Anandtech Haswell review is up!

[krumme]

My two cents to the discussion:


- It seems a chip built from the scratch for mobile. It packs a lot of punch, both in graphics and CPU power, all that in a power sipping power envelope. In

- I'm kinda impressed by the top iris part. 70-80% of the GT650 + Quad Core Haswell in a 47W power envelope.It will be nice for very small form factors.

- Overall, this is APU doing right, at least from the consumer POV. Once Intel gets to 14nm, it will be APU done right from a cost POV too.

- It might be able to establish itself on tablets.

Selling cpu above 150usd for both the consumer market and the business market, outside of server, is just plain uninteresting from an economic point of view.

The processors we have seen until know is just not needed. They dont give any user benefit compared to IB imho.

I think we will have to wait until 14nm to see if this arch is really good.

And we need to see the more relevant mobile and ulv i5 chips to get any sense of this, and the tablet variants. I am sure they will impress, because it must be designed for lower voltage.

Its obvious iris colapses at resolutions above 900. Give it full hd and i think its going to jitter.
Give us some minimum framerates. Give us better assessment of the solution.
I dont think the technology is ready yet. Intel seems to agree, because it prices it like a halo product. I mean seriously a cpu for what 600usd? - Anand can say Apple all day long, for sure, they are not going to pay even a third of that. Secondly, why should they want a solution that just colapses instantly at retina resolution.

Secondly the efficiency on the serverside must be really disapointing. Not that it matters because there is not any competition, and when there comes a little competition here, they are already at or near 14nm.

Btw the desktop review was the most biased i have read on anandtech for 14 years. Haswell is probably a nice cpu. But why just not write its meant for the mobile market instead of this "fan of" and "up to 20%". It just hurts the eyes, and its embarrasing.

 

[Arachnotronic]

"The only downside is cost" yeah you're not joking about that.

Cost is the #1 factor in being a viable solution for anything - if you can't make it cheaply then it has to be worth the price.

If it isn't worth the price? Well I guess that's why the PC makers are ignoring it and going with discrete instead.

GT3e is garbage at the price and does not affect AMD or Nvidia's bottom line in any way at all. In around 6 months time I expect AMD to match its performance with Kaveri at 35W and 1/3rd of the price. Guess which one of these will sell by the millions and which one won't?

What makes you think that AMD can get a huge lead like this? Further, if AMD can deliver a superior product, then why would it charge 1/3rd of the price?

Is AMD a charitable organization?

 

[insertcarehere]

The big question for mobile will be how significantly memory bandwidth affects the performance of the gt3 igps, as these will be used in lower tdp and (hopefully) lower priced chips as well, only without the edram on die.

 

[R0H1T]

The big question for mobile will be how significantly memory bandwidth affects the performance of the gt3 igps, as these will be used in lower tdp and (hopefully) lower priced chips as well, only without the edram on die.

AFAIK there's no GT3 based chip without an eDRAM, be it on a desktop or mobile platform, I could be wrong though :\

 

[mrmt]

The processors we have seen until know is just not needed. They dont give any user benefit compared to IB imho.

I can see your point in desktops, but not mobile. Haswell is a descent improvement for mobile.

I think we will have to wait until 14nm to see if this arch is really good.

I'm pretty confident that Broadwell will be a hit. We'll probably see more hardware and more L4 memory outside of the top SKUs, and premium SKUs with even more L4 cache.

Its obvious iris colapses at resolutions above 900. Give it full hd and i think its going to jitter.

Where did you see this?

 

[insertcarehere]

AFAIK there's no GT3 based chip without an eDRAM, be it on a desktop or mobile platform, I could be wrong though :\

Notice the "GT3" labels on the 28W and 15W TDP graphics levels, I think the "e" in "GT3e" refers to the embedded edram.

 

[SiliconWars]

What makes you think that AMD can get a huge lead like this? Further, if AMD can deliver a superior product, then why would it charge 1/3rd of the price?

Is AMD a charitable organization?

Selling powerful APU's is AMD's strategy, it's not Intel's strategy. If it was then why are they only selling GT3e in expensive parts that are being ignored by OEM's?

The GT3e chip makes no difference to AMD. It does not make their APU's slower than what they are competing against so how can it affect them? Do you think OEM's are going to stop buying current AMD chips because of this? No.

When Intel has a compelling product vs AMD in AMD's market segments, then AMD can be worried. When Kaveri comes out the gap will be bigger than it's ever been where it matters to AMD. Intel can keep the $400 and up APU market.

 

[R0H1T]

Notice the "GT3" labels on the 28W and 15W TDP graphics levels, I think the "e" in "GT3e" refers to the embedded edram.

Yes the GT3e parts would incorporate eDRAM but looking at that slide I'm wondering how segmented the new mobile SKU's from Intel, Haswell onwards, would be keeping in mind the insane number of IVB models currently available, it'd take a mammoth effort to get them under a single chart

 

[ShintaiDK]

Its obvious iris colapses at resolutions above 900.

The benches on anandtech at 1680*1050 shows its not collapsing.

 

[CHADBOGA]

And we all know why is that !

Yes, People value CPU performance, and the niche where Intel's IGP isn't good enough AND a discrete isn't required, is too small to make much of a difference to AMD's fortunes.

 

[AtenRa]

can any iGPU game as well as even a $100 GPU like a 7770? No. Not really even close.

Obviously not, HD7770 die size is 123mm2 at 28nm + 1GB GDDR5. It has 640 GCN cores when Trinity/Richland at 32nm only has 384 VLIW4 cores and DDR3. Also, HD7770 has a 80W TDP alone when Trinity/Richland APU (CPU + iGPU) tops at 100W TDP.

 

[AtenRa]

What makes you think that AMD can get a huge lead like this?

Haswell GT2, Core i7 4770K cant even touch Trinity A10-5800K in games, Richland will be a little faster and Kaveri will bring at least another 50% in iGPU performance. Unless Intel will bring a socket 1150 Core i3 with GT3, it has no chance in Hell to even be competitive in the sub $150 market.

Further, if AMD can deliver a superior product, then why would it charge 1/3rd of the price?

Is AMD a charitable organization?

AMD needs to sell high volume sub $150 products with its Desktop APUs, but they could bring higher priced Mobile/embedded APUs.
Edit: Also, AMD cannot and will not sell its APUs higher because it will make its APUs irrelevant when you will get higher performance dGPUs with lower price.

 

[Riek]

GT2 ties Trinity in mobile. GT3e destroys it. Even if Richland is 30% faster than Trinity in mobile, which is VERY optimistic (personally I think 15%), it will not be enough.
In the desktop space, GT3e will be faster than the A10-6800K as well, so Intel will have the fastest solution all around. The only downside is the cost, of course. But given Intels confidence that they will make an even larger jump with Broadwell, the whole situation is very discouraging for AMD. It was only a matter of time until Intel puts their process advantage to good use.

Ofcourse it will. But it comes at a high price... e.g. power consumption.


With a TDP of 55W it is far away from AMD 35W APU. While we can compare those for benchmarking, they shouldn't be seen in the same laptops since those are not designed to handle >50W of the cpu/igp alone.

Would be interesting to see what intels GT3 can do at lower TDP values. But seeing the drop in performance for the 47W in anand test i would assume performance will drop to the point where the difference will be alot smaller between AMD and Intel. (e.g. 35W and/or lower).

At least intel suprised me in the positive way with their best IGP.

 

[USER8000]

With a TDP of 55W it is far away from AMD 35W APU. While we can compare those for benchmarking, they shouldn't be seen in the same laptops since those are not designed to handle >50W of the cpu/igp alone.

The test was done with the CPU configured for a maximum Turbo TDP of 69W in a desktop chassis.

Hopefully,we will start to get more reviews in actual laptops and see how closely performance tracks the Intel test system.

 

[Lepton87]

!

And 128MB of eDRAM is useless compared to even 1GB of GDDR5. Utterly useless. Unless everything you're doing fits in the cache (doubtful), you're still going to hit the shared DDR3, which brings things to a screeching halt in comparison to even a crappy discrete card like 7750/650. Bring a real discrete card like 7850 or 660 to the table and ALL iGPU looks like the garbage-bin stuff it really is for real gaming usage.

Don't get me wrong, improvements in iGPU are always welcome, but the predictions even waaaaaay back when AMD bought ATI, and the Larrabee hype have ALWAYS fallen flat, and nothing is going to change that for the forseeable future. Discrete is a moving target, and if anything, the gap is only growing as time goes by rather than getting closer. All these years later, can any iGPU game as well as even a $100 GPU like a 7770? No. Not really even close.

You are so overestimating bandwidth advantage on such low performance parts my brother has 650m ddr 3 and it's only 20% slower at stock, 15% after OC. Over 100GB/s(80GB/s at stock) for 384 kepler cores is a waste, 650m has what 28.8 of memory bandwidth? We compared overclocked cards, his card had 34GB/s or so (I don't remember the exact memory clock) vs over 100GB/s of 660M and the increase in performance was just 15% after OC. GPU clock was the same, both laptop manufacturers limited the OC to +135MHz offset def gpu clock was the same. And 650M/660M is decently faster then Iris.

Am I the only one who wants an unlocked GT3e on an LGA socket? I couldn't care less about the integrated graphics but that 128MB would act as L4 cache for the CPU. Anand tested this but very sloppily.

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/18

On Average he got 4.4%(3% including some unexplained loss in performance) IPC advantage. But what I really would like to know is how that 128mb of eDRAM would fare in games, games like cache. Of course with a discreet card. A shame he didn't test that

 

[mikk]

I was talking about performance per die size. But, another way to see how inefficient the Intel Iris is, have a look at the AMD Trinity A10-4600M at 35W and 1600MHz memory against the Desktop Iris 4600 in Core i7 4770K with 84W TDP and 2400MHz memory.

The iGPU in Trinity is close to 98mm2 at 32nm, Iris 4600 is close to 87mm2 at 22nm. Trinity iGPU at 22nm TriGate would be close to 55-60mm2.

Nonsense. Haswell GT2 iGPU is ~55 mm² big while Trinity iGPU is 103 mm² big. Intels 22nm scaled 1,5 over 32nm. Process adjusted it is 82.5 mm² HSW GT2 vs 103 mm² Trinity. Furthermore Gen 7_5 is a 1/4 DP design while Trinity is 1/16 DP only.

 

[Chiropteran]

All this raw performance is nice and all, but I don't really trust Intel to deliver proper Igp driver support. There will be issues, and divers won't fix them all, based on historical Intel driver suckage.

 

[Lonbjerg]

All this raw performance is nice and all, but I don't really trust Intel to deliver proper Igp driver support. There will be issues, and divers won't fix them all, based on historical Intel driver suckage.

You are basing this on the "old Intel" and their lack of interests in IGP performance....not the "new Intel" were IGP's matter...right?

 

[Xarick]

Okay.. so I need some interpretation. Coming from a q9550, is this better bang for buck than ivy bridge or no?

I am going to need a mobo, ram and proc so all those need to be taken into price consideration.

 

[ShintaiDK]

Okay.. so I need some interpretation. Coming from a q9550, is this better bang for buck than ivy bridge or no?

I am going to need a mobo, ram and proc so all those need to be taken into price consideration.

Yes.

 

[WhoBeDaPlaya]

At this moment? No, since IVB CPUs and mobos are discounted. This is especially true if you can score an IVB + open-box mobo combo at MicroCenter (picked up two 3770Ks and two open-box Sabertooth Z77s for ~$350 after tax).

Assuming similar overclocks and the same cooling solution, you'll basically end up paying ~30% more for ~10% higher performance,
new platform and extensions notwithstanding.

 

[AtenRa]

Nonsense. Haswell GT2 iGPU is ~55 mm² big while Trinity iGPU is 103 mm² big.

So, going from Core i7 4770K GT2 of 177mm2, add another 55mm2 (for 20 more EUs) will make the Core i7 4770R GT3 be 232mm2.
You missing another 32mm2 somewhere

 

[Abwx]

Nonsense. Haswell GT2 iGPU is ~55 mm² big while Trinity iGPU is 103 mm² big. Intels 22nm scaled 1,5 over 32nm. Process adjusted it is 82.5 mm² HSW GT2 vs 103 mm² Trinity. Furthermore Gen 7_5 is a 1/4 DP design while Trinity is 1/16 DP only.

Scaling is 1.63 , easily extractible from public data.

 

[Enigmoid]

Haswell GT2, Core i7 4770K cant even touch Trinity A10-5800K in games, Richland will be a little faster and Kaveri will bring at least another 50% in iGPU performance. Unless Intel will bring a socket 1150 Core i3 with GT3, it has no chance in Hell to even be competitive in the sub $150 market.

Trinity is about 30% faster, Richland performance increase is negligible. Note that these were tested on high settings (where historically intel doesn't do well) where almost every single game tested was unplayable even at 1280 x 800.

 

[mikk]

So, going from Core i7 4770K GT2 of 177mm2, add another 55mm2 (for 20 more EUs) will make the Core i7 4770R GT3 be 232mm2.
You missing another 32mm2 somewhere

Haswell Quadcore percentage for GT2 is 31% from the whole Die. The whole Die is 177 mm². Now you can do you math. If Anand measured on his own 264 mm² means the official size should stay below 250 mm² because die size measurements by hand are always bigger than what we get from Intel/AMD/Nvidia. I measured 240-250 mm² based on the first GT3e Die picture from Vr-zone. Same for Haswell GT2, we expected ~185 mm² and the official die size is just 177 mm² big. Since Haswell GT3e is no doubling (non-slice parts won't double) it should add ~50 mm² more GPU size. I don't know if there are dead areas within its die or maybe Intel had to add some die size for its edram connection.

Scaling is 1.63 , easily extractible from public data.

Link? Based on all Ivy vs Sandy comparisons I have process scaling was 1,5.