Haswell Preview Benchmark

[ldesnogu]

Originally Posted by Prime95

I created a torture test that uses really small FFTs that fit in the L1 data cache. Alas, it runs cooler than the FFTs that fit in the L2 data cache.

I guess Haswell has extremely aggressive clock gating and is able to turn L2 clock off. Perhaps prefetching from L2 or even L3 (as Ernst hinted) while processing data from L1 will increase power consumption?

[db597]

Wow, still memory bottlenecked even with DDR3-2400 memory. How much do you think we need to escape the bottleneck? 2800?

I had been looking to upgrade my 3Y old 1600 modules to 2133... but I suppose this is a signal that we might as well go all the way.

[only_human]

Originally Posted by db597

Wow, still memory bottlenecked even with DDR3-2400 memory. How much do you think we need to escape the bottleneck? 2800?

I had been looking to upgrade my 3Y old 1600 modules to 2133... but I suppose this is a signal that we might as well go all the way.

Next year's Haswell-E will be DDR4 capable. Haswell-E - Intel's First 8 Core Desktop Processor Exposed. Dunno how fast the memory will be since DDR4 is getting such a slow start. One thing I like is that DDR4 has Data Bus Inversion just like GDDR4 memory does:

Data Bus Inversion adds an additional active-low DBI# pin to the address/command bus and each byte of data. If there are at least four 0 bits in the data byte, the byte is inverted and the DBI# signal transmitted low. In this way, the number of 0 bits across all 9 pins is limited to 4. This reduces power consumption and ground bounce.

[TheMawn]

Originally Posted by db597

Wow, still memory bottlenecked even with DDR3-2400 memory. How much do you think we need to escape the bottleneck? 2800?

I had been looking to upgrade my 3Y old 1600 modules to 2133... but I suppose this is a signal that we might as well go all the way.

We ARE talking about a VERY fast chip which is overclocked. To give you another idea, I get a 20% drop in throughput when I add my fourth worker. This is an i5-3570k at 4.6GHz and DDR3-2400.

One might suspect a 20% increase in memory speed would suffice, which is about 2880MHz.

If you are running three year old RAM, your CPU must be fairly aged too and 2133MHz is probably good. You can see for yourself by picking a random exponent and testing it on one, two, three, then four workers at a time and seeing if your time-per-iteration gets lower and lower as you increase the number of workers trying to access memory.

If it DOES decrease, your RAM is the bottleneck. You can decide how badly you think your RAM is bottlenecking before buying the fairly expensive 2400MHz RAM (my exact kit has gone up $50 since I bought them), ALSO considering your system is likely unable to handle 2400MHz. AMD is particularly bad.

[TheMawn]

Originally Posted by only_human

Next year's Haswell-E will be DDR4 capable. Haswell-E - Intel's First 8 Core Desktop Processor Exposed. Dunno how fast the memory will be since DDR4 is getting such a slow start. One thing I like is that DDR4 has Data Bus Inversion just like GDDR4 memory does:

What's kind of sad about that is we're doubling memory bandwidth and doubling the worker count. All I've seen so far leads me to believe that the memory is somewhat bottlenecking but not by a hell of a lot. On the other hand, if Haswell-E uses quad channel as well, then we're set.

[kracker]

Originally Posted by TheMawn

What's kind of sad about that is we're doubling memory bandwidth and doubling the worker count. All I've seen so far leads me to believe that the memory is somewhat bottlenecking but not by a hell of a lot. On the other hand, if Haswell-E uses quad channel as well, then we're set.

Who knows when it will be released. IB-E isn't even out yet.

[firejuggler]

hmm read post 196, ivy bridge-E bench

Consider this a parting shot: Core i7-4960X is faster than Core i7-3970X and simultaneously about 30% more efficient. In the world of Xeon E5-2x00 v2 processors, that’s going to be killer.

[db597]

Originally Posted by TheMawn

If it DOES decrease, your RAM is the bottleneck. You can decide how badly you think your RAM is bottlenecking before buying the fairly expensive 2400MHz RAM (my exact kit has gone up $50 since I bought them), ALSO considering your system is likely unable to handle 2400MHz. AMD is particularly bad.

I'm running a i7-4770K at 4.2GHz as well. Didn't do a complete rebuild, so still using the old memory from my previous build in this system. It's tempting to replace the RAM, but there's quite a big price difference between 2133 and 2400. Also, we're looking at going from CL9@1.5V to CL10/CL11@1.65V between these 2 grades. Wasn't sure if it's worth it.

[TheMawn]

Correct me if I am wrong but my belief was that a higher frequency with looser timings has more bandwidth than lower frequency with tighter timings, in general.

Anyway yes the price gap is severe enough that 2133 is probably safer. With an overclocked fourth gen i7 you're going to see a huge speedup across the board with some faster RAM. You might be able to overclock above the stock specs too...

[henryzz]

Originally Posted by TheMawn

Correct me if I am wrong but my belief was that a higher frequency with looser timings has more bandwidth than lower frequency with tighter timings, in general.

Anyway yes the price gap is severe enough that 2133 is probably safer. With an overclocked fourth gen i7 you're going to see a huge speedup across the board with some faster RAM. You might be able to overclock above the stock specs too...

The frequency is what effects bandwidth. Timings probably won't matter unless they are bad enough to stop you maxing the bandwidth.

[db597]

Assuming an affordable price difference (hmm.. I'll just skip 1 lunch), would you guys go for 2133 CL9 or 2400 CL11? Trying to decide between them, both GSkill RipRaws X. So far, the P95 discussion seems focused on bandwidth and I don't hear about timings being mentioned. Does the more relaxed timing only a small secondary concern?

I suppose the best performance would be a 2400 CL10... but that cost is in another league (another 30%+ more).