Intel Broadwell-E Review | Hacker News Reader

Intel Broadwell-E Review

anandtech.com

65 pointsethana32 comments

I want to upgrade my desktop/gaming box from a 5yo but maxed out Gulftown i7-970, but this really isn't terribly compelling. I guess I wait another generation (throw a new GPU in there, but nothing else)?

Dual CPU Xeon seems better (mmm, ECC), although faster RAM is nice, and I'm not sure what a "good" workstation board is. Is there a market for ~silent/water cooled 2P Xeon workstations still? No need to overclock, but something quiet is nice -- big diameter fans at minimum.

AMD's mainstream processors support ECC and have for several years. I wouldn't be surprised in the new mainstream Zen CPUs also have ECC support.

Why would you want ECC for a gaming box?

Ah, the old Law of ECC. Say it once and everyone assumes you not only do you not need it, but you don't know how much you don't need it!

I have one machine with ECC, because it runs a backup server for all of my devices (Bacula, for anyone curious). I wouldn't mind it on all my desktop, though, and will probably have it in the next build. I remember asking around about the miniITX motherboard for the backup machine, just to double-check that worked with ECC before buying, and I couldn't get a simple yes/no answer out of anyone. Just a lot of words about how much I definitely didn't need ECC.

Why wouldn't you want ECC for everything?

Well ECC is not the be-all end-all. There are several sources of bit errors. SD-Cards, USB Flash drives, TCP/IP Packets, HDDs, SSDs etc etc. Thats just the 'bits' you have control over. If you're copying data from somewhere else, you have no control over what the other person does to it, etc etc. I wont't produce a thesis here, but you get the idea :)

Right, ECC is improving one piece. ZFS adds file checksums. SATA uses ECC for transfers. Considering the cost I think ECC is worth it.

The xeon E3-1230 is cheaper than the i7-6700K and while it has a slightly slower clock I don't think it's noticeable.

Hmm, do you happen to know what the actual improvement is, in terms of a reduction in the probability of bit errors? What I'm thinking is.. assuming there are multiple weak links in the chain, strengthening one link, might not really make all that big of a difference. But I agree with the general point you're making - something is better than nothing.

There is about a 1-2% performance overhead for ECC. Not that that many tasks are bound to memory I/O.

It also tends to only be available in "older" types of memory, with higher latency and lower bandwidth, on top of the ECC overhead itself.

For gaming, not necessary. OTOH, if I'm going to put a 1080 GTX, 2 Xeons or $1700 single-proc, 32-64-128GB RAM, lots of SSD, etc. into a box, I sort of want to make it dual-use.

Option B is just getting two boxes and multiple inputs on monitor/kvm/etc.

Annoying that the i7-6800K is gimped by having only 28 PCI-E lanes. I guess Intel needed a reason for people to buy the i7-6850K.

Interesting, it's not that much better than the available 2015 line up, for example my iMac has a 6700K in it which sits in the top 3 or 4, mind you this is typical of a tick, in a tick-tock manufacturing process.

Why did they come out with another Broadwell based processor when Skylake / Ice Lake seems to be logical next tstep?

Intel -E platforms share more with Intel's Xenon for servers than they do their consumer chips. The Broadwell-E based on the Broadwell platform so always lags behind the consumer release of the platform.

This chip isn't designed for consumer workloads, so you wouldn't see much improvement if it was in your iMac, but is a nice improvement over the previous Haswell -E platform.

While it may seem on the surface that compared to the non-E version they just added a couple more cores. But the changes go deeper, for example the Broadwell-E i7-6850K has almost doubled the amount of cache, doubled the max memory to 128 GB. But more significantly, for some workloads they doubled the number of memory controllers from 2 to 4 and increased the number of on chip PCIe lanes from 16 to 40. This increase in memory bandwidth and PCIe lanes allows for the creation of monster multiGPU data processing computers.

Here is a comparision between your iMac's CPU and the new Broadwell-E i7-6850K. As you can see it is really an Apple and Oranges comparison: http://ark.intel.com/compare/94188,88195

The -E line of processors is always one process generation behind the mainstream chips, these are processors with more cores/PCI lanes/cache on an "enthusiast" platform. Essentially unlocked Xeons.

It may be only a 'tick' in performance, but it sure is a 'tock' in price! $1723 for a i7-6950X v. $999 for a i7-5960X?

That's even more than the Broadwell-EP Xeon equivalent:

> The recently released Xeon Broadwell-EP processor list includes the Xeon E5-2640 v4: a 10-core 2.4 GHz/3.4 GHz part that runs at 90W, and is priced at $939, which compares favorably to the i7-6950X and its 10-cores at a 3.0 GHz/3.5 GHz clockspeeds.

Interesting. So why would one buy the i7-6950X instead of the Xeon E5-2640 v4?

Overclocking. And wank.

Intel chip performance improvements seem to have slowed down. They add extras like acceleration for 3d prints, USB improvements, etc. But raw horse power seems at a standstill. I wonder if this is because they don't have a fierce competitor anymore or whether they are keeping "high" multicore chips for their Xeon line. A decade ago, people kept harping we'll have hundreds of cpu cores in our desktop systems any day now. That didn't happen. We do have thousands of GPU cores .. so .. go NVidia and AMD! Intel needs to get their act together.

It's partially because of the lack of competition and partially because the process technology itself has slowed down. We're starting to run up against the limits of silicon, going from 22nm to 14nm isn't as big of a performance increase as going from, say, 90nm to 45nm was a decade ago.

I wonder how power efficient new gen. cpus are compared to their ancestors.

It's been that way since Ivy Bridge believe it or not. Sandy Bridge was the last one to gain a 30% performance boost over the previous generation. All the rest got around 5%.

it's literally quantum physics vs the world at this point so saying 'get their act together' is a bit unfair.

> We do have thousands of GPU cores

Those are more like ALUs, and very simple ones at that, than cores.

I think at this point getting even further ahead of AMD would only get them into trouble.

Actually it's considerably faster, it's twice as fast as the 6700K in many applications that actually use multi-threading properly, this is a weird review, no blender, no encoding, nothing that uses the new instruction sets when comparing to the broadwell-E CPU's, really a poor review which is surprising considering the source.

>nothing that uses the new instruction sets

There are no new instructions in Boardwell-E. It doesn't support AVX512, and AVX2/1 are both part of stock Boardwell and Haskell.

The new MPX extensions are only in Skylake as well. Furthermore MPX isn't designed to increase compute, but help make bounds checking faster.

IIRC Broadwell has several improvements over haswell both in IPC and instruction sets, broadwell has support for ADX which can improve performance in various application such as encoding and compression.

The benchmarks for the 6950X show 2 times or higher the performance in winrar and various encoding benchmarks. almost 2 times the performance boost in ray tracing, and about the same one in blender.

This review is some what flawed they've selected all the wrong benchmarks, very few real world workstation / professional applications and their gaming benchmarks were completely wrong.

A gamer doesn't buy a 6+ core CPU to get better FPS they buy it to CPU encode on OBS or other streaming surface while gaming so a benchmark for any E series CPU should include that if you want to focus on gaming. If not there is no reason to post any gaming benchmarks Mainstream Core i7/i5 would always be better because even the best multi-threaded games out there don't scale with anything more than 3-4 cores and higher single core performance is king as far as gaming goes even in DX/Vulcan benchmarks.

>IIRC Broadwell has several improvements over haswell both in IPC

Broadwell was a dye shrink of Haskell but architecturally identical. It was just moving from 14nm to 10nm. There was no IPC improvements. When you compare the 2 chips performance the difference was literally clock speed [1].

Yes there were encoder changes but the iGPU isn't part of Broadwell. It's iGPU, that's like saying the DRAM controller, or PCIe master router is part of Boardwell.

[1] http://www.anandtech.com/show/9320/intel-broadwell-review-i7...

Yes Broadwell was a dieshrink of Haswell with 10-15% improvement in ICP but it also had a few instructional changes like ADX.

https://en.wikipedia.org/wiki/Intel_ADX

Intel has retired the tick-tock model: http://arstechnica.com/information-technology/2016/03/intel-...

They're shifting to three phase: process(tick)-architecture(tock)-optimization.

AKA "tick-tock-tweak".

Reply on news.ycombinator.com