Haswell vs Kaveri

Almost all of the SA's pieces were linked here and discussed all the time before his site went paid, which was nowhere near once in a blue moon frequency. The arguments made there are relevant and merit discussion, even if they might be hilariously wrong.

If Theo/Fuad wrote the articles linked above, would there be a problem linking to them/discussing them?

I can't see how it can be considered advertising if a third party is summarizing in a line or two, a presumably rather long, paywalled article. Copy-pasting large amounts of text would be another matter, of course.

 

If the links led to actual pieces of information, I'd be fine with it.

The fact that every single link I click for S|A leads me to a request for payment for reading the whole "story" bothers me a lot.

 

The fact that they were linked and discussed at length does not necessarily say good things about us and what we choose to discuss (but there's always room for improvement, one would hope). Yes it would be a problem if suddenly an RSS feed to their work (paywalled or unpaywalled) materialised in this forum. Once in a blue-moon it might be fun to discuss something from Fox News. Once upon a time it might've been super ultra bombastically califragilistically cool to jump after any coin thrown down a well. Hopefully we can grow up from that.

As a sidenote, I find your implication (hopefully I am wrong on this one) that somebody has something against this new business model and suddenly decided to be mean utterly insulting, to be perfectly and bluntly honest. Look at it as yet another step in the attempt to at least diminish the ever increasing stream of noise.

 

nuff said ..

 

Unless there is some way for the GPU to be used as a co-processor by games developers. Thinking along the lines of the GPGPU block in PS4, if Intels integrated GPU's could server the same function in PC's that'd be pretty cool.

I'm guessing there are fairly large blockers to that though.

 

I've learnt at the moment that the regular big Intel chip is GT3 and with the added memory, it's GT3e.

This from the Anandtech report at CES on january 9 (the vs GT650M demo)

With this naming convention, I don't need to whine, GT3 will be very common but it's the GT3e you will seldom see except on high end laptops, maybe low end Mac.

 

Ah, that makes a lot more sense. I could certainly see that getting used in the iMac (assuming it is fast enough) and Mac mini in addition to laptops and potentially tablets.

Regards,
SB

 

It's not a full review yet, but the guys from the Russian OCLab.ru website managed to get the first testing results from the upcoming Intel Haswell architecture. They put the CPU through four popular benchmarks for overclockers: SuperPI 1M, SuperPI 32M, PiFast, Wprime 32M and Wprime 1024M. The CPU used was a Intel Haswell B0 stepping clocked at 2.8GHz and they compared it to an Ivy Bridge clocked at 2.8GHz. The OCLab.ru benchmark results:

Super PI 32M
Haswell: 10 minutes 30 seconds; Ivy Bridge: 10 minutes 51 seconds
Super PI 1M
Haswell: 14,1 seconds; Ivy Bridge: 13,1 seconds
PiFast
Haswell: 24,1 seconds; Ivy Bridge: 25,7 seconds
Wprime 32M
Haswell: 23,7 seconds; Ivy Bridge: 25,7 seconds
Wprime 1024M
Haswell: 7 minutes 10 seconds; Ivy Bridge: 7 minutes 11 seconds

http://www.xtremesystems.org/forums...8-Haswell-First-Benchmark-Performance-Results

Disappointing.

 

Well we kinda knew already that the "tock" is mostly for graphics, this time.

Nonetheless, changing architecture and gaining up to ~8% at the same clocks isn't bad at all.
Look at AMD, for the past 10 years, every time they change architecture they actually lose performance to their predecessor.

 

ToTTenTranz: Isn't H actually 8 % slower then IB in the Super PI 1M test?

 

SPI is sensitive to cache latency - not so much on bandwidth, especially the 1M set . L3 in Haswell is async to the core clock, compared to IVB, so that could be the reason for the slight loss here. Despite the doubled L2 throughput, the overall access latency is slightly up for Haswell throughout the memory pipeline and that fact will reflect on some legacy code. Synthetic benchmarks with small data sets are mostly prone to this. FP intensive applications should see some free boost from the reorganized ALUs and the two extra issue ports and, of course, the iterative improvements in the prediction logic and reordering buffers size.

 

7,8% faster in WPrime 32M, 7,1% slower in SuperPi 1M.

Though it's still faster clock-for-clock across the board, and it will probably consume less power, while the later models will eventually attain higher clocks (turbo or otherwise).

Given AMD's roadmap perspectives of increasing IPC by 15% every ~1 year cycle, Intel is still allowed to sit back and relax (in the laptop/desktop CPU market, at least).

 

If intel stands still, how many years do AMD need to catch up by this rate of increase in IPC 15%?

 

Depends on how you count, but Intel is not standing still, there's a net improvement here on average, which does not even include clock speeds.

And AMD won't be able to sustain the kind of performance improvement we saw with Piledriver (and soon Steamroller) for long: sooner or later, the law of diminishing returns slaps you in the face.

They're just starting from such a low point that it's not a major issue yet.

 

I don't think we can draw an overall picture with only a few Super Pi or wprime scores. And we even don't know if this Haswell model they tested was a fully functional one with SMT. I would be more interested in Cinebench 11.5 or some newer applications.

 

https://twitter.com/FPiednoel/status/296459612377468928


Waiting for some more reliable tests.

 

This. It's not difficult to improve on garbage. And if you're basing everything purely off stock performance, you're missing the point.

 

The only disappointing thing about these benchmarks is the benchmark set they used.

* They used 18 year old benchmark that uses ancient x87 FPU instrucitons and calculates things nobody EVER needs in real world workloads.
* Another program they use is 10 years old thing that calculates absolutely the same thing nobody EVER needs in real world workloads.
* Third thing they benchmark is a prime calculation. Again a micro-benchmark targeting very specific thing, which is very rarely used in real world workloads (when creating some crypto keys, but not needed when just using those)

So, one really cannot conclude much about so bad benchmark set.

 

Well IMO there is a limit as to how much ILP you can extract from a single core and we are approaching it. Haswell increases the size of the instruction window to 192, where do we go from there 256 OK. But you end up ratcheting up complexity of your core to unsustainable levels in the end we have to increase ILP, clock speed, TLP - mostly a function of the program, programmer, compiler and tool chain but can be helped with better inter core communication and caches -, and of course improving the memory subsystem - prefetching, helper threads, DDRx.

There are other things which I hope will make it like out of order retirement but again these are 10% IPC increases. My point is AMD will catch up but it could take them anywhere up to a decade and computer performance increases will mostly be coming from a better and more integrated GPU with the rest of the system an area in which AMD has a significant lead. On the other hand Intel has a lead in power consumption and process.

 

An informal benchmarks AMD APU Richland

http://www.obr-hardware.com/2013/02/some-news-from-amd-apu-world.html