Intel TSX Presentation at SuperComputing 2013

[jvroig]

This is still a few months away, but I got tired of waiting so posting now.

Evidently, for real-world HPC apps (as Intel themselves describe), TSX can be good for as much as a 40% speed-up:

In this paper, we evaluate the first hardware implementation of Intel TSX using a set of high-performance computing (HPC) workloads, and demonstrate that applying Intel TSX to these workloads can provide significant performance improvements. On a set of real-world HPC workloads, applying Intel TSX provides an average speedup of 1.41x. When applied to a parallel user-level TCP/IP stack, Intel TSX provides 1.31x average bandwidth improvement on network intensive applications. We also demonstrate the ease with which we were able to apply Intel TSX to the various workloads.

Source

HPC is an integral part of my crypto research, so now I'm really torn between getting a 4770K vs a 4770 non-K.

Previously, I was all set to get a 4770K. I didn't mind missing VT-d, VT-x is all I need to run VM's for development purposes. TSX consideration gave me a little pause at first, and while the programmer in me was interested in trying it out, my initial interest wasn't enough to sacrifice overclocking capability. Since no specific claims of TSX performance benefits were cited back then, I decided to settle on a 4770K.

Unfortunately, before I could spring up for one, I read about this Intel presentation for SC13 in November (this was about a month ago), and so I'm back to the dilemma of choosing a K vs non-K. 40% average speed-up in HPC is nothing to scoff at, it would certainly aid in my research activities + satisfy my inner programmer + I'd never be able to overclock a 4770K 40%, so for HPC purposes I might actually end up losing performance by going for the K version. I wish Intel hadn't made these arbitrary fusing off decisions for the K versions. They already sell for a premium, after all.

Oh well. Just thought I'd share. Until the actual presentation at SC13, there's little in the way of actual data, but for those of you in a similar boat as I am, wondering if choosing non-K for TSX can be worth it, this is a data point, at least. We'll have more solid data after SC13 to see how really broadly applicable the 40% claim is.

 

[sm625]

The difference between the 4.0GHz achievable on a non-K version, and the 4.4GHz achievable on a K version is only 10%. No one is going to notice that. It's not going to make any difference in anything. But 40% could...

 

[Fjodor2001]

What was the reason Intel skipped TSX in 4770K? Is there any reasonable explanation for that decision?

 

[GlacierFreeze]

Probably so they can add it into the Haswell-E version laster down the road and make people want to ditch their K and spend even more money for a couple percent gain on the E + TSX.

I just ordered a non-k 4570. Didn't feel like bothering with OC for a few percent gain this go around, especially since these run so hot.

 

[Homeles]

The difference between the 4.0GHz achievable on a non-K version, and the 4.4GHz achievable on a K version is only 10%. No one is going to notice that. It's not going to make any difference in anything. But 40% could...

Well, at least in Anand Shimpi's eyes, 10% is the mark that he defines as being a noticeable performance increase. I think it's a subjective/biological thing -- some people are more perceptive than others. I'd argue that 10% is low enough for most people to notice, but only if they're looking out for it, and not on autopilot.

I think it also depends on the context of what's getting sped up. You might notice 10% more frames a second, or if your 5 minute workload that you run daily finishes in 90% of the time, but probably won't notice if your Windows calculator application opens up 10% faster.

40% is certainly noticeable, though. You'd have to be be seriously distracted or inebriated not to notice a 40% speedup.
_____________________

Anyways, I suppose a "good thing" for overclockers that stuck with the K-series is that TSX is a really new concept in mainstream computing. It's not like multithreading, which has been around in the consumer market for over a decade. Despite the ease of supporting TSX, it likely won't start seeing significant adoption until at least a year from now, and likely longer.

Some big hitter applications will undoubtedly support it though, and at that point it will be one of those "do I buy an FX or do I buy an i5" questions. It's going to depend on what your workloads look like, and what applications will benefit more from one over the other.

At least for now, I would argue that the vast majority of overclockers should go with the K-series. It's difficult to predict adoption rates, and we don't really know how TSX will affect popular software applications. We do know how overclocking speeds up applications though, so it's the safer, more universal bet.

Unless you really have to scratch the developer itch (and by this I mean you are an actual developer, not someone taking a computer science class), go with the K-series. TSX will be aroud for a long time to come -- overclocking won't.

 

[cytg111]

This is still a few months away, but I got tired of waiting so posting now.

Evidently, for real-world HPC apps (as Intel themselves describe), TSX can be good for as much as a 40% speed-up:

Source

HPC is an integral part of my crypto research, so now I'm really torn between getting a 4770K vs a 4770 non-K.

Previously, I was all set to get a 4770K. I didn't mind missing VT-d, VT-x is all I need to run VM's for development purposes. TSX consideration gave me a little pause at first, and while the programmer in me was interested in trying it out, my initial interest wasn't enough to sacrifice overclocking capability. Since no specific claims of TSX performance benefits were cited back then, I decided to settle on a 4770K.

Unfortunately, before I could spring up for one, I read about this Intel presentation for SC13 in November (this was about a month ago), and so I'm back to the dilemma of choosing a K vs non-K. 40% average speed-up in HPC is nothing to scoff at, it would certainly aid in my research activities + satisfy my inner programmer + I'd never be able to overclock a 4770K 40%, so for HPC purposes I might actually end up losing performance by going for the K version. I wish Intel hadn't made these arbitrary fusing off decisions for the K versions. They already sell for a premium, after all.

Oh well. Just thought I'd share. Until the actual presentation at SC13, there's little in the way of actual data, but for those of you in a similar boat as I am, wondering if choosing non-K for TSX can be worth it, this is a data point, at least. We'll have more solid data after SC13 to see how really broadly applicable the 40% claim is.

- Somewhat same boat and while deciding on much vaguer data, i opted for the non K version .. The OCability of the K version is a joke, and hey, TSX showed massive promise in theory and vt-d could turn out to be a nice surprise (hypervisor, run all your vm's side by side at native speeds).
I say go non K .. 40% is just preliminary .. that number can only go up .

 

[cytg111]

Well, at least in Anand Shimpi's eyes, 10% is the mark that he defines as being a noticeable performance increase. I think it's a subjective/biological thing -- some people are more perceptive than others. I'd argue that 10% is low enough for most people to notice, but only if they're looking out for it, and not on autopilot.

I think it also depends on the context of what's getting sped up. You might notice 10% more frames a second, or if your 5 minute workload that you run daily finishes in 90% of the time, but probably won't notice if your Windows calculator application opens up 10% faster.

I would like to negate that.
10% fps ? sub 30, maybe, but not really.
Daily runs? Very few would notice..
UI responsiveness ? Yes yes yes .. Take a note from old Steve, usability wins the day everytime.

 

[Homeles]

I would like to negate that.
10% fps ? sub 30, maybe, but not really.

Yeah really. I swear, some people are blind.

Daily runs? Very few would notice..

Okay.

UI responsiveness ? Yes yes yes .. Take a note from old Steve, usability wins the day everytime.

Well that one sure went over your head.

 

[cytg111]

Well that one sure went over your head.

- It happens im afraid. (was talking about windows calc opening 10% faster, if it makes a diff.)

 

[blackened23]

Yeah really. I swear, some people are blind.

Okay.

Well that one sure went over your head.

Why are you so combative about this stuff?

 

[Homeles]

Why are you so combative about this stuff?

Why not?

 

[blackened23]

It isn't necessarily, it's just PC hardware. Although I guess i'm guilty of it on occasion too so who am I to bring it up.

 

[Homeles]

When in a learning environment like this, the truth matters to me, regardless of the subject.

Regardless of which of us is correct, cytg111 isn't contributing anything constructive by nitpicking a minor, subjective detail of my argument.

 

[Loki726]

Here's a link to the paper: http://pcl.intel-research.net/publications/SC13-TSX.pdf

 

[Homeles]

Nice find.

 

[ShintaiDK]

Its really a nobrainer with your (abnormal) workload. Then get the non K. Unless you want something like HW-E.

Had you been a gamer and OCer, then the K model for sure.

 

[jvroig]

@Shintai

As with most other things in this world, if we could perfectly categorize everything and everyone into neat little boxes, then stuff would indeed be as simple as you say.

But it really isn't as flat and straightforward as one would expect, is it? For example, can't I be both someone who could certainly benefit from TSX because of work, while also being a gamer and overclocker as a hobby / off-work persona?

Surely, these "gamers and overclockers" you speak of also have a real job, so it's not totally impossible that a % of those (no matter how small) would have workloads that could benefit from TSX outside of their hobbies.

Hence why it's not quite as a no-brainer as you describe. We (you, me, everyone else) aren't nice little robots manufactured in a production line and stamped with one role or task. My current area of research (work, although I love it so it also ends up taking a bit of my hobby time) stands to benefit from TSX, but settling on non-K would mean the other area of my life (hobby, i.e., tinkering with the hardware, overclocking & tuning) would get a raw deal. I'm not sure how easily you can relate to this, but a purchasing choice based on the question "which do I give the raw deal to, the hobby area or the work area?" isn't the most enjoyable of choices. Thankfully, it rarely happens (I can't actually remember ever having to do this for any PC-related purchase previously).

I wish I can wait for Haswell-E, but seeing as to how Ivy-E is just arriving and how long it took to arrive, that's pretty much out of the question.

There's an obvious alternative that hasn't been mentioned yet - buying both the non-K and K versions, which is bound to make Intel and related component manufacturers happier as it would make me a customer twice for what should have been a single purchase. While I previously just built one machine to service both my hobbies and work-related stuff, I'm not totally averse to building more than one. I already have more computers than I can really use. Having one more extra won't really matter. So, yeah, I guess that worked rather well for Intel's product segmentation if I do end up on this course. :hmm:

 

[jvroig]

Here's a link to the paper: http://pcl.intel-research.net/publications/SC13-TSX.pdf

That's amazing! I didn't know they published these things well ahead of the presentations. Thanks, mate!

 

[seitur]

What was the reason Intel skipped TSX in 4770K? Is there any reasonable explanation for that decision?

Same as with Vt-d. To make high-clocked non-K versions sell better and possibly to get users used to choice: Overclock possibilities OR features OR huge premium for both (-Extreme line).

 

[JimmiG]

What was the reason Intel skipped TSX in 4770K? Is there any reasonable explanation for that decision?

Greed.

 

[ShintaiDK]

Greed.

Yep, greed from endusers who wants everything without willing to pay.

 

[zir_blazer]

For anyone that intends on purchasing a Core i7 4770, do yourself a favor and check out the Xeon E3 1245 V3. 100 MHz slower, 20 U$D cheaper, IGP with the potential of using professional certified Drivers, ECC Memory support if paired with a Server Motherboard, and a cool brand name that makes you really stand out of the crowd.

The lack of VT-d and TSX on K series parts makes you ask yourself what Intel thinks that "enthusiasts" are. Seems that they think that they are nothing more than mere overclockers with no other purpose that pay more money to get MOAR MHzs in enchange of everything else.

 

[NTMBK]

Yep, greed from endusers who wants everything without willing to pay.

What? There is no top end Haswell available which has all features enabled (TSX, all virtualisation features, GT3, L4 eDRAM, and unlocked overclocking). How can people spend more to get all the features enabled, when that SKU doesn't exist?

Anyway, hopefully we'll see such a part appear in the "Haswell Refresh" next year.

 

[SammichPG]

This is still a few months away, but I got tired of waiting so posting now.

Evidently, for real-world HPC apps (as Intel themselves describe), TSX can be good for as much as a 40% speed-up:

....

Oh well. Just thought I'd share. Until the actual presentation at SC13, there's little in the way of actual data, but for those of you in a similar boat as I am, wondering if choosing non-K for TSX can be worth it, this is a data point, at least. We'll have more solid data after SC13 to see how really broadly applicable the 40% claim is.

Buy a plain 4770 with z88 board then run all cores at turbo speed (it's just a setting in the bios).

Best of both worlds and you save money in the process.

PS the cpu might throttle with the stock cooler if you use all cores set on turbo.

 

[jpiniero]

The lack of VT-d and TSX on K series parts makes you ask yourself what Intel thinks that "enthusiasts" are.

Gamers. (As if you need a 4770K @ 4.5 Ghz to play console ports)