AMD - Piledriver chitchat

[watercooled]

Like I said it seems it's replacing the stock CnQ driver; that would also explain why forcing a driver re-install reverts it.

He also states in that first post he's now dropped the CnQ mod (essentially what I said I though would happen). Interfering with such things without a good understanding of how they work and/or accounting for modern CPUs could potentially mess up newer power saving features like gating.

[CAT-THE-FIFTH]

This is interesting if true:

http://wccftech.com/amd-desktop-kave...#ixzz2fXJUsDAF

I saw this mentioned on another forum a few days ago,so it seems the news websites have been a bit slow to pcik up on it.

If true Kaveri is going to have a very fast IGP and it seems the memory controller is massively improved. OTH,the information could all be Moose droppings!!

[teppic]

What's the situation with the memory controller in FM2 Piledriver vs Kaveri vs FX Piledriver?

[HalloweenJack]

still DDR3 - but ofc rumour mill says it supports DDR4.... no DDR4 boards though

[teppic]

Rather, what's the difference in performance/bandwidth? I know that FM2 has limitations there, but don't know the details.

[scaryjim]

I believe (correct me if I'm going mad, guys) the maximum theoretical bandwidth on 128bit DDR3-1600 is 25.6GB/s. Compare that to the 72GB/s available on a 7770, and you don't have to be a genius to work out that if Kaveri really does have a massive GPU section, it's going to be horribly bandwidth starved. The existing 384bit GPU in Trinity/Richland is already bandwidth starved (as demonstrated by the well known performance scaling with fast DDR3 modules) so it would make no sense at all to produce a GPU with more than twice the power unless they've done something dramatic to the memory controller to make it *much* more efficient.

OTOH, AMD have previously stated directly that Kaveri will be compatible with both DDR3 and GDDR5 memory, and a 128bit GDDR5 memory controller, allied with 5Gbps+ GDDR5 chips, would be a much better fit for an 832 shader iGPU. I still don't see how they're going to fit that onto a sensibly sized die when they're only moving from 32nm to 28nm production, though - particularly if they want to keep a single die for both desktop and mobile.

If they're planning a move to separate dies for desktop and mobile, I guess it might be possible. But since no-one makes GDDR5 DIMMs I don't really see what benefit it's going to be (unless they bring back something like sideport, I guess?)

[teppic]

Ah, so this just concerns the GPU? I thought there was a memory limitation on the FM2 chipsets for the CPU rather than the GPU component.

Is it just the missing cache on the Piledriver APUs that hold them back in CPU performance over the AM3+ Piledriver CPUs?

[DanceswithUnix]

Ah, so this just concerns the GPU? I thought there was a memory limitation on the FM2 chipsets for the CPU rather than the GPU component.

Is it just the missing cache on the Piledriver APUs that hold them back in CPU performance over the AM3+ Piledriver CPUs?

The memory controller is on the CPU, so chipset isn't involved.

The lack of cache does make quite a difference, check out the difference in benchmarks between the old Athlon X4 and Phenom II cpus at the same clock speed to see the sort of difference.

http://www.anandtech.com/bench/product/105?vs=81

OFC if you use the built-in graphics then that is fighting the cpu cores for access to the ram as well, which makes the lack of cache hurt more.

[teppic]

Yup, I knew it had an integrated memory controller - I was just under the impression there was something up with the FM2 platform regarding memory bandwidth, it must just be for the GPU though.

Is there any reason AMD wouldn't release higher end APUs with cache enabled? Seeing as we're not seeing any new AM3+ CPUs coming yet.

[badass]

I believe (correct me if I'm going mad, guys) the maximum theoretical bandwidth on 128bit DDR3-1600 is 25.6GB/s. Compare that to the 72GB/s available on a 7770, and you don't have to be a genius to work out that if Kaveri really does have a massive GPU section, it's going to be horribly bandwidth starved. The existing 384bit GPU in Trinity/Richland is already bandwidth starved (as demonstrated by the well known performance scaling with fast DDR3 modules) so it would make no sense at all to produce a GPU with more than twice the power unless they've done something dramatic to the memory controller to make it *much* more efficient.

OTOH, AMD have previously stated directly that Kaveri will be compatible with both DDR3 and GDDR5 memory, and a 128bit GDDR5 memory controller, allied with 5Gbps+ GDDR5 chips, would be a much better fit for an 832 shader iGPU. I still don't see how they're going to fit that onto a sensibly sized die when they're only moving from 32nm to 28nm production, though - particularly if they want to keep a single die for both desktop and mobile.

If they're planning a move to separate dies for desktop and mobile, I guess it might be possible. But since no-one makes GDDR5 DIMMs I don't really see what benefit it's going to be (unless they bring back something like sideport, I guess?)

I don't believe they will introduce something sideport like as the socket has more or less the same number of pins. IIRC intel said of Crystalwell that 32MB would have been enough but they wanted the module to outlast that generation so went all out with 128MB. Maybe fo9r a reasonable cost AMD could have 32MB on a separate die in the same module with Kaveri.

I doubt it though. Kaveri will probably have 30-50% more shader/compute power than Richland and will probably be only just about adequately served by its memory interface.

EDIT: maybe with a bit of L3 cache and adequate direct communication between the CPU and GPU part will make up for some of the lack of memory bandwidth of DDR3?

[CAT-THE-FIFTH]

TH tests the FX6300 in a gaming system with a GTX760:

http://www.tomshardware.co.uk/build-...iew-32785.html

The older $650 PC used a Core i3 3220 and a HD7870XT and the older $600 PC used a Core i5 3350P and a HD7850 1GB. The older $800 PC used a Core i5 3570K and an HD7870XT.

The overclocked settings include both the CPU and GPU being overclocked.

[scaryjim]

... Kaveri will probably have 30-50% more shader/compute power than Richland and will probably be only just about adequately served by its memory interface. ...

Even Llano was only barely adequately served by its memory interface, getting performance boosts right up to 1866MHz. Trinity and Richland barely bumped up the graphical power available at all, and only then through the increased efficiency of VLIW4 - they're still barely adequately served by their memory interface. AFAIK, the memory controller doesn't change much on Kaveri. It's not a pretty picture, IMNSHO.

But going back a step (to my concern about die sizes), on Trinity the 384shader IGP took up about half of a 248mm die, so around 120mm. Bonaire's 896 shader GPU at 28nm is only a 160mm die. Now, steamroller definitely brings in more transistors to the CPU section, but it looks like they could actually keep a similar die size to Trinity/Richland and pack a big chunk more GPU cores into Kaveri. At which point we're back to me worrying even more about the GPU bandwidth available....

[CAT-THE-FIFTH]

The thing is though,the shaders are clocked much lower than the HD8670D,as the clockspeeds in the leaked are at around 600MHZ,instead of the 844MHZ for the HD8670D. Also,as I did mention the memory controller seems improved. Just look at the link!

The Kaveri memory controller using 1600MHZ DDR3 RAM is giving around 20% more bandwith than the Trinity memory controller using 2133MHZ DDR3 RAM.

Remember,Intel seems to be get more memory bandwith with the same RAM than AMD,so it could be quite possible AMD has overhauled the memory controller,which together with other improvements means more usable memory bandwith for the IGP.

Also,compute performance is easily going to be doubled over Trinity,even with the clockspeed decrease over the HD8670D. There is not only twice the shaders,but also the fact that GCN is generally better for compute based workloads than VLIW4.

[teppic]

At last a site is recommending the 6300 over the i3. About time.

[scaryjim]

... it could be quite possible AMD has overhauled the memory controller,which together with other improvements means more usable memory bandwith for the IGP. ...

That's true, but even if they've completely optimised the MC and bumped support up to 2400MHz DDR3, the absolute maximum bandwidth would still only be around half that available to a 7770/7750. So having more shaders than a 7770 makes no sense, as you simply can't get enough bandwidth to feed them.

OTOH, using a smaller number of shaders (say 640 or 512) would potentially give them enough space on the silicon to also fit some L3 cache which would alleviate some of the bandwidth problems. Then fast DDR3 + efficient caching might get you somewhere near realising the performance of the shaders (to close to DDR5 7750/7770 performance)...

[CAT-THE-FIFTH]

That's true, but even if they've completely optimised the MC and bumped support up to 2400MHz DDR3, the absolute maximum bandwidth would still only be around half that available to a 7770/7750. So having more shaders than a 7770 makes no sense, as you simply can't get enough bandwidth to feed them.

OTOH, using a smaller number of shaders (say 640 or 512) would potentially give them enough space on the silicon to also fit some L3 cache which would alleviate some of the bandwidth problems. Then fast DDR3 + efficient caching might get you somewhere near realising the performance of the shaders (to close to DDR5 7750/7770 performance)...

OTH,it could also be down to power consumption too. The clockspeed of the shaders can be lower for a similar level of performance. AFAIK,this is what Intel did with Iris Pro to decrease power consumption of the IGP section(excluding the L4 cache of course).

We still don't know how good the GF 28NM bulk process is compared to their 32NM SOI process,and it could be more a case of giving the CPU section more headroom too.

The HD7750,HD7770 and HD7790 are all running at between 800MHZ to 1000Mhz,whereas the supposed Kaveri IGP is running at only 600MHZ. Even with GDDR5,at 600MHZ and 832 shaders,I doubt the IGP will be even HD7770 GDDR5 level,due to the massive difference in clockspeed. I expect the IGP to be between HD7750 GDDR3 and GDDR5 level for games with fast DDR3.

Of course the other aspect is the IGP is going to a massive increase in computer performance over the previous one. For example look at FLACCL:

http://translate.google.com/translat...%2Findex28.php

Both versions of the HD7750 are close together in performance,and the HD6670 is much slower than the HD7770.

Even with bandwidth limitations,if Kaveri does have an 832 shader IGP it is going to one heck of a general purpose CPU.