Intel Kaby Lake 7700K vs Haswell/Broadwell-E in HFT

Recently BlackCore announced it's first server based on the new Intel Kaby Lake architecture. Targeted at specific corner case workloads requiring extremely high frequency and up to 4 Cores, this product is the first from BlackCore that utilises the latest cutting edge Intel architecture.

As an anomaly in the Intel tick-tock process (which would seem to be coming to an end), the key benefit over Skylake is Kaby Lake's ability to scale very well in terms of frequency - both on the CPU cores and the cache. This makes an interesting propostion for HFT as it can present opportunities to reduce latency on existing code optimised for prior generation x86 compatible CPUs.

Using the "small socket", which in this latest 7th generation format is LGA1151, does bring some limitations to a system. Indeed it is the first time BlackCore have forayed into this "lesser" product line from Intel. The limitations include Dual Channel memory for one, usually lower TDP limits than its big brother, and less PCIe lanes. That said this latest generation supports up to 24 lanes making it viable to have interesting PCIe configurations vs last generation supporting only 20.

It is important to understand that whilst the Kaby Lake CPU scales to 5Ghz and beyond with BlackCore cooling, the limited number of cores mean that in any multi-threaded benchmark it loses out to it's bigger brothers based around the X99 chipset and Haswell/Broadwell architecture. This is especially becuase of the 10 cores available on the X99 platform. Also whilst the Kaby Lake supports extreme memory overclocking (we are currently testing 4133Mhz memory in the labs on this system, for customer deployments in Feb) - what a lot of people forget is that this small socket only supports dual channel memory vs Quad Channel at up to 3333Mhz on the X99 Platform.

Think about it like lanes on a motorway, 4 lanes with all the cars doing 333MPH vs 2 lanes with the cars doing 413MPH for instance. If you care about ALL of the cars getting to their destination, option 1 is better, but if you only care about ONE particular car, option 2 is better. So, if transactions to memory are infrequent and carry small payloads, often dual channel could be better for your particular workload.

There is a crowning glory however, in single threaded benchmarks the Kaby Lake CPU really shines, especially if you take advantage of its AVX2 extensions and run in a BlackCore liquid cooled chassis, with the CPU at 5Ghz instead of the default 4.2Ghz, alongside increases to the cache frequency, and further BIOS/OS optimsations! Not only is the single thread performance fantastic, the overall latency in the system is also measurably reduced vs X99 under specific workloads.

Moving forward, it would be incredibly interesting if Intel would consider putting this architecture into a next generation chipset and CPU replacement for X99/6950X sometime ;)

Give us a call if you want to try one out!

Mark - CTO @ BlackCore - [email protected]