Intel's Future Chips: News, Rumours & Reviews

Page 31 - Seeking answers? Join the Tom's Hardware community: where nearly two million members share solutions and discuss the latest tech.
Toggle sidebar Toggle sidebar
Status
Not open for further replies.


Half true, I'd say. I do concede that in the broad sense, you can't optimize a certain point, so using AVX, SSE or whatever won't do magic since you'd have to re-work the code itself around a certain "feature" to be used. BUT, for regular code that *already has* code paths that allow for this kind of "easy" speed up, it is worth it. This comes from the first X87 vs MMX talks.

Cheers!
 


Yes, but typically, performance for applications that aren't tuned SPECIFICALLY to use special instructions pretty much peaks after SSE2. That's why, for most apps, AVX support means NOTHING.
 


Let me see if I got it right... So you're saying that, if a piece of software already has been optimized for SSE2 / AVX or even SSE1, you won't see big gains moving to AVX2 or AVX512?

Can we conclude that it will only matter for time critical applications and since Cazalan has said it might not come to the consumer space in a good while, AVX512 is a moot point to us?

Cheers!
 


Pretty much. There's only so many opcodes the compiler can automatically drop in during the compilation process, and after SSE2, those opcodes start getting REALLY specialized. So unless you are specifically tuning an application against those opcodes, they simply aren't going to get used.

The last round of compiler benchies I have is from 2012, which among other things, compared X86, SSE, SSE2, SSE3, and AVX. And averaging the results, the SSE2 build was the fastest overall.

Granted, compilers are likely a bit better now, but it's REALLY hard to use some of those instructions. They really don't matter, which is why I chuckle every time I hear people saying how AVX512 is going to double bandwidth or something.
 
Colorful Unveils First Skylake Z170 Chipset Based Motherboards - Features Socket LGA 1151, Officially Launching in September
http://wccftech.com/colorful-unveils-skylake-z170-chipset-based-motherboards-features-socket-lga-1151-officially-launching-september/

 


AVX512 makes little sense for the comsumer side, which is probably why it's disabled or removed to save die area. At least the last 8 or so leaked Skylake slides (Desktop/Mobile) don't mention it.

For HPC Intel wants the AVX512 for Xeon and Phi. What ends up happening is the first generation of code is worked out on Xeon. If the code/project is deemed useful it gets optimized for Phi. If it ends up getting used a lot they will optimize further. Which makes sense. Scientists need preliminary results quickly to see if their hypothesis is on the right track. Why spend time optimizing if you're barking up the wrong tree?
 
Those leaks paint a really sad picture, mainstream performance CPU´s still coming on 4 cores with the HT unit disabled is plain unacceptable, this means we will still have i3 dual core chips costing 170$ in 2017 being easily outperformed by sandy bridge 4 core chips that costed 180$ back in 2011 and considering that DDR3 memory was a lot cheaper back then, price/performance ratio hasn't changed a bit in 6 years, newer instructions that nearly no software uses and small IPC improvements mean very little in multithreaded apps and newer chipset even cut overclocking and PCI lanes in the low end, and DDR4 doesn't even improve the bandwidth per clock.
Also due to different manufacturing processes we can pretty much forget about a possible Haswell price drop.
Best price/performance CPU is getting an used i5-2500K
 
Why i'm so negative towards Intel's competition i'm never satisfied AMD needs to up their game^^^^^^^ I want to see this
I7 Extreme 8 core+HT only
I7 6 core+ HT
I5 4 core+ HT
I3 4 core
Pentium dual core +HT
 
Intel Haswell/Broadwell Power Use On Linux Still Moving Lower
http://www.phoronix.com/scan.php?page=news_item&px=Intel-HSW-BRW-Power-Drop
Intel Is Making Some Progress With Compute Shaders
http://www.phoronix.com/scan.php?page=news_item&px=Intel-Mesa-CS-Progress
Intel Xeon E5-2620 v3 Haswell Performance
http://www.phoronix.com/scan.php?page=news_item&px=Xeon-E5-2620-SNB-HSW-Linux
Intel Bay Trail & Cherry Trail CPUs To Run Faster With Linux 4.1
http://www.phoronix.com/scan.php?page=news_item&px=Intel-Bay-Cherry-Faster-4.1

 
We got a full suite of benchies now:

http://www.pcfrm.com/intel-i7-6700k-vs-i7-4790k/

3D mark: 13.1%
Cinebench: 13.2%
PC Mark: 4%

GTA V: 2.7%
BF4: 5.7%
Crysis 3: 0%

So architecturally, somewhere between 10-15% IPC gains when not bottlenecked by some other component, or about what you'd expect.
 

the gains seem only in multithreaded workloads. i wish the site had published something on the test platform, so i'd find the benches more encouraging.
 
PC Mark is bound by the system, so you'd expect lower gains, and games are bottlenecked by the GPU, so low to no gains are expected. 3D Mark and Cinebench are about as CPU agnostic as you can get, and those show basically the same performance gain. So that 13% is probably the "best case" IPC increase in pure CPU workloads.
 
They were not even trying to hide it.

"We have prepared the following about the potential performance graphics processors Skylak curve. Intel i7 6700 benchmark test that takes place in this comparison were prepared on the basis of performances given by Intel's processors in the previous year."

 


I didn't either until someone (Abwx) in another forum pointed it out.

*Look pretty graph must be legit* :no:
 
Apparently not everyone planning to go with lots of little cores. SGI expanding to 16 and even 32 socket servers.

32 x E7-8890 v2 (15-core) = 480C/960T

"The SGI UV 300H architecture is ultimately designed to scale to 32-sockets and up to 24TB as a single-node system."

http://www.hpcwire.com/off-the-wire/sgi-uv-for-sap-hana-expands-to-12-and-16-sockets/

 
ASRock Develops Mini-ITX LGA2011v3 Motherboard with Quad-Channel Memory
http://www.techpowerup.com/212204/asrock-develops-mini-itx-lga2011v3-motherboard-with-quad-channel-memory.html
4x so-dimm slots. drool.

Intel will turn your modem into a server
http://www.fudzilla.com/news/processors/37670-intel-will-turn-your-modem-into-a-server

Falcon Northwest releases micro-tower Tiki PC with 18 cores inside
http://www.pcworld.com/article/2916549/falcon-northwest-releases-micro-tower-tiki-pc-with-18-cores-inside.html
moar cores, but better.
 
If the biggest gain with skylake would be in multithreaded workloads, wouldn't this suggest improved SMT (in the sense of a wider core) and better inter-core communications?
 
here is a monotonously linear prediction of future intel cpu model numbers:

skylake:
core i3 61xx,62xx, 63xx up to 3.8 GHz dual cores with HT or 2x dual cores with shared 512KB L2 cache or whatever (like the atoms).
core i5 64xx, 65xx, 66xx from 2.7 GHz base to 4 GHz max. turbo
core i7 67xx from 3.5 GHz base to 4.3-4.4 GHz max turbo

broadwell-e:
core i7 6820k-6970x up to 3.9 GHz max turbo

cannonlake:
core i3 71xx blah blah
core i5 74xx blah blah
core i7 77xx blah blah

skye:
core i7 7820k-7970x GHz Edition

*Yawn* i hope i'm wrong.
 
Status
Not open for further replies.

TRENDING THREADS

Latest posts

Moderators online

Share this page

COMPANY
RESOURCES
FOLLOW
Top Bottom