At last week’s Hot Chips, it was beginning to seem like we wouldn’t see too much in the way of raw speed. Hot Chips is one of the key conferences for chip designers, but hot's no longer in fashion, or even really feasible. Nowadays, as transistors have gotten leakier and mobile computing's become big business, the focus is on energy efficiency: How to get the most bang for your watt (or, fraction thereof).
So I do have to hand it to IBM for carrying the torch. In the last talk of the conference on Wednesday, senior technical staff member Chung-Lung (Kevin) Shum revealed some specifics on the “hearts and guts” of Big Blue’s new zNEXT mainframe processor, which will ship with a speed rating of 5.5 GHz. According to the company’s press release, that makes the processor either the “world’s fastest chip” or, more cryptically, the “world’s fastest chip running at 5.5 GHz” (there seems to be a missing comma, or this is a very limited boast).
The basic stats: the chip measures less than a square inch and was made with a 32nm chip manufacturing process giving the processor denser features than its 2010 predecessor, the 5.2 GHz z196, which was made using a 45nm process). The new processor boasts 6 cores to z196’s four and winds some 7.68 miles of wire through 15 metal layers. Shum says the chip was designed under the “same power constraints as prior years,” suggesting the company has increased the clock rate without drawing significantly more power.
Getting that speed boost involved more than just shrinking transistors. Shum listed extensive changes to the microarchitecture (so many, that afterward the session chair commented that Shum covered pretty much everything you might read in a textbook on the subject). Like the BlueGene/Q supercomputer chip IBM unveiled last year, this new mainframe chip implements transactional memory, which helps reduce locks on memory that can slow down computation. Some changes Shum outlined were fairly basic, such as reducing, for example, the physical separation of caches to cut down on latency.
As The Register noted, these mainframes are still a very profitable venture for IBM: pour $1 billion into chip development; get $7 billion back in sales. But given “Big Iron’s” small role at a conference dedicated to fast chips, you have to wonder how long the effort will continue. Some analysts say the traditional mainframe’s days are numbered (no, really, this time!) as it competes with systems based on cheaper chips.
Rachel Courtland, an unabashed astronomy aficionado, is a former senior associate editor at Spectrum. She now works in the editorial department at Nature. At Spectrum, she wrote about a variety of engineering efforts, including the quest for energy-producing fusion at the National Ignition Facility and the hunt for dark matter using an ultraquiet radio receiver. In 2014, she received a Neal Award for her feature on shrinking transistors and how the semiconductor industry talks about the challenge.