4 ms·
The RK3399 doesn't come anywhere close to this. The BPi board looks to be 24 cores on a single soc (or at least running a single system image). The closest you
by subway 8y ago
The RK3399 doesn't come anywhere close to this. The BPi board looks to be 24 cores on a single soc (or at least running a single system image). The closest you could get is to hang 10gig ethernet off of the pcie bus on 4+ RK3399s. Then you also have to start thinking about fun things like cluster scheduling, and the overhead of running an OS on each device.
- contingencies 8y agoYes, these were carrier boards with numerous (8-16?) sub-boards each of which had an RK3399 CPU, each of which has 6 CPU cores (2xCortex-A72 cores, 4xCortex-A53 cores) plus a NEON coprocessor and Mali T860 MP4 GPU. The article discusses 24 cores. The Firefly hardware I saw would therefore have at least 6x8 = 48 CPU cores, or double the ARM64 CPU cores discussed in the article, before even counting the coprocessor or the 8-16 GPUs also included.
- subway 8y agoDepending on workload, pure core count can be irrelevant when your interconnect is trash and you have no shared memory. You might see OK performance if the interconnect somehow supports RDMA, but that seems exceedingly unlikely. Almost all these cluster boards are using gigabit ethernet interconnect. Not to mention you spend a core or more per SoC to handle your os and interconnect/network interrupts. NEON is implemented on a per-core basis, not as a separate coprocessor. Every A53 (and your handful of A72s) supports NEON. I'll give you the GPUs... if you really want to put in all the effort for that sweet sweet OpenCL 1.2 action on an out of date vendor kernel. An RK3399 cluster, and a single many-core system is a nonsensical apples to oranges comparison.
- floatboth 8y agoBut RK3399 has much better single core performance with the A72 cores (that you can run at 2.2ghz), so if you're not running parallelizable tasks all the time (/not looking for a cluster) it's a better chip.