David Kanter on the Haswell SOC

[Phynaz]

Informative higher level article on what Haswell brings to the SOC world.

http://arstechnica.com/gadgets/2013/05/a-look-at-haswell/

 

[Charles Kozierok]

For those who don't know him, David Kanter has been the head dude at Real World Technologies for a while, and seriously knows his stuff. Interesting that he's at Ars now.

 

[Carfax83]

The most significant ISA extension is TSX, which has been extensively discussed in a previous article on Haswell's transactional memory. In short, TSX separates performance from correctness for multi-threaded programs. Programmers can write simple code that is easier to debug, while the hardware extracts concurrency and performance.

I don't get it. If TSX is the most significant ISA extension, why isn't it enabled in the K versions?

 

[CHADBOGA]

For those who don't know him, David Kanter has been the head dude at Real World Technologies for a while, and seriously knows his stuff. Interesting that he's at Ars now.

I think he is still at RWT, but freelances some of his stuff to Ars.

 

[piasabird]

Just run some benchmarks. A demonstration of how it works better might be nice. If the processor can handle multitasking better maybe testing it with a sorting multiple thread sorting algorithm might be a way to test it. This is just a suggestion of one simple test. Another test might be opening several programs that can run at the same time, or compiling animations.

If you simply stated what functions run better we might see how it works. Most people will not understand exactly how a processor runs or what good one new instruction will do. Modern processors are very complex.

I would kind of like to see a benchmark test on that new PICO motherbord made in India with the qualcom quad core processor(?Snapdragon). That might be interesting.

http://www.tomshardware.com/news/Embedded-Snapdragon-600-Krait-Inforce,22307.html

Speed is not everything when an i3 is fast enough for most people. People buy tablets that seem to run like a snail and like them.

 

[2timer]

Article's title really fails. SoC is barely mentioned, it's just a marketing piece on the new microarchitecture.

 

[Ajay]

I don't get it. If TSX is the most significant ISA extension, why isn't it enabled in the K versions?

AVX2 is the most important ISA extension for consumer CPUs. The 'K' series are aimed at extreme overclockers, which isn't TSX's target market, IMO.

 

[Ajay]

For those who don't know him, David Kanter has been the head dude at Real World Technologies for a while, and seriously knows his stuff. Interesting that he's at Ars now.

Man, I remember him from comp.arch discussions over a decade ago.

 

[Carfax83]

AVX2 is the most important ISA extension for consumer CPUs. The 'K' series are aimed at extreme overclockers, which isn't TSX's target market, IMO.

So consumer programs won't benefit from the TSX instructions?

 

[ShintaiDK]

I don't get it. If TSX is the most significant ISA extension, why isn't it enabled in the K versions?

Who says, besides Toms , that the K model wont have TSX?

 

[Tuna-Fish]

I don't get it. If TSX is the most significant ISA extension, why isn't it enabled in the K versions?

Because Intel is being really, really stupid. I really do hate them for that choice. Not having it on the best desktop CPUs means that it will be very long until I can depend on having it when writing programs.

So consumer programs won't benefit from the TSX instructions?

All heavily multithreaded ones potentially could. The decisiong to segment based on it is probably done to extract more profit from the server CPUs.

 

[Carfax83]

Who says, besides Toms , that the K model wont have TSX?

It seems to be a common notion among the tech media at large, but that's not to say that Toms hardware wasn't the only source of this information.

Personally, I find it rather suspicious so I'm hoping it's wrong.

All heavily multithreaded ones potentially could. The decisiong to segment based on it is probably done to extract more profit from the server CPUs.

But if that's the case, why does the 4770 non K version still have it enabled? Assuming Tom's is right of course.

 

[ShintaiDK]

It seems Toms got TXT and TSX mixed up

 

[NTMBK]

Article's title really fails. SoC is barely mentioned, it's just a marketing piece on the new microarchitecture.

To be fair, Kanter may not be responsible for the title.

 

[ViRGE]

Article's title really fails. SoC is barely mentioned, it's just a marketing piece on the new microarchitecture.

Not to drag this off topic, but go read the volumes and volumes of work over at RealWorldTech before you call anything by Kanter a marketing piece. Kanter is second to only Anand himself in his knowledge of CPUs, and the sheer awesomeness of the resulting articles. I don't think you appreciate just how much information is being given to you.

 

[grimpr]

After Kanter said, and his word counts, that TSX is the most important new extension i guess its game over for buying and building with unlocked Intel chips at least for me, i'll buy the extension instead of the unlocked multiplier.

 

[Phynaz]

For those who don't know him, David Kanter has been the head dude at Real World Technologies for a while, and seriously knows his stuff. Interesting that he's at Ars now.

From David:

The Ars folks and I are trying to work out something where I can write here from time to time. So if you like it, please ping Ken or Eric : )

 

[Idontcare]

I don't get it. If TSX is the most significant ISA extension, why isn't it enabled in the K versions?

Just speculating here but one possible reason could be that TSX is more difficult (finicky) when trying to find Vcc_min for stable and reliable operation at any given clockspeed and Intel is worried that OC'ers will give TSX a black-eye by OC'ing their rigs but not giving it enough Vcore resulting in instabilities all over the map in apps that are TSX-conscience.

So if they disable the feature entirely then at least they retain some control over the mindshare perception that slowly but surely develops for it. Consider how hyperthreading's reputation has been for 10yrs now based on the poor performance and reception of its initial implementation in Prescott.

 

[cantholdanymore]

Just speculating here but one possible reason could be that TSX is more difficult (finicky) when trying to find Vcc_min for stable and reliable operation at any given clockspeed and Intel is worried that OC'ers will give TSX a black-eye by OC'ing their rigs but not giving it enough Vcore resulting in instabilities all over the map in apps that are TSX-conscience.

So if they disable the feature entirely then at least they retain some control over the mindshare perception that slowly but surely develops for it. Consider how hyperthreading's reputation has been for 10yrs now based on the poor performance and reception of its initial implementation in Prescott.

Besides "K" chips tend to be "discontinued" by their owners a lot sooner than non enthusiast chips, so a lot of them won't be on a "main pc" when TSX becomes common is software

 

[lagokc]

I don't get it. If TSX is the most significant ISA extension, why isn't it enabled in the K versions?

A few workstation features are intentionally kept out of the K versions because Intel doesn't want sneaky people from buying a K CPU, clocking it to the sky, and holding onto it for years instead of buying a more expensive Xeon system every couple years. Previously virtualization was the primary feature left out.

 

[Charles Kozierok]

Kanter is second to only Anand himself in his knowledge of CPUs, and the sheer awesomeness of the resulting articles. I don't think you appreciate just how much information is being given to you.

With respect to Anand, I believe Kanter knows quite a bit more about CPUs. I'm pretty sure he was a CPU architect. I often have trouble even understanding what he's talking about as it is very technical.

Who says, besides Toms , that the K model wont have TSX?

I realize that Tom's-bashing is a pastime around here, but for the record, I asked Chris Angelini about this and he maintains it is not a mistake and that he has reliable sources. We'll see.

From what I understand, TSX is mostly important in business applications. It's not essential, it's a performance optimization, and probably matters a heck of a lot more to a database server than a guy playing 3D shooters. I'm not sure it really matters much if they omit it from the "K" chips.

 

[guskline]

Despite years of fooling around with computers, building and breaking them as a hobby and following cpu development, I've had a chance to read many articles on the substance of a cpu. A lot of what David Kanter wrote is way over my head. However, what he said overall about the Haswell makes me believe that this architecture will have a long term positive affect on Intel and it's ability to continue not only in the server and desktop arena but make a BIG dent in the tablet arena, if it hasn't already. Impressive article.

 

[kukreknecmi]

This is not a start-up article, it is the washed up / shortened version of what he wrote on RW. Even ars may had modified it.

I kept saying where did i read this sentences, must have red somewhere, seems familiar.


http://www.realworldtech.com/haswell-cpu/

 

[Fx1]

All i see is the numbers going up and performance remaining static and the power use going down.

Not a fan but what choice is there

 

[lagokc]

All i see is the numbers going up and performance remaining static and the power use going down.

Not a fan but what choice is there

Judging by the features added I kind of get the impression that even if Haswell is only a few % faster in today's tasks, it will show more significant improvement once compilers are more optimized for it.

Don't we already see this sort of thing? Sandy Bridge was only a bit faster than Nehalem on release day but compare them with applications today and the gap widens a bit?