Fusion die-shot - 2009 Analyst Day

Quite many sites have released their first benchmarks on Zacate:

AnandTech
Tech Report
PC Perspective
Hot Hardware
Legit Reviews

 

It's quite competitive today. But I think CULV SNB will steal AMD's thunder. Again. It's at-least 6 months late.

 

I find it surprising how many sites insists on benchmarking Cinebench and h.264 encoding on a chip which primary focus is slates, netbooks and low power laptops.

I'd like to see multitasking of media playback, heavy browser usage and various Office-like apps instead. - And I would like to see the result in relation to power usage/battery life.

Cheers

 

Sure, but still it is better compared to the brain damaged stunts that these sites pull by testing linpack on smartphones.

 

Cinebench is quite nice though as it illustrates very nicely how much it's faster in singlethreaded tasks compared to atom, but multithreaded a dc atom can almost close the gap with HT (really helps atom a lot).
I think it's a pity though amd lowered the ram speed from ddr3-1333 from the early demo systems to ddr3-1066. I think ddr3-1333 would be good enough for about 15% faster graphics performance, and ddr3-1600 for about 25% better graphics performance (I don't think it would make much of a difference for cpu bound tasks). This is mostly based on looking at HD4350 vs. HD4550 performance (the latter is very close to HD5450, and the only difference between the HD4350 and HD4550 is ddr2-500 vs. ddr3-800). Considering even low-voltage ddr3-1600 is available today easily, this just doesn't sound right, though I don't know the reasons (one site speculating it didn't really validate at ddr-1333).

 

Looking at the PC Perspective tests it gives a pretty nice picture of performance. Basically SU2300 + better than ION graphics with Atom D510 + NM10 power consumption.

This is great for nettops and good enough for entry laptops. Not to forget GPU part of Fusion will gain on importance with time.

I agree it would be nice to have DDR3 1333 option, still maybe some manufactures will use only 1 DIMM but give 1333 setting somewhere in BIOS (as I suppose the reason for gong with 1066MHz on a final platform was due to validation with two DIMMs running)?

 

Heh, absolutely.

My problem is that Cinebench isn't representative of the kind of applications that are likely to be run on Bobcat devices. Cinebench is FP heavy and SSE optimized; Bobcat only has 64-bit wide FP units (ie. two cycles to do 128bit SSEx instructions) and looks comparatively worse, IMHO.

Cheers

 

Well I see that as a synthetic benchmark (illustrating both sse and single/multi-thread performance quite nicely) so I don't really have a problem with it. I agree though it's not quite representative what people are likely to run. I think though quite a few applications nowadays (games included for example) use a fair amount of sse nowadays - obviously AMD felt 64bit SSE units were a good compromise.
FWIW atom's sse units aren't quite 128bit neither. Ok they are but that's mostly for logical ops / shifts / shuffle. It can also do packed float (or int) add in one cycle but multiplier is only 64bit. Packed float (or int) mul has half the throughput to scalar one. And don't even get me started on doubles - atom is a disaster with packed doubles (throughput 1 for scalar add, 1/5 (!) for packed one, 1/2 for scalar mul, 1/9 (!) for packed one), the optimization guide actually recommends just using scalar ones, and seeing those throughput numbers it's easy to see why (if you're wondering why packed doubles are not just half performance of scalar ones, it's apparently because those operations aren't pipelined - ouch!).

 

I would like to see Bobcat core compared clock for clock to first K8 Sledgehammers and K7 in both 256KB and 512KB L2 cache variants! Throw in some PIII as well for good measure because Anand already showed that P4 needs around 3.4GHz to equal two Bobcats

 

Many people will use Bobcat APU as their only computer (millions do with Atom and Intel graphics already), some will use it at their first computer and others will use it as a massive upgrade or a sidegrade from desktop.

So even these kinds of benchmarks feel legit to me.

 

AMD should add to its APU designs the same kind of encoding accelerator as found in SnD asap.

 

If the transcoding on SB is as fast as intel says it is, it doesn't matter if you add a GF110 or Cayman, you'll always lose to SB.

 

But you might win in flexibility and supported formats.

 

Why would that happen if AMD were to add comparable video encode hw?

 

Because it will have UVD and not anything else?

 

Which is what I am asking, why not a "UVE"?

 

Spoiler

That's what I asked but it looks like I was not clear enough

 

they have AFAIK:

http://blogs.amd.com/fusion/2010/09/27/part-the-clouds-ii-cloudy-with-a-chance-of-high-definition/

 

Good news I hope that software support will follow tho.

 

Is there a main fusion thread or is this it?

Anyway, I made this post on S|A just now.

http://www.semiaccurate.com/forums/showpost.php?p=83713&postcount=106

Is there anything at all to suggest Zacate is or will be seriously struggling with memory bandwidth or is Anand scaremongering or what?