Nicolas Capens (nicolas.capens@gmail.com) on 2/16/11 wrote:
---------------------------
>Hi Seni,
>
>Seni (seniike@hotmail.com) on 2/11/11 wrote:
>---------------------------
>>>>>So instead of a multi-core CPU with an IGP you might as well have a CPU with a couple more cores.
>>>>
>>>>I'm wondering why you think there is a viable market segment that is too cheap
>>>>to pay $5 for an IGP and yet wants more cores.
>>>>Outside of the server market, some workstations, and a few other niches, the extra cores are not all that useful.
>>>>And the things they work for are things that GPUs do too - media transcoding, offline rendering, games, etc.
>>>>Why would someone who wants no GPU want more cores?
>>>
>>>First of all, an IGP does not cost 5 $. A quad-core Sandy Bridge costs at least
>>>177 $ and the IGP takes 18% of die space. That's over 30 $.
>>>
>>
>>You can't just divide and get the cost like that. The actual cost of the die area
>>is quite small - around $20 for the entire Sandy Bridge chip. The rest of the cost goes to:
>>- cost of the fab
>>- package/pins
>>- testing/binning
>>- cost of the process R&D
>>- cost of the chip design
>>- cost of validation & verification
>>- cost of the mask set
>>- cost of writing drivers
>>- cost of marketing programs
>>- and so on.
>
>If this was a valid argument then a Core i3 would cost only 5$ less than a Core
>i7. Clearly things like testing/binning, R&D/design/validation/verification, mask
>set and drivers also affect the IGP cost.
>

Price is not the same as cost.
Intel can charge different prices for products that are almost the same cost.

>Comparing the area really isn't a bad approximation of the real cost. In fact is
>should be higher since it's entirely different testing/binning and R&D/design/validation/verification
>compared to just varying the number of CPU cores.
>
>>The cost of the die area for the IGP really is about $5. But compared to the way
>>they did it before, moving the IGP onto the CPU probably saves more than $5, so really it's free.
>
>Making one thing a little cheaper doesn't mean some other expensive part suddenly
>became free. The total cost for the IGP is a lot higher than the ~5$ of the chip
>cost alone, and you can save a significant part of this total cost. A 6-core CPU
>would be cheaper than a quad-core with IGP.

The IGP's share of the other costs (fab buildout, etc.) is not entirely clear.

The fabs are the #1 cost. But how much of that cost can be spread out per chip depends on how many customers you can find to buy one chip each.
If there are lots of customers, smaller chip are gonna cost less.
If there are few customers, you can only utilize the fab fully by making big chips, in which case the extra area per chip has no effect on fab costs.
If you have multiple generations of fabs at once, then the cost of idling a fab is more complicated.
An increase in the area of one product at a new fab may cause another product to be shifted to an older fab that would have been idling.

So, for example, by moving Sandy Bridge's IGP on chip, they consume 32nm capacity at a greater rate and free up some 45nm capacity.
They then use up the extra 45nm capacity by delaying the 32nm shrink of Atom.
Would it be any cheaper if they made IGPs on 45nm and both Atoms and non-IGP Sandy Bridges on 32nm? Hard to say.
And if they did that, but replaced the 45nm IGP with software and idled the 45nm fabs?

The IGP's cost of design, v&v, drivers, and so on is shared between current IGPs and any future products that use the same IGP (e.g. future Atom),
and to some extent future products with next-generation IGPs based on modified versions of the current IGP. (e.g. Ivy Bridge)

I think mask costs are pretty much a constant term + a linear term for chip area.
So the mask is more expensive with an IGP than without, but sublinearly, and if the die area were spent on cores or cache instead, the cost would come out the same.
Package and pincount cost is pretty small for an on-chip IGP, but is dramatically higher for a northbridge IGP or a discrete video card.
Software rendering would probably have comparable pincounts to on-chip IGP, except if it makes inefficient use of bandwidth or power (which it almost certainly would).

>>If they did it your way and didn't design an IGP at all, they could save both
>the $5 and that last little bit of R&D...
>>Except that instead of spending money on IGP design, they have to spend it on software renderer design.
>
>How many developers do you think it takes to write a software renderer? Pixomatic
>was written by two people. Now compare that to the number of specialized people
>you need to perform all of the above described tasks to get an IGP onto the CPU.

>Note that a big strength of the move to software processing is that there are millions
>of software developers *eager* to develop for powerful homogenous chips. There's
>no significant difference in cost to design a game for the GPU versus the CPU. In
>fact today a great deal of time is wasted trying to figure out how to get the GPU
>to do what you want. Note that taking detours also results in lower effective performance.
>
>So in the long run what you'd get with a homogenous CPU with FMA and gather/scatter
>is a highly competitive software ecosystem with lots of new exciting products, at

"Competitive software ecosystem" sounds like you want to ditch the existing software ecosystem and start over.
Is that your plan? Abandon backward compatibility with DirectX and OpenGL?

>a lower cost to the end user. No need to pay Intel for something you migth not use
>much. Instead, you get more powerful hardware for your money.
>

At this point, I doubt I can convince you regarding either the cost or the performance.
I dispute all of the following:
-a lower cost to the end user by $5, if even that? who cares?
-No need to pay Intel and where do the extra cores and gather/scatter units come from? Not Intel?
-something you migth not use much gpus are used more than you think even at the low end. The Vista/Win7 gui uses them. Also a few non-game programs are gpu-accelerated.
-more powerful hardware for your money unlikely, as explained by pretty much everyone else commenting.

>>>Secondly, not having an IGP doesn't have to mean replacing it with CPU cores. You
>>>can also opt to save 30 $ and still enjoy having a fast CPU which can adequately
>>>take care of 3D graphics should you need it.
>>
>>A cheap CPU wouldn't be adequate for texture filtering even if it had 2 more cores.
>
>Yes it would. The HD Graphics 3000 IGP has just 2 samplers. A 6-core CPU with FMA
>and gather/scatter would be like having up to 12 slightly slower samplers.
>

Each sampler does a 4-texel quad at a time, so 2 of them do 8 texels per cycle.
Each bilinear-filtered texel lookup requires 4 reads from texture memory, so that makes 32 reads per cycle.
Each Sandy-Bridge-like core (without gather/scatter) has a dual-banked L1, dual ported TLB, etc. and cannot handle more than 2 loads per cycle.
So you would need at least 16 cores to match the IGP at equal clock frequency. Or 8 cores if you figure a 2:1 difference in clock. Just to keep up with the current somewhat-crappy IGP.

If you beef up the gather unit so it can pull multiple source texels from a single L1 cache access in a single cycle, you could catch up a little.
But this would require a swarm of enormous multiplexers in the middle of the L1D/LSU datapath, which would probably add 2 or 3 cycles to L1 latency and would devastating to all non-graphics code.
Also, it's not clear that other supporting parts of the chip could keep up, in particular, the parts of the LSU and TLB associated with enforcing memory ordering, coherency, memory protection checks and so on.

So basically, what we're looking at is that gather/scatter alone needs to make the cores 4 times faster to hit your target of two cores=1 IGP.
But it can't do it without a whole lot of bulky supporting hardware which is probably not compatible with a fast OOO x86 core.

>Note also that while the IGPs samplers can be either a bottleneck or running idle,
>a unified design can achieve higher peak throughput and only pays for delivered throughput.
>

While this is true the benefit of sharing two hardware pools is at best 2:1, from reclaiming the half of the units that would have been idle.
If the benefit of specializing the hardware is more than 2:1, then unifying is a net loss.

I'm still not entirely convinced that unified shaders were a good idea, but at least in that case, the shader code is nearly identical.
Unifying it further with the texture units or the framebuffer blending units for example, would probably be a mistake.

>>>Last but not least, people who use a discrete graphics would much rather have a
>>>6-core CPU with gather/scatter than a 4-core CPU + IGP.
>>
>>The people who would want to trade away the IGP for two extra cores, they already
>>can. There is a separate product for them - namely 6-core Gulftown.
>
>Gulftown doesn't have AVX, and requires a socket 1366 motherboard. So Gulftown
>sales are not a useful representation of the market for a future homogenous CPU with FMA and gather/scatter.
>

I think you completely missed the point here.

The point is there is are two separate the market segments here served by separate product lines.
Gulftown is high-core-count, high TDP, no IGP product, appropriate for mainly for certain uses. (High-end Desktops, Servers...)
Sandy Bridge is a moderate-core-count, moderate TDP, CPU+IGP product, appropriate for other different uses. (DTR Laptops, IGP Laptops...)

You are complaining that Sandy Bridge has too much IGP and not enough cores, but that's because it's not for you. You are in the other market segment.
When the next generations come out, there will still be high-cores/no IGP product and a IGP-based product with less cores.
That is to say, Gulftown will have a successor, and so will Sandy Bridge.
If you're not in the 4-Core+IGP market segment, then you don't buy a Sandy Bridge. You either get a Gulftown, or you wait for Gulftown's successor.
There's nothing to complain about except that you don't like the staggered timing of the product releases.

>>>>Other than misinformed consumers, I expect that market hole is awfully small.
>>>>Generally, consumers either need graphics for gaming, or need neither graphics
>>>>nor cpu power and would be better off with a netbook-class processor.
>>>>The consumer who buys an eight-core processor to browse the web is a fool.
>>>
>>>If that were a valid argument, then there would also be an awfully small market
>>>for Sandy Bridge as-is.
>>
>>Well, firstly, there are many misinformed consumers.
>
>That's true for any product. Most consumers will buy a bigger car than what they
>actually need for daily commutes.

Look at how car sales changed with the rise and fall of gas prices. SUVs and pickup trucks were selling like hotcakes, and then they weren't. And then they are again. And soon they probably won't be again.
So some people buy a car that's more than they need. But they only do it under certain market conditions that are subject to change.

The thing is, misinformed consumers can go in any direction.
They are unpredictable in the longer term and can change behavior with the whims of marketing & fashion.

If your plan is to make an inferior or overpriced product and hope people buy it anyway, it might work. Or it might not.
Making a good product and selling it to informed customers is a more reliable plan (unless you fail at marketing it).

To make a good product, you want to match it to the customer's actual needs, not what they think they need.
And their actual needs, in most cases, is either:
-1 they need a both CPU and a GPU, or
-2 they need basic capability at about a Celeron+IGP level.

In neither case is scrapping the IGP to pack in more cores going to actually help them.
They may not know that yet themselves, but there's a chance that they will catch on. What do you do then?

>So call it misinformation or whatever, people
>will buy the system with the highest performance for the money, even if they could
>do with a little less performance most of the time.
>

They do now. Will it continue? For how long? Who can say for sure?
All it would take is a strong ad campaign from someone people blindly listen too (Apple maybe?) to tell them that multicore CPUs are so uncool, and that some other new thing is the new thing.
Just look at the crash and burn of the Pentium 4/"clockspeed is everything" marketing plan. Suddenly people care about efficiency. The Pentium 4 is now everyone's scapegoat.

>They're still getting a more compelling system. In particular, it should be more
>future proof. History shows that software developers have always come up with ways
>to push the hardware requirements up. The Althon X2 I once paid 500€ for, is now
>a real pain to work on. And my 2500€ laptop with a Core 2 Duo 2.0 GHz is overdue
>for a replacement. And mark my words, multi-core software development is only at
>its infancy. A 6-core CPU is going to look average in just several years time.
>

Right. Performance demand grows over time. But not just demand for cores, but also demand for graphics performance.
You have argued in some of your other posts that you think the demand for graphics performance will saturate, much like the demand for audio performance did.
I don't think this is going to happen anytime soon, given my experience with offline renderers and vast quality gap between them and games.
There is just too much room for further improvement in graphics quality in games.

The wildcard here is the stalled console cycle. This hasn't happened before since the 1970s, IIRC, so it's unclear what effect it will have.
If pc games continue to be mostly console ports, even when there is no new console for 7+ years in a row, it could create a temporary plateau effect.
There might be no popular game demanding enough to need more than an IGP, thus killing the GPU market altogether.

>>Others think before they buy the computer that they will never play games on it.
>>Then they change their minds after it is too late.
>
>It's exactly the same for the CPU. Most still buy a dual-core CPU, based solely
>on the fact that most applications don't take advantage of more cores just yet.
>But they're not taking into account that the software world moves relatively fast
>and they're best off planning ahead 3-5 years.
>

I agree.

>That said, for a desktop system it's never too late to change your mind about gaming
>performance. You can start with a 6-core CPU performing graphics in software, and
>when it turns out you do play more games than anticipated, you can buy a discrete
>graphics card. If instead you started with a CPU+IGP, you'd have wasted money on
>the IGP. And if you need more CPU power, the CPU+IGP solution requires you to upgrade sooner.
>

You assume that GPU upgrades are possible. On desktops, they are. On laptops, they are not.
The desktop market is slowly collapsing.
You don't upgrade laptops. You buy a one, use it for a while, and then buy a whole new one.
This is the direction the whole market is going.


>>>Noone who calls himself a gamer is going to buy a Sandy
>>>Bridge CPU and use the IGP.
>>
>>You might think an IGP would be useless to someone with a proper GPU. It's not.
>>
>>I bought a laptop in 2008 or so, with a Penryn 6M and a Geforce 8800M. It also has an IGP.
>>It offers a low-power mode that is useful when you don't want to waste power using the real GPU.
>
>That's again something you can get from a homogenous CPU as well. When not running
>games, the near-idle power consumption of a modern CPU is lower than that of a hefty
>GPU.
>And if that changes in future designs due to better power gating in the discrete
>GPU, you still got a more powerful CPU while the IGP became useless. Honestly, falling
>back to the IGP because the GPU takes too much power at low load isn't a feat of
>the IGP but poor power management of the GPU. There's no reason it shouldn't be
>able to shut off partially instead of entirely to achieve a good balance. So using
>the IGP instead is just a short-lived band-aid.
>

I basically agree with this. It is a short-lived band-aid. But it is a band-aid that covers a genuine wound, and only costs $5.
It would be better that there be no wound at all, but as it is it's an ok deal.

>I'm happy for you your setup gives you a pracical advantage for the current reality,
>but it's not a long-term argument pro IGPs.
>
>>More importantly it's a backup in case the Geforce dies. Which in fact, it did.
>>So now I use the IGP. If not for the IGP, the system would be unbootable until
>>I got a replacement GPU, which for laptops is a huge pain.
>>Am I glad I spent the extra $5 to get an IGP? Hell yes!
>
>With all due respect that's just a silly argument. It only shows how unreliable
>GPUs really are. Would you buy a car that comes with a free bike, "in case it breaks
>down"? Personally I'd rather buy from a manufacturer who gives me a higher quality
>guarantee. And note this 'bike' isn't actually free.
>

The relative prices and values matter here.
If I buy a known unreliable sports car on account of its speed and general awesomeness, and for just $5 extra, the dealer throws in a perfectly good motorcycle, I'd say that's a good deal.
Well, maybe buying an unreliable sports car at all is not a good deal, but in this case, the options are sorta limited, as all other cars on this market have problems at least equally bad.

With desktops, if your GPU dies, you can just buy a new video card swap it in.
With a laptop, if you GPU dies and you have no IGP, the entire system is pretty much dead, and the only way to fix it is to send it to the manufacturer who hopefully has the spare part.
I think a backup is worth at least $5.

>Seriously, IGP+GPU systems have no future.
>
>>>>Non-consumers are different, as they have non-graphics servers, and non-graphics workstations.
>>>>But there are separate processor lines for servers and whatnot, so they don't really count here.
>>>
>>>It does count for something because designing an IGP has a high R&D cost. In other
>>>words a 6-core CPU can be cheaper than a 4-core CPU + IGP even if the die size is
>>>exactly the same. Furthermore, having different production lines and testing methods
>>>and all that reduces the efficiency so not having to produce CPUs with an IGP not
>>>only saves R&D but also production cost.
>>
>>Having multiple products is necessary anyway (2-core netbook vs. 4-core laptop
>>vs. 6-core workstation vs. 8-core server)
>>The R&D cost can't be reduced unless they completely eliminate all IGPs from their
>entire product line, including Atom.
>
>Yes, multiple products remain necessary, but not another combination of CPUs with
>and without IGP. Reducing the number of product lines, reduces cost per part.
>
>And you appear to imply that Atom uses the same IGP design as Sandy Bridge.

No, what I'm saying is that they are spawned from different generations of the same product lines made by the same hardware design team.
Intel's current netbook IGPs are based on the previous generations of mainstream IGPs.
All of intel's future netbook IGPs will be based on hand-me-down designs of current and future mainstream IGPs.
So, most of the work gets done once, and is shared over time between the two product lines.
If you shut down the mainstream IGP development, you would also lose your netbook IGP development. It's one and the same.
If you keep the netbook IGP development going at all, then you still pay the full development price that you would have paid for both whether you use it twice or not.

So any plan to replace mainstream IGPs with software rendering is workable only if it also works for netbooks.
When do you expect netbooks to have 8+ cores & gather/scatter support?
Right now Atom doesn't even have a dual ported LSU, and most netbooks are still single-core+HT.
They've got a long way to go to catch up with even the terrible IGP they come with.

> That's
>incorrect. Pineview has a GMA 3150, which is the old Direct3D 9 generation GMA 3100
>shrunk to 45 nm. It does vertex processing in software! If all desktop CPUs eliminate
>the IGP and several years later Atom runs out of IGPs to shrink down, Intel can
>still license cheap IGPs from PowerVR. The cost is lower than designing such an

One problem here is that PowerVR IGP drivers are from hell, and any possible fix seems to be blocked by licensing issues that Intel has little control over.
The other problem is having a PowerVR IGP means having an IGP at all, which is what you are arguing against in the first place, isn't it?
Surely if there is cost savings to be had, the netbook market is the ideal place for it.

>IGP in-house since the PowerVR design is licensed to multiple manufacturers. And
>it's definitely cheaper than continuing to design IGPs suited for the desktop market.
>

>So once again ditching the IGP saves money, which translates into more valuable homogenous CPUs for everyone.
>
>Regards,
>
>Nicolas