I too agree with many of your statements Ian,

(however i was too nervous to come out and disagree with Linus )

Especially regarding the bandwidth required for modern GPUs to do their work.

However i think that the bandwidth issue will be solved pretty quickly once we start seeing decent market penetration of combined CPU / GPU chips.

One - look at the xbox360 - by use of embedded DRAM for exclusive use by the GPU, they are able to get by with half of the external memory bandwidth that the PS3 has. Obviously the 360 doesn't compare very well against current top end GPU's but i think that the concept is solid, and that in the future we will likely see similar implementations in some combined CPU/GPU. My personal hope is that the AMD fusion products will go this route.
It's also important to note that the biggest limitation of the 360 is the size of the embedded DRAM, not the external bandwidth to/from the CPU/GPU. If the DRAM was 20Mb instead of 10, most 360 games would be fine @ 1080p, instead of the 720p they are now.

Two - The way modern games are moving is towards computational complexity, not requiring purely more bandwidth. As modern techniques such as procedural textures and geometry, more complex AI, and more complex rendering systems such as Deferred rendering/Shading, and other Global illumination systems. This will favor computational power of the GPU over memory bandwidth.

I reckon that the market will end up in 3 segments:
Low-end fully integrated systems. 60-70% of market.
Mid-rage CPU with integrated GPU, where GPU has some exclusive cache. 10-20%
Top end systems with ( multiple) discrete GPUs ~~5%


We are already looking for ways to use the extra space + transistors allowed by die shrinks, adding more CPU cache is getting lesser and lesser returns, it seems logical that GPU exclusive cache is an easy way to make of of an extra few million transistors.


- James

Ian Ameline (ian.ameline@gmail.com) on 9/30/10 wrote:
---------------------------
>I knew the day had to come -- Linus has made a post that I cannot agree with at all.
>
>Linus, with graphics, it is at least as much about bandwidth to RAM as it is FLOPS.
>
>GPU's are bandwidth hogs -- they need as much memory bandwidth as they can get.
>GPU users are prepared to sacrifice flexibility (socketed and expandable memory) for this. CPU users are not.
>
>Current CPUs (including SB/IB and BD) have only a fraction of the memory bandwidth needed by a decent GPU.
>
>Unless there is a breakthrough in pin counts or bandwidth per pin, CPUs will continue
>to have a fraction of the bandwidth of *current* GPUs. (such a breakthrough would
>apply to discrete GPUs too, so they would maintain their lead.)
>
>This factor alone means that integrated graphics will only play in the low end of the market.
>
>Yes, the fact that GPUs keep a separate (and non CC) address space from the CPU,
>and the pipe between those spaces is not wide enough or low enough latency for many
>purposes is a real PITA -- one I beat against every day.
>
>But the current integrated gfx from Intel does not run my software at all, and
>while the next generation may (assuming the driver quality improves substantially),
>it will not have sufficient performance to interest most of my paying customers.
>Who are pushing around models with 50 to 100 million polygons with texture datasets
>that are scary huge -- many gigabytes. Integrated gfx helps a little with the latter
>problem, but both ATI and Intel have the same bandwidth limitations -- To make integrated
>gfx acceptable, we'd need to see a 8x increase in memory bandwidth.
>
>I'm not holding my breath.
>
>Aggressive proxying/LOD techniques only get you so far, and when users want to
>push around tens of millions of vertices, or billions of texels -- my app is essentially
>a vertex and texel editor -- at some point you have to do it to the real data, not the proxies.
>
>
>So to paraphrase someone I respect;
>
>You're just full of crap. Ranting against high performance GPUs is just a symptom of how fundamentally wrong you are.
>
>:-)
>
>Go ahead -- tell me I'm full of crap.
>
>-- Ian.
>
>
>Linus Torvalds (torvalds@linux-foundation.org) on 9/29/10 wrote:
>---------------------------
>>mpx (mpx@nomail.pl) on 9/29/10 wrote:
>>>
>>>As I've said it's an industry-wide problem: inability to sell quality products
>>>at reasonable prices to mass consumers. A discrete GPU is a superior product, most
>>>people have money to spend in computer price ranges, yet industry will convince them they only need an Intel IGP.
>>
>>You're on drugs.
>>
>>A discrete GPU is absolutely not a superior product.
>>
>>It's extra expense, it's a slow interconnect, and it's
>>extra power and size that most of the market neither needs
>>nor wants.
>>
>>So your whole basic premise is just f*%^ed up. You think
>>the market is illogical, but you don't see that it's your
>>own opinions that simply don't reflect reality. The market
>>is acting fairly rationally, as you'd expect.
>>
>>Integrating the GPU with the CPU allows for size reduction,
>>power reduction, better communication between the two, and
>>allows for the two to share resources like the cache or
>>the thermal headroom. To top it off, it has some obvious
>>huge manufacturing advantages, which makes it cheaper than
>>the discrete version.
>>
>>Sure, discrete chips allow for a bigger total die size and
>>thermal envelope, and it's why things have traditionally
>>been discrete. It was why floating point units and even
>>memory management units were done as separate chips.
>>
>>But on the whole, the disadvantage of discrete chips are
>>much bigger than the advantages. The fact that you
>>think that discrete graphics is somehow fundamentally a
>>better thing shows how crazy your whole argument is.
>>
>>So you're just full of crap. Ranting against the market
>>is just a symptom of how fundamentally wrong you are.
>>
>>Linus
>