Xbox One (Durango) Technical hardware investigation

That's all true, but it is a meaningless comparison if you don't know how it impacts the final performance (aka framerate).

Let me exaggerate just to make a point: Imagine that a 680 spent 85% of the frame time idle due memory latency and had only 10% for ALU work, and the remaining 5% would be anything else. In that case, while it's true that durango's gpu would never match the flop performance/output of a 680, it turns out that ALU is simply not a huge part of the frame time, so you can actually have a gpu with less ALUs but faster memory that outperforms a 680 in that frame.

Of course, in reality it's not a drastic scenario like this, but Ms has data that shows for all current games where they spent most of their time, and instead of brute forcing everything to improve performance, they seem to have designed a system that tackles those bottlenecks. It's not that far fetch to assume that for running those games that setup could perform better.

But that's not guarantee that they will always perform on par either, because games from 5-6 years from now can be drastically different from current ones and be more suited to another architecture...

It kinda happened this gen when developers started using deferred renders, which nullified the multisample advantage that xenos had in the beginning of the generation, and accentuated its shortcomings.

 

7770 has 10 CU's.

 

The flops rating matches up. Durango's GPU has more CUs (12 vs. 10) that are clocked lower, more TMUs (48 vs 40), more memory bandwidth (2 pools with 102GB/s & 68 GB/s of bandwidth that's shared with the rest of the system vs. 72 GB/s) and equivalent ROPS.

 

VGLeaks IIRC. Devs were being encouraged to hit the metal on Orbis but had to go through the APIs on Durango, leading to speculation that Durango is supporting a device-independent software platform and might present an upgradeable console (new revision in much shorter time-frame than 6 year console cycle).

 

I thought the 360 had the same requirement (a lower-level, but DX-style API which games needed to use or they would not be certified for the console)?

 

As others have pointed out, nvidia already does have some low latency memories on it's design to improve performance.

Why they don't go all in with a design that is more efficient? I can think of two reasons:

- On pc game space the safest way you can guarantee good performance is brute forcing everything. Specially if the changes in architecture means that the api and code has to support those features either. A change this drastic on the pc space could make all games that were not designed for it run like ass.
- IIRC they currently use the same hardware for their consumer level and quadro cards now, the difference being the driver. Maybe the savings they have by designing and mass manufacturing a single product out weights the gains of having a product that while is more power efficient has. Due to lower demand it would be possible that secondary design actually costs more to make.

 

Probably because they can achieve good speed up via cheaper methods. Graphics problems are largely embarrassingly parallelizable, and latencies are already hidden/optimized in existing software.

It's also easier to market big numbers and an improved visual as opposed to a more efficient GPGPU design.

 

Thx for the explanation.

I just looked at that note and didn't know what to think.

Greetings!

 

No, they wouldn't. If nvidia came to you and said "I have this new chip, it's twice the size of the old one, costs twice as much, and doesn't improve your current games _at all_, but future games might work a lot better for lower power draw" would you buy it? I wouldn't. Would game companies invest the effort to use the esram? Probably not, unless all the gpus started having it.

It's one of those cases where doing something in a console is a lot more feasible than doing it for a general purpose pc. And esram _is_ starting to be used in supercomputer applications. I predict within a couple of years there will be mainstream GPUs with esram.

 

Which is what I was banging on about but know one seemed to amswer

Edit..that being the case, although comparatively less powerfull compared to even high emd current pc parts.. durango might actually end up future proofed from the get go for next gen game engines...and perform on par with gtx 680 in dx11 type games...taking into consideration some form of hsa, and low latency esram...all coded for bespoke with much lower api over head.

Of course that all rests on whether we do indeed see such a scenario with esram...but certainly the above does seem in the realms of possibility

 

The HSA tries to reduce latency by reducing the need for copying data. Having two RAM pools seems to clash with this design philosophy, doesn't it?

 

I'm not sure what the final spec actually says (not certain that exists yet), but HSA promises at least one coherent unified memory space, which Durango appears to have.
It has a separate pool memory whose current positioning relative to the main memory space is not clear enough to know how much it may run counter to HSA.

The other component to HSA is the run-time and implementation-specific libraries and resolver, which may allow implementations to do their own thing internally, so long as they can still work with primary memory.

 

The latency referred to in the HSA case isn't the same thing, it's the latency between submitting "compute" jobs to the GPU and being able to read/process the result. Since the CPU and GPU can both read both pools it's no different than on a machine with a single pool.

When talking about the eSRAM the latency referred to is the actual time it takes to read or write a single element from the memory.

 

Vgleaks now appears to be a skeleton site showing the / directory on loadup. Wonder if they're dieing or transitioning? whois says they're France based.

 

Scaning around the internet it seems that 64MBit SRAM chips have been produced on TSMCs 28nm process with a cell size of 0.127mm2/Mbit.

For standard 6T-SRAM that's 32.5mm² for 32MB. And for 1T-SRAM it would be 6mm².

Considering the GPU and CPU selection, 6T-SRAM is certainly possible on a 28nm process.

 

If it's 1T-SRAM, seemed like they could have packed in 256mb at that size.

 

Damn, i want my kinect 2 leak XD

I though they were following the pool. By now we should have had the kinect leak yet XD

 

Current gpus could definitely be better optimized.

Just looking at the difference between power consumption during games vs furmark shows that there is quite a big part of the gpu idle during games.
Other prof is the higher power consuption during better optimized games like battlefield 3 and crysis 2.

If they manage to optimize this 12 cus it will definitely not be energy free, it will be about the same as runing furmark on a hd 7770