AMD Breaks 1GHz GPU Barrier With Radeon HD 4890

MojoKid writes "AMD announced today that they can lay claim to the world's first 1GHz graphics processor with their ATI Radeon HD 4890 GPU. There's been no formal announcement made about what partners will be selling the 1GHz variant, but AMD does note that Asus, Club 3D, Diamond Multimedia, Force3D, GECUBE, Gigabyte, HIS, MSI, Palit Multimedia, PowerColor, SAPPHIRE, XFX and others are all aligning to release higher performance cards." The new card, says AMD, delivers 1.6 TeraFLOPs of compute power.

Re:AMD CPU too

[LoRdTAW]

Digital Broke that with the DEC Alpha (Was it DEC at that time?). Wasn't popular but it was a desktop CPU for high end workstations.

Re:Why is it harder on GPUs than CPUs?

[KillerBob]

Heat. Because of the form factor, you can't put a massive heatsink on a graphics card, certainly not the kind that you see on high end desktop CPUs.

GPUs are also generally a completely different architecture than a CPU... they're usually massively parallel and optimized for working with enormous matrices, whereas a CPU is significantly more linear in its operation, and generally prefers single variables.

Re:apples to apples

[pshuke]

According to intel [intel.com] it's about 0.04.

Re:Why is it harder on GPUs than CPUs?

[zolf13]

Wide vector processing with "800 stream processing units" (or "pipes" or "cores") - it is hard to put 800 cores in one chip and not to boil the silicon.

Re:Why is it harder on GPUs than CPUs?

[mikael]

You have so much data being churned around. The high end GPU's have 240+ stream processors, compared to a handful for a mobile phone. Then there is the constant punting of video data from the VRAM chips to the LCD screens (width x depth x RGB x bits/channel Hertz. VRAM is like standard RAM memory except there is a special read channel to allow whole rows of memory to be read by the video decoder simultaneously as it is being read/written by the GPU. It would be possible to
raise the clock frequency, but they would need a larger heatsink. If you visit the overclocking websites, you will see some of the custom water cooling systems that they have. Early supercomputers like Cray used Fluorinert [wikipedia.org].

Re:1 TF/s is so 1996

[wjh31]

if you're imagining insane numbers of cards for silly TFlops, the 4TFlop nvidea Tesla has a 1U rack form, so you can shove as many of them as you like in a rack

Re:Not first ??

[Warlord88]

The 1242 MHz speed is the frequency of vertex shaders, not the core speed. Also, 1 GHz is the core speed without overclocking.

Re:Ummmm.....

[mr_mischief]

How about the fact that it runs each instruction on 800 pieces of data at once? This isn't a 1 GHz one, two, four, or even 16-way chip. It's processing up to 800 pieces of data at once, and its clock for doing that ticks every billionth of a second. You're absolutely right, the clock speed by itself means nothing. The clock speed times the amount of work done per clock does mean something. If you raise either without lowering the other, you raise the overall amount of work the chip can do.

And....

[MasseKid]

And it's still slower than a GTX 285 OC edition. Ghz != Preformance. And Nvidia, stop renaming your cards damn it!

Comment removed

[account_deleted]

Comment removed based on user account deletion

Re:It Was Epic

[fbjon]

I would think history says it's pronounced with a hard 'g', specifically greek history.