Tom's Hardware put up an initial review of Intel's (INTC) upcoming new CPU line, code-named Haswell, recently comparing the top of the line Core i7 4770K to its predecessors from the previous generations Ivy Bridge and Sandy Bridge. Last month I posted an initial impression of Haswell's performance that suggested the gains over Ivy Bridge would not be significant and that, to me, threw up serious red flags as to Haswell's place in the upcoming chip marketplace. These results from Tom's Hardware, despite the reviewer's attempts to spin them otherwise, confirm what we saw from the leaked results in early February which prompted my last article on the subject.
The first thing glossed over in the review is that the Ivy Bridge Core i7 -3770K Haswell was tested against is a 77W TDP part, while the 4770k is an 84W TDP part. For all intents and purposes there is between a 3% to 8% increase in CPU performance clock tick for clock tick between Ivy Bridge and Haswell. Since increasing instructions per clock (IPC) was not the design team's primary goal, this can be mostly overlooked. It is likely that the wider thermal envelope was allocated for the improved GPU. I will give Intel the benefit of the doubt there.
I'm not going to go into the specifics, but in the real world tests that were conducted there were some things Haswell did significantly better than Ivy Bridge, but the performance gains for these desktop SKUs look, overall, to be very bad value propositions over Ivy Bridge unless Intel introduces them at similar prices to Ivy Bridge. But then, why is Ivy Bridge still in the channel? Honestly, the desktop Haswell SKUs look DOA to me, because Ivy Bridge will survive in the channel through the end of this year and no one needs the GPU or the power savings the chips offer.
The one thing that I will give Intel credit for here is their not putting the high end GT3 GPU with or without the Crystalwell on-die graphics memory on these high-end CPUs. Anyone willing to pay $500 to $1000 for a CPU is buying a discrete graphics card to go with it, so making them pay for a big integrated GPU is really churlish and even Intel can't abuse their highest margin customers that badly anymore. So, the HD 4600 (GT2) on die GPU is what we see in the review and its performance comes in pretty much in line with expectations -- about 15% higher than the HD 4000.
What is interesting about this test is that it came out just 24 hours before the initial reviews for AMD's (AMD) Richland APUs were due to be published. And it is well known that Tomshardware has been a reliable conduit for Intel's PR department in the past, so take the timing of this report with as much salt as necessary. This is especially true when the reviewer ends the article with the following statement:
Almost certainly, however, a (mobile) part with twice as many execution units and 128 MB of L4/eDRAM at 1.2 GHz would blow Trinity out of the water in games. A comparison to Richland probably won't change much there. And we'll likely need to wait until 2014 to see how Kaveri affects AMD's position.
He's referring to Haswell's GT3 GPU with Crystalwell here and yes, it is likely that it would do exactly what he claims - given the caveat that Intel is not even making that part as the GT3 will be downclocked to 800 MHz to save power and Intel's graphics drivers are generally atrocious and the image quality on par with the drivers. But the real question is, of course, at what price would a full blown mobile Haswell GT3 with Crystalwell be offered? In what universe will that part be price competitive with a Richland A10-5750, AMD's highest end APU of the moment? Not any one that we inhabit.
Intel, in effect, is asking OEMs to pony up similar money for just the Crystalwell memory upgrade - reports are as high as $60 -- that AMD would likely charge for the entire APU. Currently the top end desktop Trinity APU - the A10 5800k - costs around $129. What could the 35W and 17W versions cost?
Making that comparison is really intellectually dishonest because only enthusiast portables and Apple's (AAPL) Macbook Pros have the margin built into them to afford such a beast. And since the market is now completely focused at the sub-$600 price points to compete with tablets, we have to be realistic and look at how the SKUs to be sold in volume stack up overall. Graphics performance for the GT2 does not come close to rivaling that of the outgoing Trinity APUs - even Tom's Hardware acknowledged that grudgingly -- and we know that at the high end Richland will offer ~20% more graphics performance over Trinity. So, knowing, at best, the GT3 will be ~30% faster than the GT2 clock for clock that puts its performance in line with Trinity.
At the lower end, however, is where AMD's initial benchmarks are showing the greatest gains for Richland with the replacement for the A6-4455 seeing a full 40% improvement in 3D Mark scores. This is due to a mix of the improved power management features in the chip along with higher clock speeds overall which allows the chip to spool up the GPU to its full turbo clock speed for far longer than Trinity could in the same thermal envelope.
Against the Core i3 and i5 SKUs these will be competing against, it looks like there is a real opportunity for AMD to take some people by surprise here as Richland should be cheaper and more overall useful than the GT3-equipped Haswell chips in the typical $350-$600 laptop you can pull down off of Amazon (AMZN) or Best Buy (BBY) without much work. For now, Richland will be competing against the 13W Ivy Bridge-Y series which it will thoroughly outclass dollar for dollar.
This is not to say that the performance gains Intel has apparently made with the iGPUs in Haswell are not impressive. They are. Along with the power management features, the chips will go a long way to improving Intel's position in the low power arena and improving notebook usability. But, and this is what is important, they are still playing serious catch up to AMD's heterogeneous architecture which is where the future of mobile computing lies.
The GT3 may have caught Trinity but click this link (and translate it from Chinese) to see what a drop-in replacement of a Richland A10-5750 did to the performance of a gaming laptop. Bottom line? 12% CPU performance increase and a 13% increase in graphics performance over a Trinity CPU mated to the 7970M discrete GPU.
The point of these comparisons is to look at the real world value proposition of a part and assess its usefulness to the paying customer. Richland is more than just a simple refresh of Trinity. It is the chip Trinity should have been. Haswell is the chip that Intel should have made two years ago to begin in earnest the competition with ARM Holdings (ARMH) but did not.
Richland, not Haswell, has the opportunity to create real value-based gaming laptops and all-in-ones. This market doesn't exist currently because the technology has not quite been there. Now it is and it will only improve with Kaveri and potentially Broadwell.
So, at this point, Haswell will launch in June as a CPU without a clearly defined market niche. The desktop parts are not compelling upgrades over Ivy Bridge. The mobile versions will be competing not only with Richland - which is a sincere tragedy -- and AMD's real stalking horse Kabini/Temash, but also with the 13W Ivy Bridge-Y CPUs that will be around until the end of the year. Overall, Intel's roadmap looks like a complete mess and as an investor I would be very worried about these benchmark numbers.
Additional disclosure: I have a couple of AMD computers operating in my house.