Update: AMD Quietly Releases The OEM-Only Radeon HD 7000 Series: VLIW5 Rides Again
by Ryan Smith on January 5, 2012 11:04 PM ESTEarlier today one of our news editors, Andrew Cunningham, came to me with an interesting problem. HP was launching their gaming-oriented Phoenix desktop, and the spec sheet listed a video card he had never seen before: the Radeon HD 7670. Initially thinking it was a typo on HP’s part we did some digging, and after finding a well-hidden product page we had our answer. It’s Turks.
AMD Radeon HD 7670 | AMD Radeon HD 6670 | |
Stream Processors | 480 | 480 |
Texture Units | 24 | 24 |
ROPs | 8 | 8 |
Core Clock | 800MHz | 800MHz |
Memory Clock | 1GHz (4GHz data rate) GDDR5 | 1GHz (4GHz data rate) GDDR5 |
Memory Bus Width | 128-bit | 128-bit |
VRAM | 512MB/1GB | 1GB |
FP64 | N/A | N/A |
Transistor Count | N/A | 716M |
TDP | N/A | 66W |
Architecture | VLIW5 | VLIW5 |
Manufacturing Process | N/A | TSMC 40nm |
There are days I’d like to think that it’s the OEM market that’s the oddball, but the truth of the matter is that it’s the retail discrete market that’s the oddball. The OEM market – mobile and discrete – is in fact the norm, as OEM video card sales absolutely dwarf retail discrete video card sales. So much of what we take for granted with retail cards – well defined specifications and formal product announcements through press releases – simply don’t happen in the OEM market. Instead the OEM market is ambiguous on its best days and secretive at its worst, and as a result OEM-only products rarely get a formal announcement, and you would never know about some products if not for the fact that OEMs list them in their system specifications like HP did in this case. Ultimately because of the ambiguity in the OEM market and the need to push specifications, some of the most ridiculous video cards are launched here.
Technically speaking AMD already diluted the 7000 series last month with the launch of the 7000M series, which saw Turks’ and Caicos’ mobile counterparts reborn as various 7000M products. However as the OEM desktop is usually at least slightly saner than the mobile market we had some hope that the 7000M rebrand wouldn’t catch up to the desktop market, but this was not to be. It typically takes AMD around 6 months to launch a complete product stack so a Turks-like GCN product would be a spring/summer affair, but OEMs don’t want to wait that long, especially with CES right around the corner.
The problem of course with a Turks-based 7000 series product is that there just aren’t a lot of similarities to the other 7000 series products to speak of. Turks and Tahiti (and Cape Verde and Pitcairn) are distinctly different products from a feature set perspective. Southern Islands was the biggest GPU architectural overhaul for AMD in the last 5 years, creating a massive divide between Turks and Southern Islands.
The following is a list of some the important attributes and major features being introduced with Southern Islands. None of which will be available with the Turks based 7670.
- TSMC’s 28nm HKMG process
- Graphics Core Next architecture
- PCI-Express 3.0
- Direct3D 11.1
- Partially Resident Texturing
- Fast HDMI
- Video Codec Engine (fixed function H.264 encoder)
- DDM Audio
- ZeroCore Power
- Anisotropic Filtering Quality Improvements
As it stands Turks is still a fine GPU, but when badged as the 7670 this insane namespace collision makes it very hard to meaningfully differentiate between products. Do you want fast H.264 encoding and Direct3D 11.1 support? Then you want a Radeon HD 7000 series card, but now you have to make sure it’s not a 7670. Turks simply doesn't have enough in common with Southern Islands to justify this kind of a model number. Ultimately the launch of the 7670 brings with it the same problem that most other rebrands do, albeit on a larger scale: it's being sold on the well-earned strength of the 7000 series name but lacks the 7000 series' features.
Long term there will also be the question of whether we'll see the 7670 Turks in the retail market. OEM products sometimes cross over - the 6770 being the most recent example - so the 7670 may not stay OEM-only forever.
Update 01/06/2012: The crew over at Tom’s Hardware found AMD’s complete OEM video card page; the 7670 wasn't alone.
The Turks based 6570 is back as the 7570. The Caicos based 6450 is back as the 7470 and the 7450 (depending on the type of RAM used). And absurdly enough, the 2 year old Cedar based 5450 is back as the 7350. The last one is particularly notable as Cedar is from the Evergreen family, not Northern Islands. So it lacks all the features Northern Islands brought, including DisplayPort 1.2 support, improved anisotropic filtering, UVD3, MLAA, and the improved tessellation unit; all of this being on top of all of the differences between Northern Islands and Southern Islands.
AMD Radeon HD 7570 | AMD Radeon HD 7470 | AMD Radeon HD 7450 | AMD Radeon HD 7350 | |
Stream Processors | 480 | 160 | 160 | 80 |
Texture Units | 24 | 8 | 8 | 8 |
ROPs | 8 | 4 | 4 | 4 |
Core Clock | 650MHz | 750MHz | 625MHz | 400-650MHz |
Memory Clock | 1GHz (4GHz data rate) GDDR5 | 900MHz (3.6GHz data rate) GDDR5 | 800MHz (1.6GHz data rate) DDR3 | 800MHz DDR3/ 400MHz DDR2 |
Memory Bus Width | 128-bit | 64-bit | 64-bit | 64-bit |
VRAM | 512MB/1GB | 512MB/1GB | 512MB/1GB | N/A |
Transistor Count | 716M | 370M | 370M | 292M |
Manufacturing Process | TSMC 40nm | TSMC 40nm | TSMC 40nm | TSMC 40nm |
Architecture | VLIW5 (Turks) | VLIW5 (Caicos) | VLIW5 (Caicos) | VLIW5 (Cedar) |
Meanwhile AMD’s web team appears to have made some mistakes listing specifications, which makes the 7000 series ambiguity even worse. The 7470 and 7570 are listed as having “Video Compression Engine (VCE)” support as a feature. As you may recall, SI introduced the Video Codec Engine (VCE), which is AMD’s fixed function encoder. Turks and Caicos of course have the same shader-based video encoding functionality as the rest of the 6000 and 5000 series, but they do not have a fixed function encoder, so we’re not sure how that ended up there. It’s worth noting that AMD had a similar problem with the 7970, which when initially published touted support for DX10 Super Sample Anti-Aliasing.
33 Comments
View All Comments
piroroadkill - Thursday, January 5, 2012 - link
Just to clarify, this isn't a fanboy comment. I don't care what card people buy.. I don't understand why anyone would.I'm simply saying that NVIDIA is infamous for rebranding the same cards over and over.
Wolfpup - Thursday, January 5, 2012 - link
As is AMD, but I don't remember Nvidia ever releasing some GPUs with a new design and some with an old design using the same model numbers. At least they had the decency to increment the numbers.Well, the Geforce 4 and and 4MX was confusing for a lot of people, since the 4MX was basically a Geforce 2.
StevoLincolnite - Friday, January 6, 2012 - link
Basically the same isn't the same.The Geforce 4 MX was more like a Geforce 2 on steroids.
The Geforce 4 MX advanced over the Geforce 2 in multiple areas like the then-new bandwidth and fill-rate saving technologies and in an improved memory controller.
One of the larger issues with the Geforce 2 was how badly it was memory starved.
This was evident when it was pitted against the technically inferior Radeon 7500 which had HyperZ to save on Fillrate and Bandwidth.
It was also able to hardware decode video which made nVidia comparable to ATI for the first time in that regard, this is something that the Geforce 4 Ti and Geforce 2 lacked.
ATI was also not immune to the "Re-Branding" issue.
Take the Radeon 9600 for instance. It would go on to be the:
Mobility Radeon 9600, 9700, Radeon x300, x600, x1050 with all the other variations like the Radeon 9550, SE and XT and others.
Plus... I remember reading awhile ago that 28nm was going to prove more expensive than 40nm for awhile, so with this in mind... The more cost-conscious segments will probably stagnate for awhile.
And nVidia hasn't exactly been competitive in the low end either as of late which doesn't help things, in fact... After the Geforce 240 GT and GTS they took a step backwards in regards to performance with the Geforce 430, 440 and 530.
Mortius - Friday, January 6, 2012 - link
The Geforce 200 series comes to mind. It was a mixture of G9x and GT2xx based parts For example:GTX280 - GT200
GTS250 - G92b
GTX280M -G92b
GTS250M - GT215
Not only that, the capabilities differed. GT21x parts supported DX10.1, everything else only supported DX10.
Then there is the GT230. It shipped had G92b and G94b variants. The former having twice as many shaders.
It isn't completely dead. The entry level G210 lives on as the G310 and the G405.
bennyg - Friday, January 6, 2012 - link
A complete list to relive the full extent of the chicanery at and around that time. It is even more complex if you include the while 9000 and 9000M series as there's even more overlap.GT230 - mix of 48 shader G94b and 96 shader G92b
GT230M - GT216
GT240 - GT215
GT240M - GT216
GTS240 - OEM only G92 based
GTS250 - G92b
GTS250M - GT215
GTX260 - GT200
GTX260 "Core 216" - mix of 55nm/65nm GT200a/b
GTX260M - G92b
GTX280 - GT200
GTX280M - G92b
GTX285 - GT200b
GTX285M - G92b
GT215 lived on in GT335M/GTS350M/GTS360M.
Radeons have a while to go before they get THAT bad. But they're building up a nice head of steam, with 6970M/6990M parts which arent Cayman-like (i.e. VLIW4) at all, and 68xx parts which are just the same Junipers as 57x0 and 58x0M, and now these "new" 7xxxM parts ...
And I guarantee if Kepler runs late Nvidia will start doing the same thing...
Dribble - Friday, January 6, 2012 - link
In none of those examples did nvidia do an exact replacement at the same model number like 6770 to 7670. That is perhaps a new low.Mortius - Friday, January 6, 2012 - link
What about the following pairings?8800GT/9800GT
GT240 (GDD5)/ GT340
G210/G310
The last one was indeed mentioned earlier.
coldpower27 - Friday, January 6, 2012 - link
Yeah the renaming game is getting annoying I don't mind though if it's only DX10.1 -> DX10DX11 parts though should be named as such..
There is also the 9800GTX+ -> GTS 250
GTX 200 (GT200) Series was still largely the same tech G92 based parts which is still largely the same as G80..
It wasn't until Fermi and it's derivatives that we got a major architectural change...DX11 with GTX 400/500 Series based parts.
I miss the days of clear nomenclature..
Geforce 5 was pure Direct X 9.0 Shader Model 2.0a
Geforce 6/7 was Shader Model 3.0 w/Direct X9.0c
the Direct X 10 gen was too many cards
Geforce 8, 9 and the 200 Series and even some 300 Series for the low end.. WAY too spread out.
At least we have a cleaner break with the Geforce 400/500 Series with Direct X11 Support.
chizow - Saturday, January 7, 2012 - link
I think that's largely why Nvidia's rebadges were justified in those cases, but not not so much with these AMD parts.Nvidia has stayed remarkably consistent with their feature sets over a generation of GPUs, perhaps to a stubborn fault. They resisted DX10.1 on the majority of their parts as it would clearly bifurcate their product line and create inconsistencies and market confusion.
Not to mention their shader architecture today is also largely the same as G80. No small feat considering AMD is going on its 3rd major change in 2 years. So for the most part, the feature set and compatibility on Nvidia parts has remained the same while the price and performance regardless of rebrand has remained intact.
chizow - Saturday, January 7, 2012 - link
8800GT to 9800GT was a die shrink from 65nm to 55nm.The other two parts were OEM only for the short-lived and incomplete 300 series. Technically the rebrand was justified as these were Nvidia's only DX10.1 parts and also the first to support limited HD audio and probably fit better in a 300 series rather than being lumped into the DX10-based 200 series.