AMD - Piledriver chitchat

[CAT-THE-FIFTH]

It does look too bad TBH,as Jaguar seems to have the same CB 11.5 score as a SB ULV CPU at the same clockspeeds!

[DanceswithUnix]

Kaveri supposed to have an option to use GDDR5 memory:

http://www.brightsideofnews.com/news...ets-gddr5.aspx

Guess that fits in with the PS4 using GDDR5 as well.

Not sure what they are harping on about with limits to memory being so low though. Charlie has tweeted that the PS4 uses a 256bit interface for its GDDR5 memory, so twice the width of Kaveri to provide 8GB of ram. That says to me Kaveri can provide 4GB on a 128 bit interface soon if not now. Not huge, but should be enough for most people.

[CAT-THE-FIFTH]

Kabini or Kaveri dev board picture:

http://www.xbitlabs.com/news/cpu/dis...tion_APUs.html

HD7950 for Mac sighted:

http://www.insanely-great.com/news.php?id=12819

This is interesting:

http://investing.businessweek.com/re...AMD&ticker=AMD

[kompukare]

GDDR5 would be interesting in a laptop or ultrabook but there's no GDDR5 infrastructure like SODIMMs etc. so it would have to be soldered on. And - although it's not something computer users think about - there may even be issues with timing, signals and termination. That is the GDDR5 standards may actually only work if soldered onto a PCB.

In a way, the on-package interposer rumoured for some Haswell SKUs may be lot easier to implement. Although GDDR5 as main memory would a lot faster. Can't help but think that AMD should be putting more effort into their IMC since despite being first AMD's IMC is now being outperformed by Intel.

[CAT-THE-FIFTH]

If you read through some of the earlier bits of this thread,you will already find that they are working with Amkor on interposer technology. It seems the PS4 might be using the same technology.

[watercooled]

GDDR5 is faster than DDR3 in bandwidth, but it actually has higher latency, which could harm performance in some CPU applications.

[DanceswithUnix]

Unless the DDR3 cpu has latency from waiting for the graphics part to clear off the bus, which might happen sooner with more b/w.

Besides, 5 is better than 3 right, so it should sell

Trawling through Charlie's tweet feed was quite interesting today (once you apply the appropriate frothing rant filters). From John Carmack: "I can’t speak freely about PS4, but now that some specs have been made public, I can say that Sony made wise engineering choices." Guess it can run Quake then

[badass]

GDDR5 is faster than DDR3 in bandwidth, but it actually has higher latency, which could harm performance in some CPU applications.

Latency as in more nanoseconds or latency as in cycles

[watercooled]

Nanoseconds AFAIK i.e. higher latency relative to DDR3. Couldn't argue the point though.

[CAT-THE-FIFTH]

The Core i3 2100 is actually slower than even an Athlon II X4 in the new TR game:

http://gamegpu.ru/images/stories/Tes.../TR%20proz.jpg

[kompukare]

The Core i3 2100 is actually slower than even an Athlon II X4 in the new TR game:

http://gamegpu.ru/images/stories/Tes.../TR%20proz.jpg

Yes seems to be a trend in a lot of new games. Looks many threads is getting common which finally means some competition because while AMD may not be able to compete with Intel for single-threaded IPC, they are willing to sell true quads or better for a lot less than Intel and their CPUs are able to overclock too.

Mind you, GameGPU are a bit strange with their CPU scaling because at 1080P the scores are GPU limited (one the get to 100+FPS). Will have to wait for pcgameshardware.de or similar to see their CPU scaling results.

[Terbinator]

That really is an odd result, especially with the gulf in frequency.

[scaryjim]

That's got to be an instruction set thing. In the latest 3DMark an i3-540 is just ahead of my Phenom II X4 905e on physics, and my 905e (2.5GHz, 6MB L3 cache) should be faster than the Athlon II X4 620 (2.6GHz, no L3 cache). Those graphs are showing the i3-530 trailing way behind, when it should be at least level with the Athlon.

[CAT-THE-FIFTH]

It could be more a case that HT does not work well with the engine.

[scaryjim]

Well, that's basically the same thing - some instructions hyperthread well, some don't. The TR engine is clearly using some instruction set that doesn't hyperthread well. Be interesting to know which one though - it's obviously something that Bulldozer/Piledriver is particularly good at, given the FX4100 hammering a Phenom II X4....

[watercooled]

Maybe some high ILP integer?

TBH I think looking for CPU scaling at something like 800x600 is missing the point, I don't really care how it performs there as long as performance at resolutions you actually use is acceptable. People always seem to assume it scales perfectly linearly, which simply isn't the case i.e. CPU A might be ahead of CPU B at low res, but take it up to something reasonable and CPU A starts to fall behind. This has been proven a few times by people who bother to check (and mentioned previously on this thread IIRC).

Ironically, the argument doesn't really work the other way anyway - people seem to assume increasing quality settings/resolution has no impact on CPU load, or as I said above, scales perfectly, which is plain wrong. But assuming for a second it's true, these benchmarks churn out something daft like 250 vs 300fps. Yeah, because you're actually going to hit that, or be able to notice the difference even if you could...