Haswell vs Kaveri

[Rangers]

Somebody on H is speculating that with 4GB RAM that Kaveri may be in single channel mode. In other words, leaked benchmarks are meant to be salty.

If AMD really did actually decrease FP performance though, that's horrible, since it's the weak point of the architecture to begin with.

 

[pjbliverpool]

AMD would argue that the FP performance of the chip is vastly higher than anything before it thanks to HSA. It's the app that isn't properly using that performance. They certainly have made statements previously about FP performance stagnating or even reducing in CPU's but this being no big deal thanks to HSA so it wouldn't come as a great surprise to see FP performance lower in Steamroller. Big disappointment though considering Intels strength in this area.

 

[fellix]

AMD would argue that the FP performance of the chip is vastly higher than anything before it thanks to HSA. It's the app that isn't properly using that performance. They certainly have made statements previously about FP performance stagnating or even reducing in CPU's but this being no big deal thanks to HSA so it wouldn't come as a great surprise to see FP performance lower in Steamroller. Big disappointment though considering Intels strength in this area.

I think Intel is still hedging on that matter. While they are too pushing their IGP offerings, the traditional FP/SIMD pipeline is still apparently first-class development target, not only for the server/HPC product class.

 

[pjbliverpool]

Indeed. 1 TFLOP under turbo seems pretty realistic for Skylake CPU SIMD in 2015. That's more than the whole Kaveri APU, GPU and all.

 

[Paran]

Somebody on H is speculating that with 4GB RAM that Kaveri may be in single channel mode. In other words, leaked benchmarks are meant to be salty.

It doesn't matter for Cinebench R15 CPU score if it run in single or dualchannel mode.

Indeed. 1 TFLOP under turbo seems pretty realistic for Skylake CPU SIMD in 2015. That's more than the whole Kaveri APU, GPU and all.

As far as we know consumer Skylake won't get AVX-512.

 

[pjbliverpool]

As far as we know consumer Skylake won't get AVX-512.

Is that conclusion based on a lack of evidence supporting AVX-512 on Skylake or some specific source saying that it will not be featured?

 

[Chabi]

 

[Paran]

Is that conclusion based on a lack of evidence supporting AVX-512 on Skylake or some specific source saying that it will not be featured?

When Intel announced AVX-512 they claimed it will come for Knights Landing and Xeon. Nothing about consumer Skylake.

http://pics.computerbase.de/5/2/9/5/3/3.jpg

Intel AVX-512 will be first implemented in the future Intel® Xeon Phi™ processor and coprocessor known by the code name Knights Landing, and will also be supported by some future Xeon processors scheduled to be introduced after Knights Landing.

http://software.intel.com/en-us/blogs/2013/avx-512-instructions
http://software.intel.com/en-us/forums/topic/404037#comment-1748254

 

[pjbliverpool]

Is it likely that Intel would remove this feature from a consumer version of the Skylake CPU while leaving it intact in the server version? I'd have thought it would be too fundamental to the Skylake design to seperate out like that but then I'm hardly an expert on CPU design!

There is this (now pulled) slide linked from this page:

http://www.bit-tech.net/news/hardware/2013/07/05/intel-skylake/1

Which at least suggests Skylake Xeons will feature AVX 3.2 (512).

 

[Paran]

The leaked slide was a HPC Roadmap, refers to EP/EX and Xeon Phi. They should get AVX512 of course. Whether or not consumer Skylake has AVX512 integrated but disabled depends on the Xeon E3 entry CPUs. If they don't get AVX512 then I would assume there is no AVX512 integrated.

 

[Rangers]

http://linustechtips.com/main/topic/97361-first-kaveri-apu-benchmarks-surface/

Alleged Kavari benches.

 

[liolio]

Nice improvements overall, though I'm more and more wary about AMD investing so much silicon on the GPU, actually I'm close to think that it is a waste of money.

 

[Alexko]

They need to differentiate somehow…

 

[pjbliverpool]

If that's really representative of GPU performance then Broadwell is going to wipe the floor with it. At least at <=65w

 

[Alexko]

It is certainly not representative of Kaveri's graphics performance, which is hardly surprising since AMD still hasn't released any drivers for Kaveri.

 

[Wynix]

http://linustechtips.com/main/topic/97361-first-kaveri-apu-benchmarks-surface/

Alleged Kavari benches.

This seems to be for the 65W version, would that be more likely to throttle?

 

[Alexko]

For Trinity/Richland, the top 65W SKUs were very close to the top 100W ones. I'd expect them to be as close or even closer with Kaveri.

 

[UniversalTruth]

Nice improvements overall

Relatively good improvements on the CPU side with between 10 and 40 % but the new GPU scores are embarrassing at best.

It is certainly not representative of Kaveri's graphics performance, which is hardly surprising since AMD still hasn't released any drivers for Kaveri.

How did they perform the benches without a Kaveri driver? There should be something put, shouldn't it?

Please, if you mean something, clarify, so people understand you

 

[Deleted member 13524]

Relatively good improvements on the CPU side with between 10 and 40 % but the new GPU scores are embarrassing at best.

It could be the rumoured testing platform with single-channel memory. It would explain the good CPU scores and abysmal GPU scores.

 

[Alexko]

Relatively good improvements on the CPU side with between 10 and 40 % but the new GPU scores are embarrassing at best.



How did they perform the benches without a Kaveri driver? There should be something put, shouldn't it?

Please, if you mean something, clarify, so people understand you

They probably just used the latest Catalyst, which do not support Kaveri. Obviously it still works, but the poor performance should not come as a surprise.