Hacker News Reader

Comment by frisco | Hacker News Reader

Hold on, hold on a second. Let's put this in perspective. This brand new ATI flagship card has a whopping 24 cores clocked at 880 MHz and a memory bus capable of 5.5 GB/sec. There's no general purpose API here; everything is shaders. Let's compare this to the latest NVIDIA desktop GPU, the GTX 580, capable of both shaders (graphics-native code) and general purpose programming: 512 cores clocked at 1.5 GHz with 192 GB/sec in memory throughput. What? And this isn't even NVIDIA's most powerful card. How does ATI get out of bed in the morning?

View on HN

There is a rational debate to be had about the relative merits of the two architectures and programming ecosystems; this isn't it.

1) A GPU "core" is loosely defined. Those "512 CUDA cores" are 16 streaming multiprocessors with 32-wide SIMT.

2) Big parts of GF100/110's SMs are double-clocked; most of the rest of the chip runs at 750MHz (for the clock rates in your example).

3) 5.5 Gigabits per second per data pin (faster than GF100/110). 256 data pins (fewer). 176 Gigabytes per second per chip (close).

4) OpenCL. There are and have been others, but OpenCL is probably ATI's bet for general purpose programming. You could say it's inferior to CUDA (and I'd agree), but to act like it doesn't exist cheapens the debate.

This does speak to the immense power of marketing though; it's easy to lay down such a thicket of buzzwords that you can spin a product any way you want.

Exactly, it's all how the two companies market their architecture is all, that and branding. Technically speaking, AMD does more in less silicon area but also has to run at higher clock speeds to do so. The power draw is about comparable between similar price/performance ratios from each camp.

I think you need to compare total memory bandwidth to total memory bandwidth, rather than memory chip speed. Skip the diagram and get to:

The memory speed on the 6970 tops out at 5.5Gbps (1375MHz), while the 6950 hums along at 5.0Gbps (1250MHz). That puts memory bandwidth in the 175GB/s and 160GB/s ranges for the Radeon HD 6970 and 6850, respectively

The key numbers there are 175 GB/s and 160 GB/s (not quite as fast as nVidia still, but not 2 orders of magnitude away either). Even 1998's Voodoo 2 had 2.2 GB/sec of memory bandwidth :-)

Reply on news.ycombinator.com