By using this site, you agree to our Privacy Policy and our Terms of Use. Close
greenmedic88 on 01 February 2014
Edit Thread
drkohler said:
Turkish said:
globalisateur said:

Real benchmarks during ideal conditions have been done on both machines:

- 140-150GB/s averaged at 145GB/s in best conditions using the best test available for Esram. Those are Microsoft ideal/best condition numbers and work for rare utilizations, most cases will reach maybe 100GB/s in fact.
- 160GB/s done with a similar test on the GDDR5 ram by some developers, they can reach 91% of the total 176GB/s which is compatible with similar, on best conditions tests on GDDR5 ram on PC, maybe like 135GB/s in less ideal conditions.

Hmm this is interesting, where were this tests done?

In the MS voodoo laboratory, as noone ever was able to explain the numbers. Let's bring some points up to clear up stuff and obfuscate clear stuff:

1. There is a poster here who bravely states "esram is only a bottleneck if it is used"

Fact: That is a funny idea because what happens if you DON't use the esram? That is prety easy to figure out: the gpu and the cpu and the dma engines all access the main ddr3 dram. Since the ddr3 has a 256bit interface, only one of the 256bit dma/memory controlers can access the main memory. This gives us 953MHz*256bit = roughly 25GByte/s bandwidth for the gpu. Good luck making a game that does more than 10fps at any resolution...

2. There is a number of 204GB/s (which would correctly be 218GB/s after the clock uptick) for esram and real measured 145GB/s bandwidth number stated by unknown testers under unknown conditions.

Fact: The first number is complete and utter bullshit. This bogus number comes from a pr/tech guy who is frequently quoted by journalists as "we can read and write at the same time into the esram". No vlsi engineer has ever come forward to explain the number and how that actually works, and unless the layout of the mmu crossbridg in the apu is considerable more complex (and I mean considerably more complex) than anything ever done before (this includes supercomputer mmus), I stand by my bullshit meter hitting the fan. The second number, 145GB/s, on the other hand, is completely believable, though. However, this number does not come from the gpu data side. It is the bandwidth reached in the mmu crossbar switch. It is achieved when the gpu accesses the esram at 109GB/s (by either reading and/or writing to the four individual 256bit memory controllers of the four separate 8MB esram blocks) and simultaneous access (parallel in the mmu crossbar) of the cpu mmu of the ddr3 main ram. This gives us 109GB/s + 40GB/s = 149GB/s data rate (or 177GB/s if the cpu could use the full bandwidth (cpu access is stated lower by ms).

Again the memory layout of the XBox One apu is extremely complicated (it can get a lot, really a lot, more complicated when the gpu accesses esram blocks and ddr3 at the same time, concurrently with a dme engine accessing esram, and cpu concurrently accessing the ddr3 ram. That would probably be the messiest access pattern I see in the XBox One, but it would get the number upto (109+68) = 177GB/s, the maximum memory thorughput possible (unless there IS novel voodoo technology inside the apu).

Good post. Should be pinned. 

No voodoo here; just numbers.