Haswell vs Kaveri

[Deleted member 13524]

Does this mean that Kaveri will not have any form of memory bandwidth enhancement like eDRAM, additional DDR3 channels or dedicated GDDR5 memory in the motherboard?

The picture came from this presentation.

AMD is going to stick a 512 GCN shaders, ~1TFLOPs iGPU to some 25GB/s of memory bandwidth shared with 4 CPU cores?
That'll be a bit depressing IMO. Just look at the difference between the 512 GCN Cape Verde DDR3 vs. GDDR5:

It'll be even worse if the memory needs to be shared with the CPUs, no matter how you put it.

 

[jimbo75]

Would be ~40 GB/s with 2600 MHz RAM, which is a lot better than the ~24 GB/s on the card in that review. But yeah, still it's not wonderful but it's what been expected for a while I think.

 

[Alexko]

Custom desktop builds may get DDR3-2400, but aren't likely to get official support for faster RAM. And I'd be surprised to see many laptops with anything faster than DDR3-1866 (even that may be optimistic), i.e. just under 30 GB/s.

However, this only means that Kaveri will be compatible with regular DDR3. It doesn't mean it won't accept any other kind of memory, possibly with specifically-designed motherboards.

 

[Deleted member 13524]

Would be ~40 GB/s with 2600 MHz RAM, which is a lot better than the ~24 GB/s on the card in that review. But yeah, still it's not wonderful but it's what been expected for a while I think.

So people are expected to pay $75 more for the RAM in order to increase the performance of an iGPU which costs $65 in its higher performance version?

I thought one of the biggest points of the APU was to be cost-effective.

 

[swaaye]

What's really neat are the notebooks that ship with a single DIMM (ie single channel). My dad has a Trinity notebook like that. Though to be fair you don't notice doing typical desktop stuff. It needs a SSD more than additional memory bandwidth.

 

[onQ]

Does this mean that Kaveri will not have any form of memory bandwidth enhancement like eDRAM, additional DDR3 channels or dedicated GDDR5 memory in the motherboard?

The picture came from this presentation.

AMD is going to stick a 512 GCN shaders, ~1TFLOPs iGPU to some 25GB/s of memory bandwidth shared with 4 CPU cores?
That'll be a bit depressing IMO. Just look at the difference between the 512 GCN Cape Verde DDR3 vs. GDDR5:

It'll be even worse if the memory needs to be shared with the CPUs, no matter how you put it.

Not sure but the PDF also show GDDR5 in the memory hierarchy slide.

 

[OpenGL guy]

So people are expected to pay $75 more for the RAM in order to increase the performance of an iGPU which costs $65 in its higher performance version?

I thought one of the biggest points of the APU was to be cost-effective.

You're comparing 8GB of high speed RAM to a GPU board with 1GB of memory. Do you expect to run your whole system on 1GB of memory including the part the GPU uses?

 

[Homeles]

You're comparing 8GB of high speed RAM to a GPU board with 1GB of memory. Do you expect to run your whole system on 1GB of memory including the part the GPU uses?

That doesn't invalidate his point.

 

[jimbo75]

So people are expected to pay $75 more for the RAM in order to increase the performance of an iGPU which costs $65 in its higher performance version?

I thought one of the biggest points of the APU was to be cost-effective.

It's not really all that much different from what Intel is doing with Iris Pro.

 

[Wynix]

It's not really all that much different from what Intel is doing with Iris Pro.

Iris Pro targets a different market, not a good comparison.

 

[jimbo75]

Iris Pro targets a different market, not a good comparison.

The point is to look good in benchmarks.

 

[OpenGL guy]

That doesn't invalidate his point.

Of course it does. You don't need 8 GB to run your machine, for one thing. Secondly, does the discrete card come with a CPU? Didn't think so.

 

[DuckThor Evil]

Of course it does. You don't need 8 GB to run your machine, for one thing. Secondly, does the discrete card come with a CPU? Didn't think so.

His point was that the graphics portion of the APU is next to useless with only DDR3.

 

[yuri]

Does this mean that Kaveri will not have any form of memory bandwidth enhancement like eDRAM, additional DDR3 channels or dedicated GDDR5 memory in the motherboard?

There are 4 DCTs (DRAM ConTrollers) in some 15h Models 30h-3Fh SKUs. Kaveri having 30h signature is the lowest/first model from this line. So maybe later...

NOTE: if channel = 3, then alias it to 1.
This is because, in F15 M30h, there is support for
4 DCT's, but only 2 are currently included in the architecture.
They are DCT0 and DCT3. But we have read all registers of DCT3
into pvt->csels[1]. So we need to use '1' here to get correct
info. Refer F15 M30h BKDG Section 2.10 and 2.10.3 for clarifications

Linux kernel mailing list

 

[Albuquerque]

Of course it does. You don't need 8 GB to run your machine, for one thing. Secondly, does the discrete card come with a CPU? Didn't think so.

Yeah, I think it's pretty apparent that the contention doesn't surround total quantity of ram, but the actual speed / bandwidth / latency of that ram. Sure, you can "carve off" 2GB of ram for use by the IGP for capacity sake, but that memory bus is still being used by four CPU cores at the same time.

So not only is it DDR3-slow, it's actually even slower.

 

[Blazkowicz]

There can be bandwith savings, if even in regular games the drivers insure there are no unnecessary copies between "system ram" and "GPU ram". So the situation can be less dire than with a regular IGP.

Price of ddr3 increasing is a problem, you're now seeing again ddr3 1333 or 1600 pushed at price points you could have ddr 1866 or even 2133 before but the situation is not so dire.
I've found fast RAM for 15 euros more than the cheapest RAM (ddr3 1600) in my usual go-to store, considering 2x4GB. That's 20% more expensive RAM for 50% more bandwith.

Of course cost still is a problem and AMD sells the 6800K as a crazy high price (internal competition.. they have to get rid of 5800K, and want to push Richland on laptops I guess). So what is the market exactly.. a gamer ought to buy a 750K and 7790 instead (much faster that the 7750), a "multimedia" user who wants a big CPU for dealing with high megapixel pictures and high res video can get an i5 Haswell - i5 4430 plus ddr3 1600 is just slightly more expensive than 6800K plus ddr3 2133 or 2400.

So, it's still for a narrow range of "family computers".
In truth it's quite powerful though for people who will upgrade from old low end stuff.

 

[OpenGL guy]

Yeah, I think it's pretty apparent that the contention doesn't surround total quantity of ram, but the actual speed / bandwidth / latency of that ram. Sure, you can "carve off" 2GB of ram for use by the IGP for capacity sake, but that memory bus is still being used by four CPU cores at the same time.

So not only is it DDR3-slow, it's actually even slower.

Wow, you guys are really grasping at straws here. APUs are cheaper for the following reasons:
- No reason to buy a discrete graphics card with its own memory
- Shares power with the CPU
- Shares memory with the CPU

Since you don't need a discrete card for graphics, this allows for smaller form factors as well as other cost savings (power supply, board connectors, case, etc.).

APUs are targeted towards a cost-conscious and power-conscious market segment, if you want higher performance, there are plenty of discrete card options.

BTW, have you actually analyzed how much the CPU degrades memory performance? There's a reason CPUs have large caches.

 

[Albuquerque]

Wow, you guys are really grasping at straws here. APUs are cheaper for the following reasons... <snip>.

I don't think anyone is suggesting otherwise. Your minor tirade about how APU's are supposed to be cheap and easy is a strawman to the conversation at hand. That, or else you're still missing the point entirely and are simply unaware of it. The contention is, again, what good a larger more capable GPU will do in an APU design that is already provably choked by memory bandwidth contention.

Yes, I see you saying that APU's are destined for cheaper machines with a different set of tradeoffs.

Yes, I see you saying that APU's will never perform like a discrete card and people who assume otherwise will be perennially disappointed.

Yes, I see telling us that maybe we don't understand the amount of bandwidth required by a modern CPU.

None of these things are on topic, they are all tangential to the core question. Why would AMD snap in a whole bunch more shading and ROP power, only to link it to a memory subsystem that we can already prove is insufficient? That's the question.

Have any thoughts on that?

 

[jimbo75]

Why would AMD snap in a whole bunch more shading and ROP power, only to link it to a memory subsystem that we can already prove is insufficient? That's the question.

Have any thoughts on that?

Higher performance with lower clock speeds in mobile would be one reason. The fact that they basically can't do anything else is another. They are spending mm2 for power gains, but they need to do it to remain relevant.

 

[Putas]

Why would AMD snap in a whole bunch more shading and ROP power, only to link it to a memory subsystem that we can already prove is insufficient?

Your "proof" assumes that shader workload and graphics bandwidth requirements are constants. And yet quite few things can change.