4 chips in 6 months reaches its end today, with the launch of the final chip in AMD’s Evergreen stack: Cedar. Cedar, the baby of the family, will be powering AMD’s bottom-tier cards. Today we’re seeing the  the first of what we expect will be a couple of Cedar cards with the launch of the Radeon 5450.

  AMD Radeon HD 5670 AMD Radeon HD 4670 AMD Radeon HD 5450 AMD Radeon HD 4550
Stream Processors 400 320 80 80
Texture Units 20 32 8 8
ROPs 8 8 4 4
Core Clock 775MHz 750MHz 650MHz 600MHz
Memory Clock 1000MHz (4000MHz data rate) GDDR5 1000MHz (2000MHz data rate) GDDR3 800MHz (1600MHz data rate) DDR3 800MHz (1600MHz data rate) DDR3
Memory Bus Width 128-bit 128-bit 64-bit 64-bit
Frame Buffer 1GB / 512MB 1GB / 512MB 1GB / 512MB 1GB / 512MB
Transistor Count 627M 514M 292M 242M
TDP 61W 59W 19.1W 25W
Manufacturing Process TSMC 40nm TSMC 55nm TSMC 40nm TSMC 55nm
Price Point $99 / $119 $60-$90 $49-$59 $35-$55

It should come as little-to-no surprise that Cedar and the Radeon 5450 finally deviate from the rule-of-2 that has marked the difference between the other Evergreen family members. Whereas all of the larger Evergreen chips have effectively been ½ of their bigger sibling, Cedar cuts right to the bone. It’s half as many ROPs as the Redwood-powered Radeon 5670, but 40% of the texturing capacity, and a mere 20% of the shader capacity. As has always been the case for video cards, once you drop below $100 you have to start sacrificing a lot of hardware to meet lower price targets, and Cedar is no different.

For throwing all of those functional units out along with GDDR5 capabilities, Cedar comes in at a slender 292M transistors with a die size of 59mm2.  This is a little less than half the transistor count of Redwood while being a little more than half the die size. In this case the limited reduction in transistor count in spite of the significant reduction in shader capabilities is an excellent example in what makes cutting a design down to budget-levels such a tricky proposition. AMD won’t release a die shot of Cedar (or anything else of Evergreen for that matter) but it’s a safe assumption that most of Cedar is occupied by fixed and semi-fixed units such as the PCIe controller, UVD2.2, and AMD’s fixed function rendering pipeline. AMD can’t scale down any of these units like they can shaders, hence shaders had to take the brunt of the cuts to get a sub-300M transistor chip.

Attached to Cedar is a 64bit memory bus, which as we stated before drops GDDR5 memory support. Instead Cedar will be paired with DDR2 and DDR3 – with today’s launch card being a DDR3 variant clocked at 800MHz. This also makes the 5450 the first card to launch with something other than GDDR5, which has otherwise been paired with everything from the 5970 to the 5670.

Compared to the RV710 chip at the core of the Radeon 4350 and 4550, Cedar and the 5450 are virtually identical to those parts. It has the same number of functional units and the same memory interface running at the same speeds, making it the closest thing yet to a 4000-series card with DX11 + Eyefinity functionality. Cedar is even pin-compatible with RV710, so that manufacturers can drop it in to existing Radeon 4350/4550 designs. And just to put things in perspective, in spite of these similarities Cedar is 50M transistors larger than RV710, which means AMD spent most of their gains from moving to the 40nm process on adding Evergreen family features and getting a slightly smaller chip. This also means that it’s a safe bet that we’ll see AMD double-up on functional units for the next die shrink.

One of the advantages of throwing out so much shader hardware and dropping GDDR5 is that the power usage of the card comes down significantly, playing well in to the low-power nature of budget video cards. AMD specs the 5450 at a mere 19.1W TDP, and an idle power usage of 6.4W. This is more than 2/3rds lower than the 5670. Lower clockspeeds also play a part here, as the 5450 is the lowest clocked 5000-series card yet, at a core clock of 650MHz.

It goes without saying that as a budget card AMD is not targeting hardcore gamers with the 5450, instead the target market is a mix of buyers who need their first real GPU on a tight budget. This means pushing the Radeon 5450’s UVD/HTPC capabilities, Eyefinity, GPGPU acceleration, and it’s significantly improved gaming performance over IGP solutions. AMD is making sure to tag the 5450 as a DX11 card too, but as we’ve already established from our 5670 review, cards this slow are too slow to take advantage of any of DX11’s wonder features – the tessellator is probably going to be the only DX11 feature to see any action on cards of this performance level.

AMD is framing the 5450 as competition for NVIDIA’s bottom-tier GeForce, the GeForce 210. From a power and form factor standpoint this is a good matchup, however the 210 uses an even smaller GPU than Cedar along with DDR2 memory, which means there’s certainly a performance difference but also a pricing difference, since NVIDIA should be able to build the 210 for less. Pricing-wise the 5450 is in competition with the DDR2 GeForce GT 220, the Radeon 4550, and the Radeon 4650, all of which can be found for around the same price if not lower in the case of the 4550.

Meet the 5450
Comments Locked


View All Comments

  • Lifted - Thursday, February 4, 2010 - link

    The first graph on each of the benchmark pages lists a 5670, the second graph lists a 4670. Typo or are you actually using different cards?
  • Ryan Smith - Thursday, February 4, 2010 - link

    It's not a typo. We never ran the 5670 at 1024x768, there was no reason to. It's more than fast enough for at least 1280.

    The 4670 data is from the GT 240 review, which we used 1024 on (because GT 240 couldn't cut the mustard above 1024 at times).
  • 8steve8 - Thursday, February 4, 2010 - link

    should have had the clarkdale igp in there for good measure, if you aren't gaming I'd guess that igp would be the way to go
  • MrSpadge - Thursday, February 4, 2010 - link

    Would have been interesting to compare idle power consumption: Clarkie + IGP vs. Clarkie + 5450.
  • Ryan Smith - Thursday, February 4, 2010 - link

    Testing a Clarkie requires switching out our test rig, so the results wouldn't be directly comparable since it means switching out everything including the processor. Plus Anand we only have a couple of Clarkies, which are currently in use for other projects.

    At this point Clarkie (and any other IGP) is still less than half as fast as 5450.
  • strikeback03 - Thursday, February 4, 2010 - link

    That brings up the point though that with a card this low on the totem pole it might be nice to include a benchmark or two of it paired with similarly low-priced hardware. I understand the reason for generally using the same testbed, but when it is already borderline playable it would be nice to know that it won't get any slower when actually paired with a cheap processor and motherboard.
  • Ryan Smith - Thursday, February 4, 2010 - link

    Testing a Clarkie requires switching out our test rig, so the results wouldn't be directly comparable since it means switching out everything including the processor.

    At this point Clarkie (and any other IGP) is still less than half as fast as 5450.
  • kevinqian - Thursday, February 4, 2010 - link

    Hey Ryan, I'm glad you are the first reviewer to utilize Blaubart's very helpful deinterlacing benchmark. I would just like to note that with ATI, it seems memory bandwidth plays a big part in deinterlacing method as well. For example, the HD 4650 DDR2 can only perform MA deinterlacing, even tho it has the same shaders as the (VA capable) 4670. The only bottleneck there seems to be the DDR2 memory bandwidth. On the other hand, with the HD 4550, though it has DDR3, it is limited to 64bit memory interface, so that seems to be a limiting factor.

    I have an old HD 2600Pro DDR2 AGP. When I OC the memory from 800mhz stock to 1000mhz, VA gets activated by CCC and confirmed in Cheese slices.

    Nvidia's deinterlacing algorithm seem to be less memory intensive as even the GT220 with DDR2 is able to perform VA-like deinterlacing.
  • Ryan Smith - Thursday, February 4, 2010 - link

    Yeah, I've seen the bandwidth idea thrown around. Unfortunately I don't have any additional suitable low-end cards for testing it.
  • ET - Thursday, February 4, 2010 - link

    I think I remember reading that the interpolation of input values in the pixel shader was moved from fixed function units to being done by the shaders.

Log in

Don't have an account? Sign up now