THE BATTLE FOR PROCESSING SUPERIORITY
What does the future of high-end tell us about the destiny of mainstream processing?
NOTHING’S BETTER for the tech industry than a little competition, and that’s something Ryzen has very much brought to the table over the last two years. Consider this: Since the beginning of 2017, we’ve doubled the number of cores available to mainstream processing parts. From the very bottom rung, with Pentium and Athlon CPUs, all the way up to the Core i9s and Ryzen 7s, Intel and AMD have produced an epic lineup of chips, with twice as many threads as before. Look at high-end desktop processors, and that figure has almost quadrupled. Crazy. In early 2017, Intel’s 10-core Core i7-6950X would set you back $1,723; today, you can buy AMD’s TR 2990WX, with an incredible 32 cores, for just $6 more.
Are we likely to stop here? We doubt it. With AMD announcing its EPYC 2, Rome, server processors, due to launch in the second quarter of this year, the core war is still very much in full flow. Packing 64 cores and 128 threads on a single processor, with a far superior memory interface design than previously, supporting PCIe 4.0, and on TSMC’s 7nm manufacturing process, AMD’s next-generation server parts look as though they’re going to be phenomenal computational monsters. There are even rumors of a 29.4 percent performance increase in singlecore IPC, all down to architectural advancements and floating point calculation optimizations, rather than simply increased clock speed and voltage.
EPYC 2 is important for a number of reasons. Firstly, because it runs off the successor to the ridiculously good Zen architecture, the 7nm Zen 2, which will also be found in the third-generation Ryzen parts (also due this year). And secondly, as the first-gen Threadripper processors were essentially cut-down variants of the first EPYC server parts, it gives a good guess as to what Threadripper 3 might look like, too.
The big question is: What’s going to transpose down? Apart from eight-channel memory and the PCIe 4.0 enhancements, which we’re hoping make it, what will be more intriguing is how many cores are brought along with it. We imagine AMD will likely keep the overall design of the two SKUs the same, to help reduce manufacturing costs, but as we’ve already seen 32 cores on the mainstream Threadripper platform, it does leave questions as to how many cores will make the jump. At a guess, we may see a highend 3995WX, likely the same price as Intel’s unreleased 28-core HEDT part, potentially touting 48 cores and 96 threads, for the true amateur renderers and CAD-CAMmers.
We also know this is the first time AMD’s deviating away from a singlesize transistor design across the chip. Thanks to its Infinity Fabric, the company will be going to a modular transistor size architecture design, with things such as the ridiculously fat I/O controller staying on 14nm, while the main core complexes will be on 7nm.
Hopefully, most of that should filter down to Ryzen 3. We’d love to see a 16-core Ryzen 7 3700X, with support for PCIe 4.0, and perhaps more PCIe lanes in general for better PCIe SSD support, hooked up in a similar manner to its Threadripper kin—16 for GPUs, and an additional eight for PCIe storage would be a fantastic solution, and really future-proof you as PCIe SSDs break that 3.2GB/s barrier.
THE QUEST FOR 10NM
If the reported figures are correct, and AMD’s samples of the EPYC server parts prove as mighty as they currently are, Intel may lose the top spot in multi and single-threaded applications. That would be the first
time since the early Athlon days that Intel has lost its grip on the top spot of processing performance.
It all rests on whether Intel can hit its 10nm volume production targets in its current timeline. Although originally outlined to make it to market by 2016, delays and complications have lead to multiple refreshes of Team Blue’s 14nm architectures, with its current roadmap suggesting the company should be able to make bank on its volume production targets by the second half of 2019, leading to a late fall release at the earliest. That said, there is one 10nm processor currently in existence: the Core i3-8121U. It’s an ultra low-power mobile part by design, but in a rather bizarre twist, the integrated GPU is switched off. According to sources, this is due to logic incompatibilities between the GPU and CPU that are currently too troublesome to rectify.
Going from 14nm to 10nm should allow Intel to increase its transistor density by a factor of 2.7. However, as in the past, this likely won’t lead to a huge performance increase, so far as IPC is concerned, and with no evidence or rumors on benchmark performance, we can’t say how powerful these parts will be. All of this leaves AMD with eight to nine months of open market to capitalize on, if it can successfully get its 7nm processors out of the door in time.