AMD - Bulldozer Chitchat

[scaryjim]

Download the developer preview and check?

AFAIK that'll be the default screen, and is the home screen of the new metro UI. Whether it'll be possible to disable it on desktop / laptop systems is (again, AFAIK) undecided, but I find it unlikely that Microsoft won't put in some way of going straight to the conventional desktop, even if it's a registry / settings hack...

[watercooled]

I've been playing with it in a VM, albeit not much because of my extreme dislike of the new UI, and I can't see an obvious way of doing it. And I'm not just picking on Windows, I hate the new Unity interface on Ubuntu. At least they should give a simple option to let the user decide, rather than making the choice for them...

[scaryjim]

haven't looked into unity much (never been a big ubuntu fan) but the little research I've done made me think I might actually quite like it. It looks like it tries to me vaguely useful on both desktop and touchscreen devices, whereas Metro looks like a tablet UI through and through. Still, can't really bad-mouth either of them until I try them, can I Just wish x86 tablets were a bit cheaper...

[watercooled]

Yeah it tries to be useful on both but they're very different markets and IMO it really lets down at least the desktop side. Granted, it's hard to cater to both with a single UI but AFAIK the vast majority of Ubuntu installs will be on a normal computer system.

[CAT-THE-FIFTH]

It seems faster RAM seems to make a difference with some games:

http://www.madshrimps.be/articles/ar...#axzz1l0Lr913D

http://vr-zone.com/articles/amd-fx-8...zer/13704.html

A tad surprised TBH.

[DanceswithUnix]

AMD did state that AM3+ was not the ideal socket for BD, so given the only change they could really make was to memory I guessed something must be up.

My guess is this: With 2 un-ganged memory channels you have a chance that 2 cores at a time can be fetching something from ram. If another core wants a shot at the ram, it has to wait for the core currently using the memory controller to finish its burst. If the memory is faster, then the latency of that waiting time improves.

That would explain the move to 4 channels. Not much need for the increased bandwidth, but if you look at it as the ram equivalent of moving from a 2 disc raid 0 array to a 4 disc array it starts looking more interesting as an accesses per second increase.

I notice talk of the next generation having 10 core parts. Thats a lot of contention for the ram!

[watercooled]

A tad surprised TBH.

Same, it looks like the dual-channel RAM limitation of AM3+ is starting to become a bottleneck. Kudos for keeping backwards compatibility for so long, but beyond a point it looks like it's going to harm performance.

The 3D06 @ 1333MHz looks like an anomaly though.

[watercooled]

My guess is this: With 2 un-ganged memory channels you have a chance that 2 cores at a time can be fetching something from ram. If another core wants a shot at the ram, it has to wait for the core currently using the memory controller to finish its burst. If the memory is faster, then the latency of that waiting time improves.

Ganged vs unganged tests would have been a nice addition to those reviews IMO.

[scaryjim]

It seems faster RAM seems to make a difference with some games:

...

A tad surprised TBH.

the two tested games below ( selected as they are known to scale with extra ram speed )

So the change in RAM speed produces improvements in games that are known to scale with RAM speeds. They go on to say you can do the same thing on Sandy Bridge. So it's not a case of Bulldozer in general being RAM speed limited, but those two particular games being RAM speed limited.

The vr-zone results look pretty inconclusive to me - Crysis 2 performance actually drops from 1600 CL7 to 2133 CL7 and in each case the "improvements" are pretty marginal (unlike the ones in the madshrimps review, but again those titles are cherry picked).

Incidentally, the 3D06 chart is misleading. The actual difference is ~ 2.5%. The chart has been scale-shifted to start at 17,300 which makes the differences *look* much bigger than they are. It's an argument I have almost daily with my boss regarding the visualisation systems we produce...

[shaithis]

Wonder if they will start rolling out LRDIMM to desktops as well then.......

[watercooled]

Incidentally, the 3D06 chart is misleading. The actual difference is ~ 2.5%. The chart has been scale-shifted to start at 17,300 which makes the differences *look* much bigger than they are. It's an argument I have almost daily with my boss regarding the visualisation systems we produce...

Ah sneaky, I didn't notice that. I must have noticed the previous graphs starting at 0 and thought they all did...

Wonder if they will start rolling out LRDIMM to desktops as well then.......

I doubt it, LRDIMMs are more about being able to add more physical DIMMs to a system, if anything they increase latency over standard DIMMs when used in small numbers.

[shaithis]

I was under the impression it also allowed greater speeds?

[DanceswithUnix]

I was under the impression it also allowed greater speeds?

Well it should stop speeds from tanking with large numbers of dimms per channel, but otherwise it introduces more silicon in the data path. That silicon costs time for the data to pass through, hence making latency worse, and is another chip that the average punter isn't going to want to pay for so drives cost up.

[CAT-THE-FIFTH]

A set of slides from the AMD Financial Analyst Day(saw this one another forum):

http://phx.corporate-ir.net/External...F8VHlwZT0z&t=1

The 3DMarkVantage score for the 17W ULV A6 CPU is 2355 and for the 25W ULV A10 CPU is 3600.

The score is close to a 100W A8-3850:

http://www.tomshardware.com/reviews/...o,2975-10.html

[scaryjim]

A set of slides from the AMD Financial Analyst Day ...

So AM3+ lives on as an entry-level server platform? Interesting...

[Biscuit]

A set of slides from the AMD Financial Analyst Day(saw this one another forum):

http://phx.corporate-ir.net/External...F8VHlwZT0z&t=1

The 3DMarkVantage score for the 17W ULV A6 CPU is 2355 and for the 25W ULV A10 CPU is 3600.

The score is close to a 100W A8-3850:

http://www.tomshardware.com/reviews/...o,2975-10.html

Impressive if its genuinely accurate, very impressive!