Xbox One in trouble.

[Maverick Jester]

Pretty sure the ESRAM was added to try to address the fact that the PS4 is using GDDR5, which offers much higher bandwidth.

Me too.

 

[B3ns]

Pretty sure the ESRAM was added to try to address the fact that the PS4 is using GDDR5, which offers much higher bandwidth.

+1

"Microsoft has also implemented an eSRAM buffer that mitigates the relatively low-bandwidth of DDR3.
"

 

[Garson007]

It depends on what you want to do with the data.

There's the crucial part of this entire discussion. I never refuted that. I'm saying the hardware is different and that the PS4 has more raw performance. What I would like to see is what MS first parties can come up with by using the architectural differences they have been given, things that are not possible otherwise.

How much worse is the latency of GDDR5 compared to the ESRAM?

I don't have the numbers, but it's a significant difference.

What is the latency tolerance of tasks that will be run on the GPU, considering your desired frame time?

You're still fixating on the GPU alone.

Also, if the dataset that needs to be computed by the GPU is larger than 32mb - considering especially that during this time, other tasks may require the ESRAM - you are out of luck.

It's unlikely, since 32mb covers 32*1024*1024/8=4,194,304 double precision numbers. It might happen for graphics, but is unlikely for the situations I'm envisioning.

And that is another problem - the ESRAM acts as a bottleneck.

If the dev tools, or the devs do I bad job - yes.

 

[Garson007]

I understand that part... Why is it needed for effective HSA functionality? I think that is too broad a statement.

Say you need many logical passes over computed data in real time - low latency will help there a lot. It creates the ability for computations to happen in real time, but not a lot of them (due to lack of bandwidth).

It is needed for Microsoft's iteration of the HSA solution, yes. The addition of eSRAM is seen more as an act to address the large imbalance in bandwidth. Latency was never the issue for the XBO.

This is where I disagree. I don't think it exists to address the lack of bandwidth. I think it exists to address the lack of latency. However there's no official word so I might be completely incorrect.

 

[Mike Hoxbig]

Seems like a few new electronic engineers pop up whenever there's a console announcement

Seems to be the case

 

[Maverick Jester]

Say you need many logical passes over computed data in real time - low latency will help there a lot. It creates the ability for computations to happen in real time, but not a lot of them (due to lack of bandwidth).

Which is why it is more appropriate to say that it is needed for Microsoft's iteration of the HSA solution

This is where I disagree. I don't think it exists to address the lack of bandwidth. I think it exists to address the lack of latency. However there's no official word so I might be completely incorrect.

Say what now?

 

[Garson007]

Which is why it is more appropriate to say that it is needed for Microsoft's iteration of the HSA solution

Correct. However this does not mean all HSA iterations are equal in what they perform well at. Hence the use of the keyword different, and not lesser, architecture.

Say what now?

I meant to say lack of low latency.

 

[Ancalagon]

Correct. However this does not mean all HSA iterations are equal in what they perform well at. Hence the use of the keyword different, and not lesser, architecture.

50% more shaders say its lesser.

Fact is, the PS4 can devote 768 shaders to graphics processing - ie exactly equal to everything the XBO has, except reportedly at higher clockspeeds - and have 384 shaders to do physics/AI calculations.

So, I'd say PS4 has this in the bag.

 

[Polymathic]

50% more shaders say its lesser.

Fact is, the PS4 can devote 768 shaders to graphics processing - ie exactly equal to everything the XBO has, except reportedly at higher clockspeeds - and have 384 shaders to do physics/AI calculations.

So, I'd say PS4 has this in the bag.

Which one makes my life easier and less cluttered? The Xbox One but that said I'm not exactly a "hardcore gamer"

 

[Garson007]

50% more shaders say its lesser.

Fact is, the PS4 can devote 768 shaders to graphics processing - ie exactly equal to everything the XBO has, except reportedly at higher clockspeeds - and have 384 shaders to do physics/AI calculations.

So, I'd say PS4 has this in the bag.

/me sighs.

 

[Ancalagon]

/me sighs.

Its more powerful, sorry but it is! No one else on the entire Internet thinks that the "advantage" offered by the much lower bandwidth DDR3 and a small pool of latency DDR3 is enough to offset the massively increased shader power and memory bandwidth.

How exactly is SLIGHTLY lower latency to 32mb of ESRAM going to make up for:
50% more shader power
TRIPLE the bandwidth to main memory.
?

 

[Maverick Jester]

Correct. However this does not mean all HSA iterations are equal in what they perform well at. Hence the use of the keyword different, and not lesser, architecture.

Agree with you, especially because of their different aims.

I meant to say lack of low latency.

Ah, gotcha

 

[Chevron]

Which one makes my life easier and less cluttered? The Xbox One

How? It doesn't replace any boxes. Just adds to them.

 

[Garson007]

How exactly is SLIGHTLY lower latency to 32mb of ESRAM going to make up for:
50% more shader power
TRIPLE the bandwidth to main memory.?

It won't. Making up for what is lost is not the point. It's going to lose at bandwidth heavy stuff. It will win with latency heavy stuff. It's different. It offers different possibilities. Neither offers the same stuff at the same performance level. There is a clear winner in bandwidth and compute units and there is a clear winner in latency. Some things will run better on the one than the other. It creates opportunity for first party developers to develop stuff on their platform that's just not possible on the other.

 

[Chevron]

It won't. Making up for what is lost is not the point. It's going to lose at bandwidth heavy stuff. It will win with latency heavy stuff. It's different. It offers different possibilities. Neither offers the same stuff at the same performance level. There is a clear winner in bandwidth and compute units and there is a clear winner in latency. Some things will run better on the one than the other. It creates opportunity for first party developers to develop stuff on their platform that's just not possible in the other.

Leave the straws alone. No need to grab at them. They didn't do anything to you.

 

[Garson007]

What straws would those be? That the architecture design is different?

 

[Ancalagon]

It won't. Making up for what is lost is not the point. It's going to lose at bandwidth heavy stuff. It will win with latency heavy stuff. It's different. It offers different possibilities. Neither offers the same stuff at the same performance level. There is a clear winner in bandwidth and compute units and there is a clear winner in latency. Some things will run better on the one than the other. It creates opportunity for first party developers to develop stuff on their platform that's just not possible in the other.

But it doesnt sound like you know why latency matters or how you measure how much latency is enough!

So you dont know what the latency is on the ESRAM, the DDR3 memory pool, or the GDDR5 memory pool? You dont know what the difference in latency is, and you want to tell me that the advantage that MS has in latency will mean that the architectures will have no clear winner, only differences?

No, I'm sorry, wrong.

First, how is latency calculated? Latency is calculated in the number of memory cycles it takes for a load or store to complete. It is measured in cycles, not milliseconds. Lets say the ESRAM is clocked at 1.6GHz, the same as the Jaguar CPUS which would not be unusual. And lets say latency is EXTREMELY low, say 2 cycles for a read. To find out how long that takes in milliseconds, you need to divide 1 second by 1.6Ghz and then multiply by the number of cycles.

When we look at GDDR5, lets say its latency is 20 cycles for a read - 10 times as many cycles. But its clockspeed is 5.5GHz - over 3 times that of our estimated ESRAM. Can you see where I am going here? The difference will drop from 10x to 3.somethingx.

And latency only matters if its too much. It doesnt help to have latency that is 0.0000000000s if you can tolerate up to 0.00001s. Its only if latency goes past your threshold that players will notice, because the time taken to generate a frame goes up, which the user will see as input lag.

If the GDDR5 latency was flipping terrible, we would have an issue. But we know its not terrible, and you know how we know that? Because GDDR5 is used in graphics cards, which is the most latency sensitive part of rendering. They even use GDDR5 for professional broadcasting (industry I used to work in). If the latencies are good enough for broadcast, its good enough for games.

EDIT: as a sidenote, the broadcast industry cant even use USB devices because their latency and overhead is too high.

 

[Chevron]

What straws would those be? That the architecture design is different?

The xbone is inferior to the PS4 no one is disputing that, but you. So you can stop the trolling now. The joke is over.

 

[Garson007]

But it doesnt sound like you know why latency matters or how you measure how much latency is enough!

Why does a CPU and a GPU have a cache? Latency. The CPU and GPU will be sharing these caches on HSA hardware. Increasing the size of these caches will improve the performance of low-bandwidth latency intensive tasks, which do exist - especially under any sort of GPU-CPU data sharing arrangement where memory will constantly have to be offloaded off of the registers and into the cache.

So you dont know what the latency is on the ESRAM, the DDR3 memory pool, or the GDDR5 memory pool? You dont know what the difference in latency is, and you want to tell me that the advantage that MS has in latency will mean that the architectures will have no clear winner, only differences?

The RAM latency would have to be reported by the manufacturer. So I'm not sure what either is. What I do know is that eSRAM is built for low latency and GDDR5 is built for high bandwidth.

No, I'm sorry, wrong.

I am not.

First, how is latency calculated? Latency is calculated in the number of memory cycles it takes for a load or store to complete. It is measured in cycles, not milliseconds. Lets say the ESRAM is clocked at 1.6GHz, the same as the Jaguar CPUS which would not be unusual. And lets say latency is EXTREMELY low, say 2 cycles for a read. To find out how long that takes in milliseconds, you need to divide 1 second by 1.6Ghz and then multiply by the number of cycles.

When we look at GDDR5, lets say its latency is 20 cycles for a read - 10 times as many cycles. But its clockspeed is 5.5GHz - over 3 times that of our estimated ESRAM. Can you see where I am going here? The difference will drop from 10x to 3.somethingx.

You're assuming that the memory calls will happen in bursts. I'm specifically talking about situations where the bandwidth does not matter and latency is more important. So the 3,4 or whatever magnitude of latency is huge (especially if it's even higher).

And latency only matters if its too much. It doesnt help to have latency that is 0.0000000000s if you can tolerate up to 0.00001s. Its only if latency goes past your threshold that players will notice, because the time taken to generate a frame goes up, which the user will see as input lag.

Latency will be too much for certain tasks. Specifically when those tasks are repeatedly going back and forth xxxx times during a frame.

If the GDDR5 latency was flipping terrible, we would have an issue. But we know its not terrible, and you know how we know that? Because GDDR5 is used in graphics cards, which is the most latency sensitive part of rendering. They even use GDDR5 for professional broadcasting (industry I used to work in). If the latencies are good enough for broadcast, its good enough for games.

It works for graphics and other bandwidth intensive computations. Graphics is not latency sensitive. That's a bull**** call, that is. Every frame generated does not need continuous calls back to memory to generate, so almost any latency faster than a frame is more than good enough.

 

[Polymathic]

How? It doesn't replace any boxes. Just adds to them.

Xbox One: Bluray player, media player and games console + can control my decoder (if DSTV allows it)