KOREAN NUMBER TWO memory giant, Hynix has joined the ranks of next-generation GDDR5 producers with its new 1Gbit memory part. It has a declared 2,500MHz frequency and 5Gbps DDR throughput.

The 32-bit wide chip can handle 20GB per second, and, if implemented on a future GeForce G100 or Radeon R680 / R700 card in a 512-bit 16-chip configuration, could give you 2GB of graphics memory with a stunning 320 GB/s memory bandwidth.

Now, do we really need this on a graphics card? Well yes, if you're running 100 fps action games or massive 3-D models on twin 3840x2400 QWUXGA (or QuadHD for the laymen) displays at whatever antialiasing you fancy - although, at these nine-megapixel resolutions, there's little need for any antialiasing as you have to strain your eyes just to see the pixels.

Somehow, this kind of bandwidth and capacity combination fits the GPGPU use far better. Combine a fast GPU running double precision computation at close to a teraflop, with this fast local memory, and a decent wide, low latency HT3 or QPI connection to the CPUs and their main memory, and you got an engine that can sustain far more of that promised peak teraflop performance - putting aside the funny programming models.

Nvidia's supercomputing Tesla GPGPUs and their DAAMIT cousins, shown at the SuperComputing 07 in Reno, would benefit from this kind of RAM. Now, when it comes to computing, we also need to start playing with the graphics memory latency, don't we? We bet the CL numbers for this Hyundai part are north of 20-20-20-something. µ