I seriously wonder what you think about the PS2 architecture. You know, lots of data in slow RAM and some in fast EDRAM...
Xbox One (Durango) Technical hardware investigation
- Thread starter Love_In_Rio
- Start date
- Status
I seriously wonder what you think about the PS2 architecture. You know, lots of data in slow RAM and some in fast EDRAM...
Unless gDDR5 can't deliver 256 bits per access during burst mode, then 176 GB/s is a real world number. Like I said you don't need to travel 60 miles over an hour to reach a speed of 60 mph. 176 GBs is a rate. If you can sustain max bandwidth over a full second you will have moved 176 GBs worth of data. But if you only deliver 44 GB over 1/4 of a second you will still have achieved a rate of 176 GB per second.
Sustained bandwidth is another rate all together and is dictated by a whole host of other variables.
Last edited by a moderator:
Why is 'we measured' not good enough? That's actually the best way to ascertain real-world performance; actually use the design in the real world and see how it performs. 150 GB/s should mean they shifted 150 GB within the span of one second. That may have been sustained 150 GB/s for a second, or a mix for 30 GB/s and 200 GB/s over that second.
I don't understand why people are arguing so much over these BW figures. We have an ESRAM bus structure, and explanation why it isn't read+write simultaneous 100% of the time, clockspeeds and bus widths, and finally some measured performance from real-world use that seems eminently plausible, favouring the middle of the road between high and low. We can't get any more info than that unless we get hold of game performance profiles from the dev tools. Maybe with XB1 being an SDK, someone will manage to do that somehow. Until then, this topic seems kinda done to death.
I don't understand why people are arguing so much over these BW figures. We have an ESRAM bus structure, and explanation why it isn't read+write simultaneous 100% of the time, clockspeeds and bus widths, and finally some measured performance from real-world use that seems eminently plausible, favouring the middle of the road between high and low. We can't get any more info than that unless we get hold of game performance profiles from the dev tools. Maybe with XB1 being an SDK, someone will manage to do that somehow. Until then, this topic seems kinda done to death.
We are not yet comparing platforms. That conversation can start when the machines are out in 2 months.
That # 4 came from somebody other than me.
Really? So a memory system with 200GB/s peak that can only achieve 10% efficiency is better than a different memory system with 200GB/s peak that can achieve 80% efficiency when both are attached to the same GPU?
If you're going to offer hints, please try to make them relevant to the argument at hand.
Please go back and re-read my posts more carefully, I've said nothing of the sort. The closest I've come to saying that is that one possible interpretation, amongst many, of Microsofts statement if that 150GB/s is the occasionally reached real world peak as opposed to average utilization.
what you are saying is that in real life, the average BW utilized is actually less, meaning that there are even more bandwidth that are available to be used.
No, that's not what I'm saying. That's merely your interpretation of it. See how there can be different interpretations of a statement? If average bandwidth consumed is less than 150GB/s it does not necessarily mean there is "spare" bandwidth up to 150GB/s just waiting to be used. It can easily mean that a particular operation uses memory in such a way that there is simply no way to get more that (say) 130GB/s out of it. And if operations of that sort make up the bulk of your game then 130GB/s is going to be closer to your average than 150GB/s.
That doesn't mean a different kind of memory configuration couldn't achieve higher utilization in the same type of operation with the same GPU. Thus your bandwidth usage goes up, your GPU bandwidth bottleneck is eased and overall performance increases.
This can apply both to esram and GDDR5 so please lets leave platform bias out of this and focus on the facts. And the facts are that no-one hear knows the true meaning of Microsofts statement.
In simplistic form, using the eSRAM as frame buffer, a 1080p frame buffer (color+z+stencil) @ 60fps needs about 6Gbps.
Each pass with blending, so add 12Gbps for both the read/write.
So yea, I think the bw on the eSRAM will chewed up pretty quickly.
Deferred shading probably use more bandwidth even more so.
As a summary of the past year.
Microsoft stated:
Maybe not literally, but this has been their focus until E3 at least.
The only thing that has changed, is that MS overclocked the cpu by 10% and the GPU by 5 or maybe 6%. No other changes have been made.
They discovered that the ERAM was (roughly) twice as fast as previously imagined.
Aside from this, nothing has really changed, hardware wise.
So either their initial reveal was incorrect, or there is some serious back-pedalling going on.
Microsoft stated:
Maybe not literally, but this has been their focus until E3 at least.
The only thing that has changed, is that MS overclocked the cpu by 10% and the GPU by 5 or maybe 6%. No other changes have been made.
They discovered that the ERAM was (roughly) twice as fast as previously imagined.
Aside from this, nothing has really changed, hardware wise.
So either their initial reveal was incorrect, or there is some serious back-pedalling going on.
You are confused, again, on the arbitrary utilization percentage.
Sure, given algorithm A that produce the exact result as algorithm B, provided everything's the same except the difference mentioned:
1) If A does it in less time, wouldn't it make A better?
2) If A does it by using less memory, is A better?
3) If A does it by moving less data, isn't A better?
Right, so that's exactly the case, so what's the problem then? The algorithm will not use the full potential, so the bandwidth is not the bottleneck, the code is running as fast as it can, isn't that the point?
The irony is unbearable.
Last edited by a moderator:
The read to write ratio is something like 2-4:1 for normal graphics workloads. If we assume we use all of the ESRAM read bandwidth (ie >95%) and 60 % of the DDR3 bandwidth for GPU reads, we get around 45-50GB write traffic to ESRAM with a 3:1 ratio.
That's 150GB/s (100 read, 50 write) of ESRAM bandwidth with aggregate system bandwidth being well over 200GB/s. Lower average latency too.
Cheers
Last edited by a moderator:
They discovered that it was designed with a full bidirectional interface. Someone somehow forgot to tell anyone.
Yes, me.
As for the conversation, I'll post a summary that'll hopefully bring it to a natural close instead of everyone delving into cyclic arguments to nowhere. Bandwidth is a measure of data capacity over time. Whenever measured, it's an average (even a peak BW measurement is an average over time). That MS measured 150 GB/s means they achieved 150 GB/s. That's with considerable probability. There's no reason whatsoever for anyone to believe that was 150 GB/s for a fraction of a second and the average BW for ESRAM access is notably less than that; that defies the purpose of the measurement as an average and a metric you'd use to inform your development partners. That sort of argument is basically looking for an opportunity to doubt the evidence in front of is, and is unnecessary.
Johnny Awesome
Veteran
This was a very nice interview with some cool information. We'll know over the coming years which system performs better than the other, but I'm pretty satisfied with Ryse, Forza 5, and DR 3 as launch titles. They look great and this gives me hope that MS has done a much better job with the hardware than MS haters on the Internet think.
Some pointy hair bosses, marketing drones, and a bunch of internet tards suddenly discovered the aggregate bandwidth numbers.
The engineers who designed the thing didn't do it based on assumptions and luck. They examined various workloads and existing bottlenecks, designed the thing to have separate read and write buses to/from the ESRAM knowing all along what typical aggregate bandwidth could be expected for various graphics engine implementations.
As did the ones working on PS4.
MS ended up with more aggregate bandwidth and lower average latency. Sony ended up with a lot more computational grunt. That's because they looked at different workloads and designed for different price points.
Cheers
You're right, I am confused, but not about what you think.
You've missed the point again but this is so far off the initial argument that this is the last thing I'll say on it. What if there is no algorithm B? What if you simply have algorithm A running on GPU A that demands 200GB/s to match out the GPU's computation resources? And what if you have 2 memory systems (A & B) that both peak at 200GB/s theoretical but A can attain 90% utilization in algorithm A and B only 50% in that same algorithm. Clearly bandwidth utilization is higher with memory system A which is a good thing since it's closer to that the GPU needs to saturate it's computational resources.
Indeed. So let me put this question to you. Which of these statements do you actually think Microsoft is making and why:
1. We measured 150GB/s average utilization over the course of a 30 minute gaming session
2. We measured 150GB/s peak utilization over the course of a 30 minute gaming session with the average being somewhat lower
Both statements clearly have very different implications but as far as I'm aware, the only part of those statements that Microsoft have made is the bolded part. So please go ahead and explain which of these you hold to be true and what evidence you have to support that belief.
The first one. There'd be no point in having a 1024 bit write bus if it is utilized significantly less than 50% most of the time. They'd be better off with 1280 bit read and 512/768 bit write buses.
Cheers
6Gbps+12Gbps = 2.25GB/s?
This happens more often than you think: people are working on different specs and they communicate through email. Imagine if a MS engineer sent a message: but it went like this:
It's possible that the MS Hotmail servers picked the bolded part up, and marked it as spam. So nobody ever got the email, so that is why they didn't know.
That's not too farfetched. MS is a really big company, and MS Hotmail is not so good at recognising spam.
Eh, logically if you are going to put more ifs on my ifs, it's already an illogical argument.
What does this even mean? You'll just finish the computation slower?
Again, if A and B can produce the same result at the same amount of time, clearly B is far superior.
I'll save the rest since I simply can't respond to conspiracy theories that essentially look away from the evidence being presented.
It's a typo, it's Bytes, you can correct my math if it's wrong.
- Status
Similar threads
