Hi Brett,

Brett (ggtgp@yahoo.com) on 2/20/11 wrote:
---------------------------
>Nicolas Capens (nicolas.capens@gmail.com) on 2/19/11 wrote:
>---------------------------
>>Brett (ggtgp@yahoo.com) on 2/17/11 wrote:
>>---------------------------
>>>Today yes, that is looking true, but the RAM is moving on chip, bandwidth is about to explode by 1,000 times.
>>>An Intel CPU with 1,000 times the bandwidth would only be twice as fast, the ATI
>>>chip will become a ~hundred times faster.
>>
>>The GPU wouldn't become faster. What you gain in efficiency, you lose in compute
>>density due to adding more storage on-die.
>
>You care about graphics, a nice size frame buffer is only 2 megs, add 2 megs for
>the Z buffer, and 2 megs for a light accumulation buffer and 2 megs for the full screens effect buffer.
>Total 8 megs, just about what the L3 size is today on a CPU.

Even with the correction in your other message, you still missed double buffering and anti-aliasing. This easily adds up to over 100 MB.

That may not be a huge challenge for eDRAM, but the problem is eDRAM is still DRAM. It has a high latency even when it sits right next to your execution cores.

>The difference is that a GPU would tile this RAM per pipeline to multiply the bandwidth,
>and get you the 1000 times bandwidth for 1000 pipelines.

One problem with that is you also need 1000 memory controllers. For comparison, a GTX 580 has 6 x 64-bit memory controllers, which take up considerable die space. Even though eDRAM controllers can be somewhat smaller because they don't need the high power to go off-chip, they still need deep buffers to handle requests for the duration of the latency.

>If you are smart (and ATI/NVidia are) you only need about one meg that is streamed.

What do you mean by streaming? Going off-chip again? How is that different from today, with GPUs having ~1 MB of cache memory?

[snip]

>We have another decade worth of semi-annual GPU performance doublings.

I seriously doubt that. A decade ago, a high-end GeForce 3 consumed 30 Watt. We're almost at 300 Watt now. So to keep scaling at the same pace we'd need 3000 Watt next decade.

And it's not just power consumption which is slowing down the performance scaling:

R600 XT: 700 Mt, 475 GFLOP/s - 0.68
RV790 XT: 959 Mt, 1360 GFLOP/s - 1.42
Cypress XT: 2154 Mt, 2720 GFLOP/s - 1.26
Cayman XT: 2640 Mt, 2703 GFLOP/s - 1.02

See what happens there? First the GFLOPS/Mtransistor went up, and now it's going down again. That's because in the early days (well, actually only 4 years ago) the percentage of die area spent on shader cores was relatively low. Then they aggressively scaled that up so shader cores dominate the die. But now they can't scale the compute density any more since they need the rest of the die to keep the cores fed with data.

What's worse is that the workloads are getting more complex, and there isn't infinite parallelism. So you need more die area for fast storage, and some workloads can only be processed faster by exploiting ILP.

>Bolting scatter/gather onto a CPU going nowhere, cannot compete with the cost/performance
>doublings of graphics pipelines.

Actually gather/scatter is the single biggest thing still separating the CPU and GPU, in terms of effective efficiency at throughput computing. FMA would also bring them closer together. And as the transistor budget grows so does the number of cores. So with the GPU converging closer to the CPU as well, it's not that unlikely that first the IGP will make way for more CPU cores and software rendering, and then bigger CPUs start threatening discrete GPUs...

>>>Todays desktops have 4 CPUs, a decade from now the CPU count will not exceed 8 for most users.
>>
>>Don't be foolish. AMD Bulldozer desktop CPUs will already include 8-core models
>>this year. Haswell will have 8 cores, by default, in 2013. By 2020 you'll only find 8-core CPUs in the museum.
>
>~Half of all desktop PCs are bought by businesses to be used as little more than smart terminals to a database.
>It would be irresponsible to buy even a quad core chip for such a task.

Those systems don't need a powerful GPU either. A small CPU with gather/scatter would suffice for all graphics needs.

Besides, note that there's over an order of magnitude difference in performance between low-and and high-end graphics sold today, whereas it's practically impossible to find any single-core CPU. The reality is that even business systems need reasonable CPU performance to run smoothly. Software is getting ever more sophisticated (it's a competitive world) so it continues to push the CPU.

Cheers,

Nicolas