Maxwell Architectural Discussion & Performance Analysis.

[TechFan@ic]

I've cooked up a very brief analysis of Maxwell compared to Kepler.

Maxwell employs a distinct set of improvements over Kepler.
#1 Redesigned GPU clusters, control logic for each GPU cluster (SMM vs SMX) has been increased by 50%.
Total number of CUDA cores per mm² went down by 20%.
This improves performance per clock of each CUDA core & allows for more fine grained control.

#2 Significantly larger L2 Cache (8X larger), this reduces the reliance of the GPU on using the dedicated GDDR5 and having to pass through the memory buss which both burns unnecessary energy & inflicts latency.
This improves performance significantly in bandwidth constrained environments & significantly reduces power consumption.

Performance analysis is based on data from Tom's Hardware review.
Performance efficiency is based on data from PCPer's review.

Non-bandwidth sensitive scenario.

You can clearly see that this is not a bandwidth constrained scenario, because both the R7 265 & the R7 260X are performing exactly the same, both of these chips have pretty much have the same 1.9GFLOPS of compute performance with the main difference being that the R7 265 has 80% more bandwidth enabled through its larger 256bit memory buss.

 

Balanced compute/bandwidth scenario.

Here we see typical scaling between the R7 265 & the R7 260X which indicates that this is indeed a bandwidth-limited scenario.

 

Average performance across all tested games.

 

Performance boost per clock per mm² = 4.5% in non-bandwidth constrained environments. (AC4)

Performance boost per clock per mm² = 9.7% in bandwidth constrained environments. (FC3)
This tells us that at lest half of the performance per clock improvements come from the larger L2 cache.
 

Die size per CUDA core went up exactly 20% compared to Kepler of the same die size.
Performance per CUDA core improved by 29.5%. compared to Kepler of the same core count.

Kepler allows for more CUDA cores in the same area as Maxwell,
in this case performance per mm is more appropriate to compare both architectures.

 

Total average performance improvement per clock per mm² = 7.9%.

 

NOTE: All performance metrics were obtained by dividing the frame rates of the 750 TI over the frame-rates of the 650 Ti after adjusting clock speeds of both to match.
Die size of the GTX 750 Ti is 147mm² , theoritical die size of a 650 TI can be obtained by dividing the GK104 die size which is 294mm² by two (this makes sense because GK104 has exactly twice as many functional units as the GTX 650 Ti in regards to CUDA core count, texture units & ROPs), giving us exactly 147mm² as well.

Maxwell seems to operate at higher frequencies than most Keplers, barring from the 770 & the 760.
It's not yet clear if this improvement is a direct result of the matured 28nm process or if the architecture itself has some positive influence on clock speeds.My guess is that it's likely a mix of both.

 

 

Now to move towards performance efficiency.

Average power while gaming.

According to various power measurements the 750 Ti on average draws 62-64 watts while gaming.
So we'll put it at 63 watts, this puts the 260X at 94 watts, the 265 at 104 watts, the 660 at 137 watts and the 650 ti at 74 watts.

Keep in mind that average power consumption and TDP are actually quite different, TDP is Thermal Design Power and indicates how much heat the chip generates, while power consumption is the wattage the system is drawing at the wall.
Most Kepler & AMD GCN cards don't actually reach their full TDP when gaming & only when stress testing or doing compute intensive workloads.
Maxwell seems to be the first to break this trend, running almost always at its maximum TDP level, this indicates that Nvidia actually did a lot more work towards power management with Maxwell.

So lets look at efficiency.
We'll set the GTX 660 as the standard.

Summary :
Total average performance improvement per clock per mm² = 7.9%.
Total efficiency improvement per FPS over Kepler = 42%.
Ok so what does this mean ? this actually means that if Nvidia were to rebuild all of it's 28nm Kepler GPUs without increasing the chip sizes & keeping the clock speeds constant, you will see an average gain of 7.9% in performance.

[Ryan Leech]

So the question remains, do I wait for the GTX 880 or do I buy the GTX 780 now?

[Guest Gazzer96]

So the question remains, do I wait for the GTX 880 or do I buy the GTX 780 now?

Well, you could get a GTX 780 now, wait for prices to come down then get a 2nd GTX 780.

[Faa]

So the question remains, do I wait for the GTX 880 or do I buy the GTX 780 now?

Hint: GK110 = gtx 680 O wait our 660ti outperforms the 7970 lets make that one the 680. Hopefully we won't see this again

[TechFan@ic]

So the question remains, do I wait for the GTX 880 or do I buy the GTX 780 now?

Going by the architectural improvements you can expect that if Nvidia does the same thing it did with the 600 series, i.e. introduce GM104 first, then you should expect roughly a 60% performance boost going from a 680 to an 880 and about a 30% boost going from a 780 to an 880.

 

Hint: GK110 = gtx 680 O wait our 660ti outperforms the 7970 lets make that one the 680. Hopefully we won't see this again

This actually happened due to poor yields with 28nm, not some sort of conspiracy.

And is likely to happen again with 20nm. Pretty much the same reason why AMD had to release the 290X so late.

[JupiteL]

What I've been wondering about this next gen card is the pricing. Is it going to be priced like 700 series?

[TechFan@ic]

What I've been wondering about this next gen card is the pricing. Is it going to be priced like 700 series?

If Nvidia's history is anything to go by then expect the GTX 880 to arrive at $500 and be roughly 30% faster than the 780.

Don't expect 20nm Maxwell to arrive anytime soon though, manufacturing will only begin in the summer and cards might start popping up by September,

[BrightCandle]

If Nvidia's history is anything to go by then expect the GTX 880 to arrive at $500 and be roughly 30% faster than the 780.

Don't expect 20nm Maxwell to arrive anytime soon though, manufacturing will only begin in the summer and cards might start popping up by September,

 

I know that TSMC announced that mass production volume was started on 20nm in January. Seeing as how they typically turns into real products 5-6 months later I would put it more June/July. But do you have information I dont that suggests Nvidia/AMD did not buy early wafer starts and still haven't started production? Because September would suggest their wafer starts are in March.

[TechFan@ic]

I know that TSMC announced that mass production volume was started on 20nm in January. Seeing as how they typically turns into real products 5-6 months later I would put it more June/July. But do you have information I dont that suggests Nvidia/AMD did not buy early wafer starts and still haven't started production? Because September would suggest their wafer starts are in March.

Mass production for high performance chips (GPUs) will begin in 2H of 2014, i.e July the earliest.

http://www.fudzilla.com/home/item/32578-20nm-high-performance-chips-in-2h-2014

[solarflare]

My problem with nvidia right now is. 

 

780Gtx - released at 750+$ .. as soon as amd released their cards it droped to 500 and i do have a problem with that considering it was supposed to be a flagship card for Nvidia.

 

Then they drop a 780TI which is what the 780GTX should have been for the prices they were asking for them on release. 

 

Nvidia plays to many games. 

 

At-least with AMD when you buy a flagship card its typically as good as they get for that series the only step up is typically a dualgpu  model of that card aka 7990. 

 

I prefer to know i'm getting a actual flagship card when i shell out 600+$ on something and i think the rest of us should prefer / demand that too. 

 

Its like paying top dollar to buy A Shelby GT500 mustang and then 2 months later they release a GT501 with 250 more HP and 190 more foot lb of torque for the same price. and my GT500 is now $20,000 cheaper, as someone who just made a decision to buy the BEST of something available i would be quite upset and probably never buy another from that company. 

 

I think Nvidia needs to start delivering on their products they have enough die hard fans that i think they owe them that much. 

[BrightCandle]

I think AMD set the early pricing schemes however by making the 7970 be really expensive. When Nvidia released its 680 it was priced competitively to their main competitor and pricing this generation has fallen out from those initial prices. It may also be a lot to do with the increasing cost of wafers. 28nm while offering more performance was also quite a bit more expensive to produce a chip than 40nm for the same size. 20nm will also increase in cost per wafer for similar size chips compared to 28nm. That is one part of the historical Moore's law that is breaking down, it was an economic argument that the transistors would double at the same price point, but actually its doubling at the same die size but the price of that die size is rising as the process costs increase.

 

The price of components is really variable, not least because both companies are targeting margins well in excess of 40% as is customary in the tech market. If one releases a long way ahead of the other I fully expect them to price gouge until the competition arrives, slotting the card into the existing market rather than its economic historical replacement of the lineup. I don't know why that changed but it seems to have.

[CarlosRex]

Following the conclusions shown above is it safe to assume that the GTX 880 will be faster than the 780 Ti ?

[SlaughterSmurf]

So the question remains, do I wait for the GTX 880 or do I buy the GTX 780 now?

Get EVGA 780, then use step up

[Ryan Leech]

Get EVGA 780, then use step up

I already did something else

[thenoelb]

that great moment when you need help understanding maxwell's architecture for a report and you find a post in the forums explaining specs. you da best @[member=TechFan@ic]