Intel 18-Core Haswell-EP Xeon E5 v3 Preview


Workstation Performance, Conclusion

The Xeon E5 v3 family isn't exclusively a server product, even if many of its workloads focus on that market. Intel is also hoping to spur workstation adoption (it calls this Technical Computing). In fact, the jump from 10-18 cores is a huge leap forward for so-called "Embarrassingly parallel" applications like Cinebench -- if you work in rendering, this kind of leap is unprecedented.



To put this in perspective, a single-socket 8-core Haswell-E pulls a Cinebench 15 multi-threaded score of 1312. The 3732 multi-threaded score from this dual Xeon E5-2699 v3 server shatters that by nearly 3x.  Here's a quick-take of it in action...


Intel's other published benchmarks back this up. Take a look at relative performance between the old dual-socket 12-socket and newer dual-socket 18-core systems.



Here's a different set of gains -- this time we're showing the boost from moving to a pair of 10-core Xeon E5 v3 processors from a single eight-core system (albeit still a v3 CPU).


Why the Workstation Boost Matters

There's a specific reason we want to talk about the workstation performance uptick, and it hearkens back to the days when dual-core CPUs were just being introduced. Back then, if you wanted more than one CPU core, you had to pay an enormous premium for the CPUs, motherboards, and the ECC / registered memory that was almost always required. If you actually needed four CPUs, you could kiss workstation-class features like PCI-Express or modern chipsets good-bye -- the handful of four-socket motherboards on the market were exclusively server-class products.

Then AMD launched dual-core Opterons, and suddenly users didn't have to pick between a quad-core workstation and a quad-core server -- you could have the best features of both platforms at a fraction of the initial price. By shifting to an 18-core per socket CPU, Intel is offering its users something similar. Before now, the most cores you could pack into a dual-socket motherboard on an Xeon E5 v2 system would've been 24. Now, that's increased to 36. Double that number if you count the virtual cores enabled with HyperThreading.

Granted, we're still talking about extremely expensive workstations, and in the absence of robust competition from AMD, no one is seriously suggesting that Intel's prices will drop in the near future. Nonetheless, this kind of kick matters for long-term performance improvements and the development of more advanced software and hardware.

Die Sizes, Availability: 

Xeon E5 v3 processors will be available immediately, in 4-18 core configurations. Don't look for the top-end chips to ever crack what we'd call "affordable" -- the 14-18 core chips have 5.5B transistors and measure 662mm sq, the 6-12 core processors are 3.8B transistor parts and 483mm sq, and the 4-8 core processors clock in at around 2.6B transistors and 354mm sq. With transistor counts rivaling top-end GPUs at the high end, Intel is incurring significant costs to build these parts.
 


Intel Xeon E5 v3 Server Processor Line-Up -- Click To Enlarge



Intel Xeon E5 v3 Workstation Processor Line-Up -- Click To Enlarge

Meanwhile, at 661 mm sq. the 18-core Xeon E5 v3 isn't the largest CPU Intel has ever built -- but it's close. Only the Itanium-based Tukwila which was built on a 65nm process, has been larger, at 698mm sq. With die sizes this huge, Intel's per-wafer yield is going to be low in terms of total chips per 300mm wafer -- and that's going to put a fundamental limit on just how cheap these chips can be.
 

Related content