The main reason for adding standalone processors to the radios in Broadcom's XStream architecture is to offload all wireless tasks from the main CPU. All the central BCM4709 processor has to do to support the radios is to forward Ethernet packets to and from them. The main use of the newfound central processing power will be to improve storage performance, according to Broadcom.
Broadcom XStream Block Diagram
Image credit: Broadcom
NETGEAR encouraged reviewers to put the R8000 to a Maximum Performance / Stress test. They even provided enough R7000s to use as bridges to ensure maximum performance. Although I had enough R7000s, I didn't have the six computers NETGEAR suggested in its test diagram. So the modified version pictured below has a deleted PC3. I found that a single computer connected to the R8000 was plenty powerful to stream traffic to both one 5 GHz radio and the 2.4 GHz.
Stress Test Setup
I first ran IxChariot throughput.scr scripts using TCP/IP and test file sizes of 9,000,000 Bytes for 5 GHz and 5,000,000 Bytes for 2.4 GHz. Each radio was hit with a single IxChariot stream. There was no wired routing traffic in this test.
The screenshot below is a composite of each radio run by itself to establish a baseline. Total throughput here is 1096.15 Mbps. This result would have been higher if I could subtract out the low throughput at the start of the 2.4 and 5 GHz hi runs caused by an IxChariot quirk.
NETGEAR R8000 Wireless Stress Test - Individual run composite
The next screenshot shows all three streams run simultaneously. Total throughput is about the same, as is the 2.4 GHz radio's. But the hi-band 5 GHz radio seems to have gained throughput at the expense of the low-band 5 GHz.
You have to eyeball the two traces, but the average hi-band radio trace above is around 650 Mbps, while it jumps to around 750 Mbps in the simultaneous run below. The low band 5 GHz radio in the plot above averages around 425 Mbps in the plot above and drops to just under 300 Mbps in the simultaneous run below. I was able to reproduce these results on multiple runs and when swapping shared computers around.