AMD - Piledriver chitchat

[watercooled]

I wonder why we're seeing far less FCAT measurements now? http://semiaccurate.com/forums/showp...6&postcount=19

[Noxvayl]

What are FCAT measurements? The graphs clearly show AMD doing better, assuming less spikes means better, but I can't make sense of the numbers with the graphs labeled in another language.

[KeyboardDemon]

The graphs appear to show render times per frame, I googled 'lagre ar bettre' from the vertical axis on the graph and it means 'lower is better' (it's in Swedish). It does appear that AMD cards have the most consistent draw rates as there is a much smaller spread which should translate to smoother playback.

[kalniel]

There are plenty at sites like techreport which show mixed results - some titles are ok on the AMD cards, but most are still worse.

[watercooled]

What FCAT aims to do is, rather than simply counting the amount of frames output per second, measures the consistency of frame rendering i.e. the difference in time from one frame to the next. Consistent time from one frame to the next is desirable and large variance can appear as stutter.

It's generally a more common issue in multiple GPU setups (SLI/Crossfire) as they currently use alternate frame rendering where the GPUs render the frames in turn, but driver misconfiguration can lead to bad pacing so you get e.g. two frames in quick succession followed by a large gap before the next two. An FPS counter wouldn't show this problem but the game could be subjectively worse to play.

Nvidia made a big deal about it and a load of sites popped up comparing SLI to Crossfire, and at the time Crossfire suffered poor frame pacing in some games. However it really didn't take that long for AMD to fix it and in some cases even surpass Nvidia's consistency but relatively few sites seemed to make a big deal about this turnaround.

For the graphs on the link I posted, a thinner line shows better frame consistency so on the games tested there, AMD has a pretty solid win. Vertical position on the graph indicates the reciprocal of FPS i.e. lower on graph= higher FPS, but bearing in mind what I said above.

The point being, a fairly big deal was made about it when it was a good PR tool for Nvidia, but now that's no longer true, we barely hear about it.

@kalniel: That's not what I've seen at all, certainly not about 'most' being worse on AMD? Of course you'll get small differences one way or the other between test runs but I've not seen anything to suggest this is anything but a non-issue for AMD now. Even a very recent TR review seems to agree? http://techreport.com/review/28356/n...rd-reviewed/10

[watercooled]

If anyone's interested in the Broadwell-H die size, although it doesn't seem to have been mentioned in reviews so far it can be found in this Intel document (page 8): http://www.intel.co.uk/content/dam/w...mech-guide.pdf

The main die is 13.7x12.3mm so ~169mm2
And the eDRAM die is 6.8x11.7mm so ~80mm2
The total of those being 249mm2.

[CAT-THE-FIFTH]

Official slides leaked:
http://wccftech.com/amds-6th-generat...sa-10-support/

Up to 15% IPC increase over Kaveri,the IGP has the memory bandwidth compression which Tonga has too and it is HSA compliant. It also does H265 decoding and the 15w models have up to a 45% frequency increase!

The IGP is DX12 too.

Dell has already introduced a number of laptops with it:
http://www.pcworld.com/article/29302...-desktops.html

[kalniel]

@kalniel: That's not what I've seen at all, certainly not about 'most' being worse on AMD? Of course you'll get small differences one way or the other between test runs but I've not seen anything to suggest this is anything but a non-issue for AMD now. Even a very recent TR review seems to agree? http://techreport.com/review/28356/n...rd-reviewed/10

That was the review I was looking at - but you're right, I hadn't done any actual stats, 'most' was just my impression. What seems to be the case is that AMD can do a great job on certain titles that they work on - I think it was far cry 4 that was a great example comparing the driver/game version in between the titan X and the 980ti reviews. But I dislike the way it's so dependant on individual game targeting. I used to dislike nVidia for the same thing, when you had to really carefully match up drivers to games. At least with AMD (and perhaps nVidia are heading this way too) it seems inclusive, ie you just go for the latest one and it contains all the improvements in one, but I wish both teams were a bit inherently efficient rather than relying on manual tweakage - to the detriment of games outside of the mainstream eye.

[CAT-THE-FIFTH]

Official slides leaked:
http://wccftech.com/amds-6th-generat...sa-10-support/

Up to 15% IPC increase over Kaveri,the IGP has the memory bandwidth compression which Tonga has too and it is HSA compliant. It also does H265 decoding and the 15w models have up to a 45% frequency increase!

The IGP is DX12 too.

Dell has already introduced a number of laptops with it:
http://www.pcworld.com/article/29302...-desktops.html
http://www.cnet.com/news/dell-inspirion-amd-carrizo/

An overview of more laptops using the chip:

https://translate.google.co.uk/trans...lse&edit-text=

So far models from Dell,Acer,Asus,HP and Toshiba have been shown.

[CAT-THE-FIFTH]

Someone on Anandtech forums spotted this one:

https://translate.google.co.uk/trans...00p-carrizo%2F

Even under the CB R15 load the clockspeed stays at 3.2GHZ to 3.4GHZ.





The chip is probably is probably set at 35W for this test.

However,the design is more optimised for 15W so there should hopefully be greater gains here,and it will be interesting to see how much power consumption is reduced too.

Laptops are going to be arriving for sale in late June and early July:

http://www.extremetech.com/mobile/207229-207229

Zotac also introducing a SFF PC based on the chip:

http://www.extremetech.com/mobile/207229-207229

[CAT-THE-FIFTH]

3DMark11 scores leaked:

http://www.notebookcheck.net/AMD-FX-....144074.0.html



The score is higher than an HD7770M GDDR5 which is a downclocked HD7750:

http://www.notebookcheck.net/AMD-Rad...M.73855.0.html

It is also massively higher than the FX7600P too:

http://www.notebookcheck.net/AMD-Rad....117371.0.html

[DanceswithUnix]

I notice that Anandtech now hail the new i5/i7 chips as having the fastest integrated graphics available. At quite a price though.

http://www.anandtech.com/show/9320/i...75c-i5-5765c/7

[CAT-THE-FIFTH]

I notice that Anandtech now hail the new i5/i7 chips as having the fastest integrated graphics available. At quite a price though.

http://www.anandtech.com/show/9320/i...75c-i5-5765c/7

In some games like GTAV there is hardly a difference though!!

Having said that if the 3DMark11 scores for the FX8800P hold true,it might not be as far off the mobile Broadwell chips as we originally thought.

[watercooled]

I can't seem to find comparable 3dmark scores for the 5775C, but this document (pdf) claims 2881. So assuming there aren't settings which make the scores incomparable, that's mighty close to the 8800P at 35W!

WRT the 8800P scores, it's impressive that they've surpassed the performance of a discrete GDDR5 GPU.

[scaryjim]

... WRT the 8800P scores, it's impressive that they've surpassed the performance of a discrete GDDR5 GPU.

I dunno, it's essentially the same core hardware as the 7750/7770M but with 2 generational enhancements and process improvements to reduce the power. With 512 shaders at 800MHz it's now core-equivalent to the original desktop 7750, and while it's only got half the bandwidth to play with (128 bit DDR3 2133 vs 128 bit GDDR5 4500) it includes that lovely bandwidth-saving delta colour compression, so it's levelled the playing field a lot.

Of course, the fact that they got all of that onto the same package as a quad-core CPU running at up to 3.4GHz, within a 35W TDP, is *very* impressive. AMD SoCs with stacked dram can't come soon enough IMNSHO (after all, windows tablets are currently pushed out with either 1GB or 2GB of DDR3 - they could easily replicate that with one or two stacks of HBM and the SoC would have access to oodles of bandwidth (although latency might become an issue* with HBM, of course....)

*EDIT: actually, I've just done a bit of reading around according to this SK Hynix presentation (so hardly a biased source ) HBM will lower latency compared to mainstream DDR3 and DDR4....

[watercooled]

Yeah it's the memory bit I was talking about really - we've seen core-equivalent integrated GPUs before but they've generally fell behind the GDDR5 counterparts due to bandwidth limitations.

And I agree, HBM seems like it would be such a brilliant fit for APUs.

I've just had a look at the 5200 Iris review to try to compare the two, however they haven't benchmarked any of the same games on them. But using other processors as reference points, the generational improvement may not be all that huge - Haswell Iris was already ~2x 4770k, and that's in line with what we're seeing with the Broadwell one? The difference being it wasn't in a socketed processor before.