Hi Brett,

Brett (ggtgp@yahoo.com) on 2/17/11 wrote:
---------------------------
>Nicolas Capens (nicolas.capens@gmail.com) on 2/17/11 wrote:
>---------------------------
>>(*) P.S.: Just in case you still don't see it, the benefit of unification is that
>>a homogenous CPU does not need to have the same compute density as the IGP, to exceed
>>its effective performance. You basically get area for free, which would have otherwise
>>remained largely idle (and you can trade it for power). If you take this into account,
>>plus all three convergence trends, and you can still provide hard proof it won't
>>ever make software rendering a viable alternative to the IGP, I'll bow down.
>
>Die area is not free, and never will be that free, the trends are misleading you.

You missed the point. Die area that you already paid for as a CPU, is free when using it for graphics.

Sandy Bridge's IGP is limited to 41 mm2. If you replaced it with CPU cores and added FMA and gather/scatter support, you'd have 110 mm2 at your disposal, for the same price.

Of course you need the CPU cores for things other than graphics while playing games, but that's only a minor workload.

>ATI can fit 1,000 pipelines in the space that Intel fits 4 CPUs, using the same power and RAM bandwidth.

Could you elaborate? Note that the definition of a GPU core is nowhere near that of a CPU core, so numbers like that don't say anything about theoretical throughput, let alone effective throughput.

Also keep in mind that GPU cores are helpless on their own. So either you get about 41 mm2 to cram in as many GPU cores as you can before the law of diminishing returns kicks in, or you get 110 mm2 and you need it to be able to run the O.S. and any and all application code.

>You argue that ATI has run out of gas, its DRAM bandwidth that is the limit.

I didn't say ATI has run out of gas. I said they can no longer keep increasing the theoretical throughput at the same pace due to Amdahl's Law forcing effective throughput down. So instead they need to start investing a larger amount of die area into latency reducing techniques.

The effective throughput will still increase from one generation to the next, but there's convergence taking place between the GPU and CPU.

>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.

>The ATI chip will start looking like Cell, much easier to program today than an ATI chip.

That would mean the compute density will be lowered. Which is exactly my point.

>Yes, the first Cell chips were hard to program for, better designs are coming.

Improving programmability means investing transistors into things other than raw throughput. Again proving my point.

>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.

Multi-core software development is still in its infancy. Currently the average CPU has only 2 cores, so the gains of investing into multi-core development are relatively small. But when 8-core CPUs are mainstream it can no longer be ignored (*). And when software can take good advantage of 8 cores, it's also going to scale to higher core counts so consumers have an incentive to buy CPUs with more cores. We're moving from coarse-grained task division to fine-grained task division.

That said, increasing the core count isn't the only way to increase CPU performance. FMA will double the theoretical throughput, and AVX is already specified to be extendable to 512-bit and 1024-bit vectors. The die area that goes to control logic is amortized by wider execution units. This is comparable to GPUs really.

>The CPU will be a spec in the corner of a huge ATI/Cell die with 100,000+ pipelines.
>Which will be surrounded by a sea of embedded RAM.

A powerful GPU is helpless without a powerful CPU, unless it becomes more programmable and more latency optimized itself, which costs compute density. Also that "sea of embedded RAM" isn't free either and again lowers the compute density. It's comparable to the CPU's caches (the L3 cache is going to be replaced by more dense memory as well).

So no matter how you look at this, the compute density of the CPU and GPU are converging. It means one day the IGP's cores are similar enough to the CPU cores to just have only CPU cores and do graphics entirely in software.

>This is why AMD bought ATI.

AMD bought ATI because in the short term CPU+IGP solutions will take over the low-end market. They have no choice but to offer competitive products or they'll lose billions more than what they paid to buy ATI.

But while in the short term the IGP kills the low-end GPU, in the longer term the CPU kills the IGP. That's because the CPU's performance increases faster than the graphics needs of people who only play games casually, and because of the convergence taking place.

Cheers,

Nicolas


P.S.: Software is a 300 billion dollar market, and growing. So when a CPU manufacturer produces cheap multi-core CPUs, software developers will take the opportunity to create more competitive software, or innovative and exciting new applications. GPUs which are merely faster are not interesting. Nobody needs Crysis to run at 1000 frames per second. Instead we need more programmability and high effective throughput at complex workloads to create something different. So an architecture which balances DLP, TLP and ILP, will prevail.