Hi David,

David Kanter (dkanter@realworldtech.com) on 2/6/11 wrote:
---------------------------
>>>No it wouldn't. Also, as graphics moves on die, you'll find the notion of adequate
>>>graphics perf blown to bits as igp takes up more and more of die area.
>>
>>The hardware doesn't determine what is considered >adequate. The consumer does.
>
>Actually, you're wrong. If your competitors have graphics that are 2X faster within
>the same cost and power envelop, you are at a serious competitive disadvantage.

You're forgetting two things:
1) You need a CPU anyhow.
2) It's not at peak usage all the time.

Exactly these two reasons made dedicated sound hardware practically dissapear, despite potentially lower cost and power consumption. So you haven't proven me wrong at all.

Graphics is a much bigger challenge, I can see that too, but the CPU's efficiency at graphics is going to increase considerably with FMA and gather/scatter. So what makes you so sure it's not going to cross the line which would make IGPs redundant?

Don't forget that the IGP also needs to invest transistors into higher programmability and higher efficiency at complex workloads, which brings it closer to the CPU architecture.

>>>>But for the market which only requires adequate graphics support, a CPU with gather/scatter
>>>>would be efficient enough at graphics to make the IGP redundant.
>>>
>>>No it wouldn't. Also, as graphics moves on die, you'll find the notion of adequate
>>>graphics perf blown to bits as igp takes up more and more of die area.
>>
>>The IGP itself came into life when Intel noted that the >majority of people didn't
>>need a powerful discrete graphics card, but they did need very basic 3D graphics
>>support. IGP's take up more than half of the market.
>
>Supposedly, what happened is that on one of Intel's older chipsets they were pad-limited
>and had some extra die area. So someone thought it would be a good idea to use the die area for graphics.

Even if that fairy tale held any truth, how does that change the fact that there was a major market for very low-end graphics which Intel (and others) tapped into? It's proof there's people who merely require adequate graphics.

It also indicates that for these people the tradeoff made with software rendering (on a CPU with gather/scatter) is likely an attractive one.

>>So the consumer won't pay for an IGP that costs twice as >much, if they don't need it.
>
>First of all, IGPs cost consumers probably under $5.

The cheapest quad-core Sandy Bridge CPU costs 177 $, and we know the IGP of that chip takes about 18% of the die space. And that's *without* taking into account that it's sharing the L3 cache and memory controllers. So no, IGPs do not cost consumers under 5 $.

>Second of all, they substantially
>improve performance and power efficiency over SW rendering. People will definitely
>pay for that. A larger battery could easily run you $30 or more, and increases the weight and size of the notebook.

There's nothing definite about that. The IGP itself could be twice as power efficient by lowering the clock and increasing the number of EUs (or keep everything the same when moving to a new process node). That didn't happen. Ironlake also featured 12 EUs, but when moving to 32 nm Intel introduced a Turbo Mode and even the possibility of overclocking. This indicates there's not much demand for solutions which are more power efficient.

But that's not all. They also introduced Sandy Bridge processors with only 6 EUs. This clearly tells us once again there's a market with very low graphics performance expectations.

Yes, software rendering would be somewhat less efficient, but there's clearly room for that. Typical laptops will have a 60 Watt power consumption during gaming (and real gaming laptops even consume 150 Watt). With a battery capacity of about 80 Wh this means less than an hour and a half of gaming. People accept that while on the move they'll have poor battery life or poor graphics performance. Given that an i7-2820QM has a TDP of 45 Watt (which includes the IGP), there's no reason to panic about the system's battery life. Other CPU intensive tasks will drain the battery just as fast. That doesn't automatically mean people are willing to pay for dedicated hardware for each of these tasks.

Serious gamers should still buy a system with dedicated graphics hardware, but those currently happy with integerated graphics can be better off with software rendering. They're hardly going to game while on the move anyway. It makes the system cheaper, or allows additional cores and clocking things lower so they have more performance at lower power consumption.

>>The demanded performance for adequate graphics support only increases slowly. Slower
>>than the rate at which software rendering performance is increasing. Sandy Bridge
>>introduced AVX, and we're going to see FMA and gather/scatter in future CPUs. The
>>number of cores increases as well. So sooner or later the CPU will overtake the
>>IGP, at which point you can simply remove it to have a >cheaper chip, or replace
>>it with CPU cores to create a more powerful CPU at the >same price.
>
>One could say the same thing about encryption HW. Except it turns out that eventually
>someone put in dedicated hardware to save the power and performance...even though
>encryption is a minuscule amount of the overall work done on modern PCs.

AES-NI is supported by server chips and some of the i7 and i5 desktop/laptop chips. It's not supported by any i3 chip. So clearly there has to be reasonable justification for including it, even for something as minor as AES-NI. That said, if you use e.g. BitLocker your disk performance can lower considerably (especially when the CPU is already occupied doing other tasks) which is obviously unacceptable given how big of a bottleneck it is during everyday use. And the impact would be even bigger with SSDs.

It's fairly generic too, in the sense that it helps a large range of applications (anything performing disk or network accesses). So it shows some similarity with gather/scatter instructions, which also speed up data accesses.

Ironically, gather/scatter would itself also speed up AES. So for desktop CPUs it's probably not worth keeping AES-NI around once gather/scatter is supported.

>In cases where dedicated hardware substantially improves performance, you will
>tend to see it appear. Graphics long ago showed that dedicated hardware was useful,
>and it will continue for a long time forwards.
>
>Keep in mind that Intel has talked about improving performance of IGPs by about
>25X in a couple of generations. I don't believe SW rendering will increase quite as fast.

Do you have a source for that? The only thing I found was slide 15 here: http://download.intel.com/newsroom/kits/idf/2010_fall/pdfs/Day1_IDF_Keynote_Otellini.pdf.. Note how the process technology for the IGP has caught up with the process technology of the CPU. This means it's not going to improve at that pace any more.

As a matter of fact, the IGP is bandwidth limited: http://www.anandtech.com/show/4083/the-sandy-bridge-review-intel-core-i7-2600k-i5-2500k-core-i3-2100-tested/13. That's right, it's at the mercy of bandwidth improvements now. So no, it's definitely not going to become 25x faster in a couple generations.

Software rendering on the other hand is catching up with the IGP. SwiftShader is currenly about a factor 5 behind on it, but by ditching the IGP you can have 50% more cores. And it's not yet using 256-bit operations. And we still have FMA to come. And gather/scatter. I have no doubt such a CPU can match the performance of its IGP (at which point it becomes bandwidth limited).

And that's not all. Such a CPU is extremely versatile. ILP, TLP, DLP, it's a very balanced architecture and it allows developers to create things we currently can hardly imagine. Things which currently are a bad fit for GPGPU, unless the GPU evolves closer to the CPU. So eventually CPU manufacturers would simply sell bigger CPUs with higher bandwidth to the enthousiast market. In the long run the discrete graphics card is doomed to dissapear as well.

Cheers,

Nicolas