NVIDIA Maxwell Speculation Thread

http://forums.laptopvideo2go.com/topic/30761-inf-v4146/

Can anybody explain why GM107 is listed as DEV.13 and DEV.17?
In the past at least the first two digits of the device ID stayed the same and only the last two digits changed for the different SKUs. Maybe two different GM107 - different fabs?

 

they say second half of the year for the second generation - huh, I thought 20 nm would be available early next year...

anyways, there should be architectural improvements also present by then

 

Which press conference are the images coming from?

I'm guessing that for mobile an 128 "core" SMM is more suitable than 192 cores, now they have smaller building blocks to build mobile solutions out of.

Remember that most of the low end mobile and desktop parts in the Kepler gen. were based off of rebranded Fermi parts.

This configuration makes much more sense. And hopefully we'll see an improvement in the amount of SMEM / FPU ratio aswell!

Can't wait for a whitepaper!!

 

CUDA 6.0 RC developer driver: http://forums.laptopvideo2go.com/topic/30763-v33259-windows-8-32bit-nvidia-mobile/

NVCUDA.dll contains interessting strings:
GK180 (was supposed to be the GPU of Tesla K40, but why has it an own codename in CUDA?)
GK210 (as mentioned above a SM_37 Kepler)
SM_52 (GM2xx Maxwell compute capability?)

 

GK210 is not for this thread but this is supposed to replace GK110, right? And if so, the GM200 is still ages from being released, so is 20 nm too

 

I think 20nm is available now, it just doesn't perform very well.

 

I meant available in a state suitable for mass production and release but I guess at current pricing and yields it is still a no go option

 

Well Nvidia is making something in 20nm.

 

NVIDIA Maxwell GeForce GTX 750 Ti and GTX 750 Official Specs Confirmed

http://wccftech.com/nvidia-maxwell-...ed-60watt-gpu-geforce-800-series-arrives-2014

 

So the 660 is still in the lineup even after the 750 Ti and 750 will be released. I guess I was too "optimistic" in thinking the 750 Ti could close that gap, given the rumored performance numbers. It seems a bit strange to me why they'd have a 600 series part in the middle of a bunch of 700 series parts (the 650 I can understand since it's at the bottom).

 

That's why I believe GM206 will come quickly to replace GK106 which has reached the end of its lifetime already. Nvidia needs a more power efficient and higher performing part for the $200 market and GK106 isn't ideal anymore 2 years since it launched.

 

I agree, this is odd [I assume by "fp" you're referring to the dp units]. Earlier I had asked where the dp units were and the logical response was that these units weren't in the block diagram in the previous release; now we're sitting here with these two charts which would seem to contradict that point. I don't think I buy the aesthetic or ignorance arguments -- not with TegraK1 throwing around the 192 term with abandon and the marketing team making crop circles. A more likely explanation would be the dilution of "192" as a magic marketing term, but even that seems like a stretch. So, are the smaller maxwell chips not dp-capable?

That seems likely at first blush -- we've violently agreed that there's no real market being fulfilled, so it seems completely reasonable to bifurcate your product line. But, then, why put the GK110 slide next to your GMxx7 slide? If there's no expectation for dp in your consumer product line, why raise the issue?

Another possibility is that these two models are somehow comparable, so the GM107 is capable of dp. Would it be possible that they made the alus capable of half-rate dp? Half of 128 alus is comparable to the 64 units in GK110, and presumably half-rate logic is cheaper by area and power.

Also curious -- the wording of increased performance per alu. Had the increase in performance been compared at the SMX to SMM level, an increase in utilization would be the reasonable assumption, but at the alu level, it implies the alu is capable of more. Can you issue separate mul & add, fp32 & int32, are there currently instructions that take more than one clock cycle to issue that can now get better throughput, or is there something else? I similarly find it odd that there is one scheduler and two dispatchers per 32 alus -- why does one need two dispatchers? 16 alu-wide dispatch? One external (tmu/sfu/???) and one internal (in which case, why co-locate them in the diagram)? Or is there some kind of co-issuing being done here? [Or, are those not dispatch units?]

Lots of questions, I wonder how many answers we'll get on Tuesday....

 

And this is exactly why rebranding an old part with a more consistent name that fits into the current product line makes so much sense.

 

If the "new" information on die size (148mm^2) and transistor density (15% more dense than GK107) is to be believed here: http://videocardz.com/49557/exclusive-nvidia-maxwell-gm107-architecture-unveiled

Then that puts GM107 roughly at 12.67 million transitors per mm^2, for a total of 1.875 billion transitors. Compared to Bonaire, which has a die size of 160mm^2, Bonaire has 13 million transistors per mm^2 for a total of 2.08 billion.

Supposedly "GTX 480" like performance will make the 750 ti around 25-30% faster than the r260x, while consuming ~20 watts less on average loads. All of this is still to be seen to be realized (and believed) but IMO Maxwell is shaping up to be an even more impressive successor to Kepler than Kepler was to Fermi (on the graphics front).

 

So nobody find really strange that Nvidia is introducing a new architecture with a low end chip ? .. when i see 1st generation on the slides, all is said.

 

Probably this is technique adopted in maxwell.
A scheduler and a register file are layered, and sm becomes electric power saving in the study by optimizing it by software.
Probably it will become the hierarchy from 2 to 3 in maxwell.
In addition, various technique will be used.
Possibly the warp size may be changed.

There is the technique to divide big warp into small warp, and to handle.

A Hierarchical Thread Scheduler and Register File for Energy-Efficient Throughput Processors
https://research.nvidia.com/publica...r-file-energy-efficient-throughput-processors
http://www.cs.virginia.edu/~skadron/Papers/gebhart_tocs.pdf

 

I think I figured it out: there will be a second generation!

 

The leaked (game) benchmarks do not really suggest that - more like 5-10% faster than r260x - mind you I would consider that quite impressive for a slightly less complex chip, with large disadvantages in raw numbers just about everywhere (mem bandwidth, alus, tmus). Power consumption though could potentially be more like a 40W advantage if that 60W TDP is any indication...
Maybe that's why AMD paper launched the r7 265 now, the gtx 750ti might beat the r7 260x but won't be able to touch the 265 (and MSRP is still quite close probably hence reviews comparing them). Of course power consumption as well as perf/w should be much better but (on desktop) this is just one factor. (On the power consumption front, if it's really that impressive I would believe they are using HPM otherwise this just seems too good to be true.)

 

About the HPM: K1 is using that and reaching a 900MHz(?) clock speed. That's really not that much slower than, say, a GK110 (at least if you ignore boost.) You'd also think that they have power efficiency in mind when they make a mobile chip and thus don't go to the limit in terms of choosing the most aggressive, power hungry standard cells, so they may leave some things on the table in terms of clock speed for HPM.

So much faster is HP really compared to HPM? Just 10%?

 

GK107 came out first with Kepler, so it really isn't strange. It's a sound strategy given the need to refresh their mobile boy's in both performance and perf/watt in the face of ever-improving iGPU's.