ONE OF THE worst kept secrets of the year, all 14 days of it, is that the ATI 56xx Redwood parts are coming out today. They are more than half of a Juniper 57xx series and have all the goodies that those chips do.
5870/Redwood cards look like this
The features are the same as the larger two cards, 58xx Cypress and 57xx Juniper, 40nm fabrication process, GDDR5, DX11 and extremely low power usage. Depending on the variant, you can also get ‘Eyefinity’, or seamless single surface display on up to four – depending on the particular card – monitors natively supported. Obviously this will be on the higher end variants, not the cheaper versions.
In terms of raw specs, things get a bit more interesting. It has 400 shaders, half those of a 5770 Juniper, but more than the 320 that the outgoing 4670 had. For memory, it has 128 bit wide GDDR5, clocked at 4GHz. Juniper has the same width, but clocks it at 4.8GHz, so Redwood is not going to be memory bound. It has an aggregate 64GBps memory bandwidth versus 76.8GBps for Juniper. Since the 5670 only has half the units to feed there should be some interesting numbers coming when the two are compared.
For raw power, it has 620Gflops in single-precision floating-point, along with 20 texture units for 25.2Gtexels/sec. Backing this up, there are 32 ROPs for a peak fill rate of 6.2 Gpixels/sec, and Z/Stencil rates of 24.8 Gsamples/sec.
For those of you with social lives that haven’t figured it out by now, the clock rates work out to 775MHz, and it packs 627 million transistors, a bit more than half of a Juniper. Redwood doesn’t halve the memory interface, so that is where most of the extra likely went. From a diagrammatic point of view, the Redwood 5670 looks a lot like it’s predecessors.
Block diagram of 5670
The most amazing part is that you have a 600+ million transistor chip running at 775MHz, and it only draws 61W peak, 14W idle, for the entire board. Just two short years ago, the 320 shader 3870 was sucking up more than 100W peak and didn’t deliver anything near the performance of the 5670, much less at the same price.
Speaking of price, the 5670 will retail for $99, but as usual, expect prices to drop as soon as initial demand fades. Given the utter lack of competition, that may take a while, especially given TSMC’s production woes of late. Since it is the only mainstream DX11 card under $100, it should be a brisk seller.
Looking at the competition, you only have the woeful GT240/GT340 for the same $99 MSRP, and that costs more than the older and faster G92 based 9800GT. AMD is claiming that the 5670 will trounce the GT240 by about 20 percent on average, and that looks quite reasonable if you go by shader counts and comparisons of higher end products. Add in DX11 and Eyefinity and the 56xx series should be a clean kill.
We have heard that Nvidia originally expected to position the GT240 against Redwood, but will now end up putting it against Cedar in the upcoming (likely) 55xx and 54xx lines. This means the GT240 prices will crater in short order, collapsing Nvidia margins in the one area where it still has profitable GPUs.
While it may not seem like a big deal on the surface, the Redwood die is around 100mm^2, and the G215 (GT240 chip) is over 130mm^2. Nvidia only has the laughably outclassed GT220 in the same die size space. Nvidia is going to have to price a more expensive to manufacture product cheaper than ATI to meet price-performance targets. This will waterfall down the line to the GT220 and G210. Renamings aside, the company has nothing to compete with until the Fermi GF100 derivatives come out.
Until then, the high end belongs to Cypress, mid-range to Juniper, and the lower end is starting to be eaten up by Redwood. When Cedar comes out in a few weeks, it is game over for Nvidia, top to bottom, desktop and mobile. The Redwood 5670 is in the mainstream where much of the GPU money is made, and ATI has no competition for the time being. Nvidia can’t match the cost, features, or anything else. ATI has a clean kill with this part.S|A
Latest posts by Charlie Demerjian (see all)
- Spin Transfer Technologies talks about their ST-MRAMs - Sep 27, 2016
- ARM adds CMN-600 interconnect and DMC-620 memory controllers - Sep 27, 2016
- Globalfoundries 7nm process isn’t even close to the name - Sep 26, 2016
- ARM upgrades realtime offerings to v8-R and adds Cortex-R52 - Sep 21, 2016
- Everspin and Globalfoundries team up for embedded ST-MRAM - Sep 15, 2016