AMD Threadripper benchmarked : Huge cores number with exceptional price performance
Threadripper 1950x offers 16 CPU cores ins for only $999. The benchmark shows that Threadripper can win Intel’s offering for much lower price. However it need the workload to be highly paralleled to gain full advantages of gigantic number of cores.
It’s no secret that this is a great year for AMD. Intel had been dominated the desktop CPU market for longer than 5 years since the debut of Sandy Bridge while AMD is struggling since Bulldozer Architecture failure. But AMD had finally made a great return with Ryzen that catch up with Intel’s Kaby Lake single core performance and completely win with higher parallel work loads with double of cores count. AMD is still in a long way to gain their market share back from Intel in High-End Lineup. Therefore, AMD had released Threadripper to win in High End Workstation market.
We live in the age where single core CPU performance can not exponentially increased like before. The path that AMD choose to built their high performance workstation CPU is high number of cores. Threadripper 1950x offered 16 Cores in one CPU dies in the price of $999. The CPU run at 3.4 GHz nominal clock and 4.0 GHz boosts clock with TDP of 180 Watts. Another big features of Threadripper is it enormous 60 PCIe lanes which can greatly benefit in multi GPU setup. From the overall specification, Threadripper should be the dream processor for both production tasks and gaming enthusiast.
Benchmark Results show Threadripper’s big win in Parallel workload
The benchmark of Threadripper shows the power of 16 cores in many applications. For the fully paralleled workload, Threadripper 1950X had advantage over Intel Skylake-X i9 7900X. An example of Blender rendering show that Threadripper is 20% faster , 36% faster for AES Encoding, 7% faster in Cinebench multicore. Threadripper also had better overall power consumption for parallel workload with higher number of cores running at lower which result in better efficiency.
There are tasks where Intel still doing better like Compiling software, web related application and gaming. In comparison, Intel still had an edge in Single core performance and implementation of instruction set for specific application. Skylake-X had AVX2 instruction with can greatly enhance performance for optimized application and many software compiler still favor for Intel.
Not a win for Gamer and the software that is not optimized for NUMA
Threadripper is not the best CPU for gamer. Modern game engine still had a hard time to utilize more than 8 cores of processor. In general, Threadripper 1950X in out of the box mode is slower than Ryzen 1800X because of suffered higher memory latency and inter-core communication that can overrun the performance. However, AMD engineer foresee this problem so Threadripper is come with Game Mode which can be used to improve Gaming Performance. It works by disabling simultaneous multi-threading or SMT (work similar to intel’s Hyperthreading) and memory access mode to Non-Uniform Memory Access. Disabling SMT make the game that is not optimized for parallel work load run better as number of threads in the hardware is reduced. The Non-Uniform Memory Access (or NUMA) is more tricky to understand.
To understand NUMA in general, we need to understand that Threadripper is made of 2 dies of CCX, each with their own memory controller which connected to different module of memories. Two CCXs are connected together with Infinity Fabric so that they can see the memory of the other CCXs with the cost of different memory access latency. The memory on CCX’s own controller is called near memory and the memory on the other CCX controller is called far memory.
Since two of these memory had difference access latency, it can impact the performance of the applications. To get the best of performance, the application and OS must know which memory is near or far, so that it can treat memory in most efficient ways. To do so need a lot of behavior consideration and low level coding which many games and today application still can not afford. The workaround that Threadripper offer is to have all the memory accessible by all core without awareness of memory. This is the default configuration for Threadripper called “Creator Mode”. However, the user can choose to enable NUMA that will help reduce memory and core-to-core latency and also disable SMT in favor of lower total memory bandwidth and less thread count.
With Game Mode enabled, Threadripper show the improved performance of around in every game tested. Overall Threadripper deliver performance better than Ryzen but still lose to Intel Broadwell-E and Skylake-X.
Should I upgrade my Ryzen 7 or i7 to Threadripper ?
Yes and No. Its depend on kind of workloads that you normally handle. Threadripper is not suitable for gaming application where its still hard to utilize more than 8 CPU cores, let alone 16 cores and 32 threads. And since the base clock for is not higher than Ryzen or i7, the gaming performance is a little worse. This is also true for the tasks that is hard to maximize the parallelization like software compilation or image processing.
Threadrippper, however is suitable for the workloads that maximize the usage of threads. Tasks like video encoding, rendering, scientific simulation or encryption can greatly benefits from 16 cores palatalization. Its kind of amazing to think that now we can have 16 cores of CPU to our home computer at the relatively affordable price.