so maybe 10in, 2 slot cooler w/ 2 DVI, a HDMI and DisplayPort weird half exhaust in the back.
Cedar is RV840, supposedly a cancelled 192bit, 960SP part.
Juniper is RV830, 150w TDP(6pin), 128bit, 16ROPs, 800SP, 40TMUs, 181mm2, performance between a 4870 and 4890, closer to a 4890. 8.5in PCB(?) single slot cooler reference(?) OC'ed/higher clocked dual slot cooler(?) 1GB. MSRP $149 and $199.
Talk of a lower binned 640SP part to replace RV740, filling the $89-$129 market(?).
Redwood is RV810, 75w TDP, unknown specs- 64bit, 320-480SPs, 24TMUs, 80-90mm2. Performance close to a 9800GT, less than a 4830(?). MSRP ~$79 and under. Low profile version(?).
-Should be a low end rv810 w/ ~40w TDP as well.
From wiki R800 article:
http://en.wikipedia.org/wiki/Compari...ocessing_units
-Sorry if some of the info is conflicting - it all comes from different places on teh webz.
I'll laugh if all this speculation ends up being bogus and like RV870 = MCM rv840, hemlock is the low-end, trillian is a 4gpu dual mcm monster with 384bit bus, and the whole thing is called the Radeon7 series,etc.... But I guess we find out on Sept. 10. If it's true I want a job at one of the rumor mill websites, because I enjoy this.
Bookmarks