Back in February, I penned a bearish article on Intel's (NASDAQ:INTC) new Haswell line of CPUs. It got a lot of attention. Back then I looked at what Haswell brought to the table from early engineering sample benchmarks, did a little conjecturing and a ton of reading between the lines of a number of technology websites and came to the conclusion that on a cost/performance basis, Haswell would not measure up to the latest offerings from Advanced Micro Devices (NYSE:AMD) or, for many users offer a compelling upgrade over the outgoing Ivy Bridge series (inventory to be sold at a discount) or tempt enthusiasts to upgrade their existing boxes. Now that Haswell is here and its top end parts have been benchmarked, we have an idea about the efficacy of the choices made and what they portend for the future.
The Good, the Bad and the Apple
In that article, I asked Intel to show me that it was capable of producing a GPU that wasn't god-awful. The IRIS Pro Graphics 5100/5200 GPU and cache has shown me that it can. All the other GPU variants for Haswell, however, are still woefully inadequate on both a raw performance and cost/performance basis when compared to Trinity, no less Richland, which is both significantly faster and more power efficient than Trinity.
There is no doubt that Haswell has mostly succeeded in improving Intel's core offering in two important areas, graphics performance and power management. This review of the i7-4770K implies that Haswell does, indeed, have some issues with power management, possibly vindicating analyst Alex Gauna of JMP Securities who called Intel out on it, but we'll have to see how the lower TDP parts fare as those have not been released yet. Anandtech pointed out in their review of IRIS 5200 GPU earlier in the week that Apple (NASDAQ:AAPL) was the driving force behind Intel changing its ways and dedicating some of its prodigious resources to graphics. Apple demanded better integrated graphics and longer battery life and Intel had to go to the drawing board and create a solution. The problem was always going to be, however, at what price was Intel going to offer up these top-end chips?
Well, we have our answer. The very best Haswells with IRIS-5200 GPU -- AKA GT3e with 128Mb of embedded DRAM as a Level 4 cache -- are hideously expensive.
Intel Core i7-4950HQ
Intel Iris Pro 5200
Intel Core i7-4850HQ
Intel Iris Pro 5200
Intel Core i7-4800MQ
Intel HD 4600
These are the best all-around performing processors on the market. Graphics performance of these parts is superior to that of even AMD's desktop Trinity APUs and will also beat out the best that Richland has to offer. Of that no one is questioning. And for Apple, who will most likely be using these in the MacBook Pro refresh and the biggest customer for these chips, it will likely be able to pass most of that cost onto its customers. That is pretty much the best news that I have for Intel shareholders here.
The Gap Into Power
Yes, Intel is closing the gap between itself and AMD and Nvidia(NASDAQ:NVDA) in terms of graphics performance and throughput. And I agree with Anandtech that making Crystalwell an L4 cache was an elegant solution that yielded excellent performance improvements due to the massive increase in memory bandwidth the cache affords. But this has come at a significant increase of die size, complexity and, ultimately cost, which Apple laughed Intel out of the room over, according to Charlie at Semiaccurate (subscription required). So, while I expect an IRIS 5100/5200 will be inside the next MacBooks, I don't expect Intel to be selling them to Apple at 65% gross margins, either.
This is a Pyrrhic victory if I've ever seen one.
As to why we won't see Crystalwell on the lower TDP parts the reason is, as Anandtech points out again (last two paragraphs on that page), the cost to performance is not viable, which means that we won't be seeing Intel competing graphically in that portion of the market where most PC sales are currently, sub $600. It is not possible to add another $40-60 to the cost of a Core i3 so that it can compete graphically with a 25W Kabini or Nvidia discrete chip and wind up with a salable device. The lack of SKUs in that segment again, confirms everything I said in February about Haswell as a product, not as a piece of technology.
The Future Path
Moreover, while Crystalwell was a good solution it is, unfortunately, not the most elegant one. Memory bandwidth is the issue for all APU-like chips as is how the GPU and the CPU share data in the same memory address. AMD's hUMA is, ultimately, the most elegant solution because it addresses both issues simultaneously.
- By allowing the CPU and the GPU to access the same physical block of memory it cuts out having to copy pages of data around into a separate block, thereby reducing not only time and improving timing, but savings scads of power by not having to power up and down redundant memory chips.
- hUMA also drives simplification of the overall chip design. The transistor savings from hUMA creates the opportunity to integrate a wider memory controller and hence wider memory bandwidth on-die while keeping the die size smaller, which improves the efficiency of heat dissipation and drastically lowers cost and should increase margins.
Now, add in the idea of spending a little extra money on your high end APU to get one equipped with a GDDR5 memory controller and building your system with that and the need for discrete graphics drops like a rock. This is just part of the reason why I am bearish on Nvidia. On the one hand, Intel will likely be throttling the PCIe interface on Broadwell along with pushing Crystalwell even farther along the performance path. On the other, we have AMD simplifying chip design, lowering cost, raising performance while offering choice of memory bandwidth based on need.
Intel and AMD are taking two completely different paths to improving graphics performance through wider GPU memory bandwidth and the next 2 years will tell us which one is more future proof. My bet is on AMD as Crystalwell looks to be a solution that will yield its biggest gains with its initial implementation. hUMA is a completely different way of building a modern SoC and because of that, opens up design options that simply were not practical previously.
So, while Intel has grabbed headlines with its massive improvement in graphics performance with the IRIS 5200 and Crystalwell, the devil is in the details.