no, max read performance on xbox one is 109GB/s + DDR3
Max write Performance is 109GB/s + DDR3
Max mixed performance is 204GB/s + DDR3
the esram and DDR3 can be used at the same time. also is also possible to read and write into esram at the same time. another difference is that you have the really high bandwidth for just 32MB Memory, while for the big memory you don't have that high bandwidth.
if you are only writing/reading 32MB in the whole GDDR5 Memory, you have a really low bandwidth for this chunk for the big memory, if the 32MB is not spread to all memory modules (you cannot access one module with all the bandwidth). you can only reach almost the same Performance (but not the low latency) for the render-target (if this is inside the 32MB) if you spread it through all memory modules. if you really use all the bandwidth of the esram has, you would have a problem achieving the same with gddr5. but normally you wouldn't really use all the bandwidth.
but at least this means, you can really really often read and write inside the 32MB, something you couldn't do with dram.
at the same time, if your render target is bigger than 32mb, you loose some speed, but you are still faster than having your whole rendertarget in GDDR5 memory (in theory).
the bandwidth shouldn't be a problem at all in the xbone, in current games it is more the drivers and the gpu, that are just to slow to use it.
but it would be interessting to know how much actual bandwidth the render target consumes in current games.
You're forgetting that GDDR5 operates at a MUCH higher clock rate, so in the same cycle that you run a read and write for the eSRAM, GDDR5 could have completed more than 6 read OR write cycles.
Remember
GDDR5 - 176GB/s (256 pins * 5500Mhz / 8 bits)
1 cycle 256 bits, done in 182 ns
DDR3 - 68GB/s (256 pins * 2133Mhz / 8 bits)
1 cycle 256 bits, done in 469 ns
eSRAM - 109GB/s+95GB/s =204GB/s (128 bytes * 853Mhz) (this is max theoretical, which is, quite frankly, undoable)
1 cycle 1024+896 bits done in 1,172 ns
in one cycle on the DDR3, the GDDR5 runs ~2.5 cycles
in one cycle on the eSRAM, the GDDR5 runs ~6.4 cycles
Thus, if we look at it from a "benefit of reading and writing" standpoint, there is pretty much no benefit.
So what if you can do them both together?
I'll just break the load up into two different cycles of read and write and there will be no issues!
And your example is a really odd example that aims to break the PS4 system by forcing it to do what the XB1 does, which probably no one will even try to do.
Why would you want to write to a small 32MB block so many times if you can just distribute it across the full 8GB. and write to lots of different 32MB blocks at once/ read from lots of 32MB blocks?
You're applying a 32MB limit to a system where it has no such limit, thus this mode of operation won't even be attempted.
This also brings up an industrial engineering POV, in comparing the types of memory.
From a bandwidth point of view, GDDR5 is actually favored over eSRAM, if other things (like size and bandwidth) is held constant.
Small batches with high frequency generally allows much better throughput than large batches with lower frequency.
Thus, if you have two hypothetical memory systems both with bandwidth of 100GB/s and running under similar conditions,
The one operating at 4000Mhz with 200 pins will achieve better throughput over another one operating at 1000Mhz and 800pins.
