Building A Radeon HD 4890
Architecturally, the RV790 graphics processor is identical to RV770. The vital specs haven’t changed one bit. It’s still a 55 nm component, though transistor count is up just slightly to approximately 959 million transistors (from 956 million). The GPU is still made up of 800 stream processors, 40 texture units, and 16 ROPs. It still sports a 1 GB GDDR5 frame buffer on a 256-bit memory bus, too.
Where it differs most is clock speed—on its core and memory bus. Stock Radeon HD 4870s employed a 750 MHz engine and quad data rate memory running at 900 MHz. This new offering cruises at 850 MHz with 975 MHz GDDR5 memory.
In order to get those elevated frequencies, ATI had to do some work to the GPU’s core. In short, the RV770 consistently had issues clocking beyond a certain point—a fact that was evident in many of our System Builder Marathon overclocking attempts, which generally fell short at the same frequency range.
The company’s engineers went in looking for slow electrical paths and re-wired them in such a way that they wouldn’t inhibit faster frequencies. Physically, the GPU is fractions of a millimeter larger due to additional capacitors that clean up power to the chip. But it remains the same ol’ design popularized mid-2008. As you can see from the table above, clock-for-clock RV770 and RV790 perform nearly-identically clock-for-clock.
| Header Cell - Column 0 | Radeon HD 4870 X2 | Radeon HD 4890 | Radeon HD 4870 | GeForce GTX 285 | GeForce GTX 260 Core 216 |
|---|---|---|---|---|---|
| Manufacturing Process | 55 nm TSMC | 55 nm TSMC | 55 nm TSMC | 55 nm TSMC | 55 nm TSMC |
| SPs | 1,600 (2 x 800) | 800 | 800 | 240 | 216 |
| Core Clock | 750 MHz | 850 MHz | 750 MHz | 648 MHz | 576 MHz |
| Shader Clock | 750 MHz | 850 MHz | 750 MHz | 1,476 MHz | 1,242 MHz |
| Memory Clock | 900 MHz GDDR5 | 975 MHz GDDR5 | 900 MHz GDDR5 | 1,242 MHz GDDR3 | 999 MHz GDDR3 |
| Frame Buffer | 2 x 1 GB | 1 GB | 1 GB / 512 MB | 1 GB | 896 MB |
| Memory Bus Width | 2 x 256-bit | 256-bit | 256-bit | 512-bit | 448-bit |
| ROPs | 2 x 16 | 16 | 16 | 32 | 28 |
| Price | $400 | $249 | ~$180 | $340 | ~$180 |
Not surprisingly, the 4890 card itself is easily mistaken for a Radeon HD 4870. They’re the same length; they both employ dual-slot coolers and the same dual-DVI plus video output configuration. Subtle differences set the two cards apart, giving away the fact that these two boards are indeed based on different GPUs. And despite the slight increase in load power consumption as a result of the 4890’s higher clock speed, ATI still gets away arming the card with two six-pin auxiliary power connectors.
As of right now, ATI doesn’t have plans for a Radeon HD 4890 X2 variant, as it likely wouldn’t offer much more than a Radeon HD 4870 X2. And the chip’s extra full-load power consumption would create additional heat that’d need to be cooled.
Overclocking
The principal benefit from moving from HD 4870 to HD 4890 would, in our minds, be overclocking headroom. Stock-to-stock, you’re looking at a 100 MHz frequency increase. However, right out of the gate, ATI’s board partners will be shipping juiced models running a 50 MHz-faster core clock. According to AMD, the new GPU layout should be capable of going even faster than that.
The driver’s Overdrive sub-routine now offers a maximum frequency of 1 GHz, suggesting ATI is fairly comfortable with its enthusiast customers running at that speed. Rather than push our card that high and run the risk of misrepresenting performance with a hand-picked sample, however, we ran our HIS Radeon HD 4890 Turbo sample at its stock 900/975 MHz speeds and compared it to the reference clocks ATI is officially launching.
Upping the core clocks to 900 MHz is good for gains between five and 10% at 2560x1600. One of these factory-overclocked boards undoubtedly enhances the value of ATI’s Radeon HD 4890 versus the 4870 1 GB. However, there will undoubtedly be an additional price premium over the reference cards, too, softening the worth of that extra performance to some degree.
There's a budget GeForce GPU selling in China that not even Nvidia knew it made — RTX 4010 turns out to be a modified RTX A400 workstation GPU
US to patch loopholes that allow China to buy banned AI GPUs from other countries — new regulations include national quotas on GPU exports and a global licensing system
