Graphics & Multimedia

Review: AMD’s ATI HD 4850 GPU

By admin ⋅ July 2, 2008 ⋅  Send this to a friend.. ⋅  Print this article. ⋅ Post a comment ⋅ Add to Technorati Favorites

The Sapphire Radeon HD 4850 you see pictured here has a core GPU clock speed of 625MHz with 512MB of GDDR3 memory clocked at 993MHz.  The GPU is connected via 256-bit memory bus.  The card is cooled by a single slot, copper fan sink.  The drivers reported an idle temperature of around 80′C, and the card was too hot to touch.  Pricing for the card is expected to be set at $199.

Radeon HD 4850 and Radeon HD 4870 go head to head with refreshed and new G92 -and GT200-based GeForces from NVIDIA, and the GPU used to create the new Radeon products is an absolute stormer.

In spite of having diminuitive dimensions for the die and a comparatively small external memory bus width, RV770 punches harder than any other GPU the Graphics Products Group at AMD have seen fit to release.

The RV670 before it, the RV770 is fabricated at TSMC on a 55nm process, which packs its roughly 965 million transistors into a die that’s 16mm per side, for a total area of 260 mm². The chip has grown from the RV670, but not as much as one might expect given its increases in capacity. The RV670 weighed in at an estimated 666 million transistors and was 192 mm².
In AMD’s new GPU is surely dwarfed by Nvidia’s GT200, a 577 mm² behemoth made up of 1.4 billion transistors. But the more significant comparisons may be to Nvidia’s mid-range GPUs. The first of those GPUs, of course, is the G92, a 65nm chip that’s behind all from the GeForce 8800 GT to the GeForce 9800 GTX. That chip measured out, with our shaky ruler, to more or less 18mm per side, or 324 mm².  The second competing GPU from Nvidia is a brand-new entrant, the 55nm die shrink of the G92 that drives the newly announced GeForce 9800 GTX+. The GTX+ chip has the same basic transistor count of 754 million.

The near 960M transistor chip is a svelte 260mm square, and packs a hardly believable 800 scalar shared processors, arranged in R600- and RV670-like SIMD arrays of 80 each, via the same 16 x 5-way arrangement.  Each SIMD array packs a data sampler unit capable of 16 scalar fetches and a quartet of bilaterally filtered results per clock.

Radeon HD 4850 uses 512MiB of GDDR3 (64GB/sec from 256-bit at 2GHz) with a 625MHz clock, giving 1Tflop/sec of FP32 compute, 20Gtexels/sec of INT8 filtering and 40Gpixel/sec Z-only writes.

Performance details include 1/4 speed FP64 compute ability (industry leading rates therefore result), half speed FP16 filtering (1/2 RV670 per sampler per clock) and full speed FP16 colour writes (2x RV670 per ROP per clock).  So the ratios of compute/sample/ROP change compared to the older generation, while sharing the same basic architecture.

The $239 GeForce 9800 GTX+ looks to be a minor waste of NVIDIA’s time, Radeon HD 4850 more than a match, with HD 4870 taking on the $399 and $649 (LOL) GeFroce GTX 260 and GTX 280 and often beating the $649 part, especially when a reasonable level of AA gets turned on.

NVIDIA are in an instantaneous position of having overpriced new hardware almost straight out of the gate, with the $649 asking price of GTX 280 a significantly difficult sell for a gamer.

(1 votes, average: 5 out of 5)

 Loading ...

You might also like...