Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[InflatableMouse]

http://www.anandtech.com/show/7007/intels-haswell-an-htpc-perspective

[Arindelle]

thanks ... good info!

[glynor]

If you're considering this, and you can hold off, I'd wait a bit for the mobile stuff to flesh out.

If (and this might be a big if) you can get something like a i7-4950HQ CPU/board and stuff it into a HTPC case, that might be a much better overall choice than one of the desktop Haswell boards, particularly if you intend to use the iGPU.  A GT3 GPU+Crystalwell in a mobile form-factor, with all of those sweet, sweet power optimizations?  Yes, please.

Of course, the issues are:

1. Those are BGA package chips, which means they're soldered onto the motherboard.  You'll have to wait to see if ASUS or Gigabyte or someone decides to sell them on NewEgg for DIY setups.  But, it shouldn't be impossible.  Intel did it, after all.  This does, of course, mean no upgrades.  But, honestly... Intel keeps changing the socket for every new gen anyway, so that ship has pretty much sailed.

2. Apple is going to buy all of them that Intel is going to be able to ship for a while.  Particularly those with Crystalwell, and I wouldn't want a gimped one without it.

So, maybe by the holidays?  If you are in no hurry to upgrade the HTPC, I'd let it settle out a bit before jumping on a desktop Haswell.  They aren't terrible, by any means, but the GPU in them is basically only a tweaked Ivy GPU.  The good stuff is all in the GT3 and the mobile parts.

[dean70]

Be interesting to see how the GPU stacks up against the current AMD APUs & if it has enough grunt to run MadVr Jinc upsampling.

[glynor]

Be interesting to see how the GPU stacks up against the current AMD APUs & if it has enough grunt to run MadVr Jinc upsampling.

Look at the links I posted...

Beats the pants off of them, largely.

[6233638]

Is Cryatalwell anything more than L4 cache?

[InflatableMouse]

I find the GPU benchmarks dissapointing, less dissapointing for Iris Pro but still somewhat. The synthetic benches look good, but the real world performance wasn't that good.

Ganesh writes that although performance has improved, its not enough to offset the recent improvements in Madvr (anti ringing, jinc3, smooth motion). It might do deinterlacing a little bit better but I wouldn't bet on it doing 1080i@60.

They fixed 23.976hz, but they don't support 25 or 50hz - how insanely stupid is that?!? They completely messed up the control panel too, shuffled options around to the most illogical way possible. No doubt they test their software, but I seriously wonder if Intel ever uses their own stuff (had some bad usability bugs with iRST too).

Then when I read about Intel Iris Pro performance, its doing better but they promised GeForce 650M level of performance. They're between 15% to 40% short of even being close to that.


Having said that, I think the power consumption figures are very good. Impressive even. But if I still have to run Madvr at the same level I do now, what's the point? Saving a few bucks on my electricity bill? No thanks. I don't think its enough to offset the expense anyways.

[glynor]

Then when I read about Intel Iris Pro performance, its doing better but they promised GeForce 650M level of performance. They're between 15% to 40% short of even being close to that.

Having said that, I think the power consumption figures are very good. Impressive even.

That's true, they didn't get their target of the 650M.

However, they do get decently close (and they blow Trinity and all other iGPUs out of the water), and they do it at less than 1/2 the power consumption of a CPU+650M setup (and way under 1/2 the power of desktop Trinity except in the worst-case).  Again, for laptops and other power-constrained (or thermally constrained) environments, that's a Very Big Deal.  If you're in a big desktop (or even mATX) case with enough room for a discreet GPU then it isn't a big deal.  But, if you're considering using onboard graphics at all, it is by-far the best option.

Also, Ganesh wasn't reviewing GT3 at all, much less one with Crystalwell.  He was reviewing GT2 (the only GPU available on the desktop variants of Haswell).  The GT2 GPU is only a modest improvement over the Ivy GPU, so his results aren't surprising at all.  That's apples-to-oranges.  The GT3 has double the compute resources of the GT2, and a number of other improvements.

Again, it is MEGA-STUPID that they didn't ship the GT3 in any of their desktop chips (even the K series).  Hopefully they'll see the error in their ways at some point.  Perhaps it is just that they expect Apple to eat all of them that they can ship for a while, so why "waste" them on CPUs where the vast majority of users will be using a discreet GPU anyway.  Maybe we'll see more here when the lower-end desktop CPUs ship later this year.  The GPU is massive (over 60% of the surface area of the die, if Anand's crappy measurements/guesstimates can be believed).  As Scott Wasson from TR put it:

Once you know that, you may be shocked to hear this: Intel has no plans to bring a GT3e class chip to socketed desktop systems. These things are slated for BGA-style packages, for surface mounting into laptops and the like. I'm not sure why no one thought, "Hey, we have a CPU with a massive 128MB L4 cache. We should sell to people who want to buy it and put it into their systems." But apparently that didn't happen—or at least that guy didn't persuade everybody else.

Likewise, it is dumb that Intel's marketing idiots decided to segment the CPUs using TSX, which will greatly blunt the impact of that exciting new ISA extension.  Why, Intel, why?!?

Also, note: They completely re-architected their drivers for the entire GT series of GPUs, so we could (if they get off their butts) see some further improvements as time goes on.

As far as PAL land framerates...  Stop living somewhere weird.  

Is Cryatalwell anything more than L4 cache?

A massive 128MB L4 cache and texture cache for the GPU.  On-die, with a very fast interconnect directly to the CPU/GPU.  It largely mitigates the impact of having no fast GDDR-based framebuffer with its associated extremely power-hungry wide interface, at least for an iGPU solution.  It is a quite clever solution, and it doubles as a massive coherent cache for the CPUs as well.  The right kind of application will absolutely love having all of that cache space to play with.

If you want to read more about it, read the Anandtech article about the Iris Pro, the Tech Report Haswell review, and there's also some good stuff in the RealWorldTech Haswell memory article (as usual, though I imagine he's working on more now that it is out).

[6233638]

Again, it is MEGA-STUPID that they didn't ship the GT3 in any of their desktop chips (even the K series).

Yep. I like the idea of being able to have a system where my dedicated GPU is only ever required for gaming purposes, but GT2 probably doesn't cut it when it comes to madVR and Jinc 3 AR scaling.
I would absolutely upgrade my hardware if they brought the desktop chips in-line with the mobile ones as far as power saving and iGPU features were concerned, combined with an Nvidia GPU that could essentially turn off when using the desktop. (rather than still consuming a lot of power)

A massive 128MB L4 cache and texture cache for the GPU.  On-die, with a very fast interconnect directly to the CPU/GPU.  It largely mitigates the impact of having no fast GDDR-based framebuffer with its associated extremely power-hungry wide interface, at least for an iGPU solution.  It is a quite clever solution, and it doubles as a massive coherent cache for the CPUs as well.  The right kind of application will absolutely love having all of that cache space to play with.

Oh, I know that it's a lot of memory, and what that means for performance. I just wasn't sure if "Crystalwell" was a whole suite of features/changes, or if it was simply branding for the L4 cache.

They fixed 23.976hz, but they don't support 25 or 50hz - how insanely stupid is that?!?

Wow, I was not aware of that. Intel has always been bad when it comes to resolution/refresh rate support, but that takes the cake.

As far as PAL land framerates...  Stop living somewhere weird.  

The funny thing is that 90% of the time, I actually want PAL content to be played back at 24Hz anyway (and I never want 24/1.001 Hz) but I've used the 50Hz setting quite a few times with games that can't quite run at a v-synced 60fps on my GPU.
Alan Wake was a good example of that - with the really nice volumetric lighting turned on, the game was dropping to about 55fps at times during gameplay. Set the game to 50Hz, and I was able to use v-sync and had a perfectly smooth gameplay experience without any screen tearing. (though the pre-rendered cutscenes were 30fps)

[glynor]

Oh, I know that it's a lot of memory, and what that means for performance. I just wasn't sure if "Crystalwell" was a whole suite of features/changes, or if it was simply branding for the L4 cache.

Branding for the GT3+eDRAM solution.

Still, very cool.  Custom silicon with a coherent memory interface, built on the same process node as the CPU (though using the SoC version of the process instead of the CPU version for power reasons).  Quite nifty, and interesting for the future.

[InflatableMouse]

Maybe AMD is a better choice this time around.

[glynor]

Maybe AMD is a better choice this time around.

AMDs x86 compute cores suck though, and their power is off the charts.

I agree with the central premise though...  If only intel had put GT3 in at least a couple desktop CPUs.  Sigh...

[InflatableMouse]

I've not seen any power measurements yet from Richland, I'm hoping its improved. Idle usage was particularly bad IIRC.

[glynor]

Richland is based on Piledriver, not the next-gen Steamroller, cores, so we know quite well how it'll perform.  It is basically a minor tweak to their existing Trinity chips.  There are some nice optimizations in there, and they should be somewhat improved performance-wise, but nothing huge (you can listen to the recent Tech Report podcast for some discussion of Richland during the Kabini section).  We already know the TDPs.  From the article you linked:

AMD sent over two Richland parts just before I left for Computex: the 100W flagship A10-6800K and the 65W A10-6700.

So, 100W, and 65W, much like their current line of chips.  And AMDs chips have been running closer to their TDPs than Intel's of late.  Don't expect that to change much.

The other issue I have with AMDs CPUs isn't with the CPUs themselves at all, but with the platform.

I used to be a HUGE AMD fanboy during the Athlon/Athlon XP/Athlon 64 days.  But, I have to say, after Nvidia stopped producing chipsets for them, I had lots of trouble (not that Nvidia's chipsets were trouble-free either, though my ASUS A7N8X motherboards were some of the best I've ever owned -- I had two different models).  Since then?  Flaky storage and USB controllers, mostly.  That's improved somewhat of late (and both are integrated into the CPU now, of course), but I've still had way more driver "fiddliness" with all of the modern AMD chips I've used.

I like their GPUs still a lot (though my next GPU might be an Nvidia -- we'll see), but their chipsets have been troublesome, and now their x86 cores are way behind and their power usage is often double a (better performing) Intel.  For servers, their "big, multi-core" CPUs are still competitive (mostly on price), but for home use, single-threaded performance still means a lot, and they're... Not great.

That's been the net effect of Intel's extremely successful Tick/Tock strategy.  Because they're releasing things yearly, with that repeating "architecture/improvement" cycle, they aren't having massive generational leaps like they used to, which leads people to get all excited about the next-release and then scream "yawn" when they come out.  But 10% year after year, with lower and lower power budgets all the time, and the "either/or" strategy they've been using on architecture details, have been killing it over the medium term.  And the old "big" architecture changes were usually 3-4 year cycles.  Plus, if they made a bad call (like with Netburst) they had to scrap everything and start over, which meant 3 years for AMD to catch up.

There's been no catching up since Tick/Tock started.  It was a very, very smart thing they did with the original Core launch.  They had to swallow a lot of pride to throw Netburst out completely and "go back" to the P3 architecture, but it was an immensely smart move when coupled with the new "refinement based cycle" they've been doing.  No one has been executing like them, and they've been doing it for years now, with only the smallest of hiccups (with Sandy).

[glynor]

The problem is, with no serious competition in the desktop space, they're lead to do things like the Haswell desktop launch, and skimp.  Why sell the fancy GT3 (with its huge increase in die area, and therefore cost) when Apple isn't demanding it, everyone else is demanding the cheapest parts they can get, and AMD is nowhere to be found (and might put themselves under all by themselves anyway)?

ARM has them scared (and they should be), so they're totally focused on mobile.

Smaller dies on the desktop saves them money and props up margins for all the lower margins they're going to be forced to swallow to kill the ARM threat.  ARM itself is still a real threat, but they have no foundries, and Intel is still a node or node-and-a-half ahead of everyone else.  The RISC vs CISC advantage gets smaller with ever generation (and microops already defeated most of it, modern Intel cores are CISC on the outside and RISC on the inside).  The main thing they need to worry about versus ARM is all of that big x86 compatibility logic in the decoders on the front-end.  But each die shrink, every other year, makes that a smaller and smaller proportion of the core (because it is relatively static, so long as they don't add any die-area-expensive extensions).

I wouldn't count them out in that space either.  Once they can get that penalty small enough, they can use their foundries and execution wizardry to pound ARM into the dirt.  It isn't decided by any means (and it is getting later and later with every gen), but bet-your-butt that Apple has a version of iOS for x86 in the labs, running in parallel, just in case.

[InflatableMouse]

I used to be a HUGE AMD fanboy during the Athlon/Athlon XP/Athlon 64 days.  But, I have to say, after Nvidia stopped producing chipsets for them, I had lots of trouble (not that Nvidia's chipsets were trouble-free either, though my ASUS A7N8X motherboards were some of the best I've ever owned -- I had two different models).  Since then?  Flaky storage and USB controllers, mostly.  That's improved somewhat of late (and both are integrated into the CPU now, of course), but I've still had way more driver "fiddliness" with all of the modern AMD chips I've used.

Same here. I had a lot of AMD stuff until 9 or 10 years ago. One of the coolest boards I owned was a 386dx40 because it was upgradable to a 486dx2 and had a Vesa Local Bus slot (which I found out the hard way didn't work with the 386). The a7n8x deluxe was the last AMD board I owned until it got zapped in a lightningstorm. It still worked but some of the onboard components died.

Anyway, there goes that idea then . Intel needs a worthy competitor. I keep hoping AMD makes a comeback and bites them in the butt once again.