Friday, January 17th 2014

NVIDIA Readies GeForce GTX 750 Ti Based on "Maxwell"

NVIDIA's next-generation GPU architecture, codenamed "Maxwell," will debut this February, with the unexpectedly positioned GeForce GTX 750. The card will launch on February 18, to be specific. Maxwell will introduce a host of new features for NVIDIA, beginning with Unified Virtual Memory. The feature lets the GPU and CPU share the same memory. Such a feature is already implemented on the current CUDA, but Maxwell could be designed to reduce overhead involved in getting the thing to work. The next big feature is that Maxwell GPUs will embed a 64-bit ARM CPU core based on NVIDIA's "Project Denver." This CPU core will allow the GPU to reduce dependency on the system's main processor in certain GPGPU scenarios. Pole-vaulting the CPU's authority in certain scenarios could work to improve performance

Getting back to the GeForce GTX 750 Ti, NVIDIA's aim is simple, to see how "Maxwell" performs on the existing, proven 28 nanometer silicon fab process, before scaling it up on the future 20 nm nodes, with bigger chips. Given its name, we expect it to be positioned in between the GTX 760 and the GTX 660 in terms of gaming performance, but we won't be surprised if it falls into an entirely different league with GPGPU. There are no specifications at hand.

Source: SweClockers
Add your own comment

44 Comments on NVIDIA Readies GeForce GTX 750 Ti Based on "Maxwell"

#1
acekombatkiwi1
I went with 4x R9 270s last week because they only pull 138w each and give me 445KH/s.

Posted on Reply
#2
Death Star
In the world of GPGPU, having access to a CPU through a very low-latency channel is a very appealing prospect. First off, this eliminates some of the normally required PCIe transfers, which are slow and costly. That's obviously always a good thing. CPU's are very good at general purpose computations. Unlike GPUs, CPUs are very good at making decisions. If you can offload a decent portion of the decision making to the CPU, that's a few percent of extra performance on *basic* GPGPU kernels while potentially much more on sophisticated kernels.

About a month ago I restructured my finite-difference time-domain code by moving as many decision-making tasks (mostly if-statements) CPU-side as possible. In doing so I removed about 15 if-statements, but added 10-50 extra arithmetic operations per if-statement that I removed. The end result was a 4% performance increase, which given the simplicity of the kernels to begin with is pretty damn good. It will be interesting to see how the addition of the ARM on-board affects all of this.
Posted on Reply
#3
TheGuruStud
Translation: 20nm is so bad that TSMC told us to go F ourselves for production. We're desperately trying to get yields up, but every wafer super, extra, mega sucks. And we gotta paper launch a card (maxwell in general) really far ahead this time LOLZ.
Posted on Reply
#4
NC37
Man invents CPUs
Man invents GPUs
CPUs become GPUs
GPUs become CPUs
Computers become self aware!!

Man invents Chobits
Posted on Reply
#5
theoneandonlymrk
by: Death Star
In the world of GPGPU, having access to a CPU through a very low-latency channel is a very appealing prospect. First off, this eliminates some of the normally required PCIe transfers, which are slow and costly. That's obviously always a good thing. CPU's are very good at general purpose computations. Unlike GPUs, CPUs are very good at making decisions. If you can offload a decent portion of the decision making to the CPU, that's a few percent of extra performance on *basic* GPGPU kernels while potentially much more on sophisticated kernels.

About a month ago I restructured my finite-difference time-domain code by moving as many decision-making tasks (mostly if-statements) CPU-side as possible. In doing so I removed about 15 if-statements, but added 10-50 extra arithmetic operations per if-statement that I removed. The end result was a 4% performance increase, which given the simplicity of the kernels to begin with is pretty damn good. It will be interesting to see how the addition of the ARM on-board affects all of this.
So 4% eh , add 10-20% and this poorly named card is not getting anywhere near a 7870 in compute never mind a 280x.
It should however play games really well and with nvidias recent cross licence with intel they might have something brewing but then again intel did keep x86 off the licence so it may yet be tricky especially since intel seam to have access to nvidias patents and much more leverage in foundry tech then nvidia.
This is panic status releasing by nvidia , , like I said ages ago maxwell (big proper one with 1tb bandwidth nested memory) is 2014 q4 earliest likely 1h 2015 on 20nm unless they have fully ditched tsv attached memory plans which would make for a lot of Bs in nvidias earlier pr slides
Posted on Reply
#6
Death Star
by: theoneandonlymrk
So 4% eh , add 10-20% and this poorly named card is not getting anywhere near a 7870 in compute never mind a 280x.
The key to the 4% was just that it was gained by getting rid of a few if-statements. There are plenty of other operations that can be used in more sophisticated kernels that would gleam higher performance gains with offloading to a CPU.

It will be interesting to see how much it actually helps with more complicated GPGPU kernels, but yeah I doubt it will approach anything like a 7870 or 280x in the vast majority of circumstances. Seems like more of a niche card.
Posted on Reply
#7
JDG1980
by: acekombatkiwi1
I went with 4x R9 270s last week because they only pull 138w each and give me 445KH/s.
What settings did you use for that? My 7870 (with the same number of SPs) tops out at about 350 KH/sec.
Posted on Reply
#8
Crap Daddy
Funny how this thread went from maxwell to mining.
Posted on Reply
#9
james888
by: Crap Daddy
Funny how this thread went from maxwell to mining.
I would rather it be about maxwell as that is far more interesting.
Posted on Reply
#10
theoneandonlymrk
by: james888
I would rather it be about maxwell as that is far more interesting.
Indeed
Posted on Reply
#11
james888
I've known about maxwell for quite a long while now and what it brings to the table. Anyone know what amd's R300 series cards are going to bring?
Posted on Reply
#13
NeoXF
by: 15th Warlock
The days when a Titan mined 350~400 KH/s and a 780ti mined 400+ KH/s are long in the past, with Cudaminer my Titans easily mine 512~535 KH/s each:



The GT430 is used to drive my display while the Titans mine :)

And yes, I know this is still a far cry from my 290X' performance (900+ KH/s) at a much lower price, but with each consecutive release cudaminer increases the mining efficiency of CUDA based cards, so don't dismiss Maxwell so easily, chances are Nvidia is going to catch up to AMD when it comes to mining efficiency, and Maxwell could well be the right architecture to do so :)
So we'll have mining-induced GPU price inflation on top of nVidia's typical one? Nice! :P
Posted on Reply
#14
Frick
Fishfaced Nincompoop
How do you mine inductance? Mining is so confusing. :(
Posted on Reply
#15
15th Warlock
by: NeoXF
So we'll have mining-induced GPU price inflation on top of nVidia's typical one? Nice! :p
You know it's coming... Newegg is already inflating prices of AMD cards, and hoarding their most sought after cards, artificially restricting supply so they can create little gems like this beauty here:


Mining Kit UKO-K140K: AMD Sempron 2.8GHz Single Core, 6 X Radeon R9 290X 4GB, 4GB RAM, 2 X 1500W 80 PLUS Gold Power Supply
All this can be yours for only: $4,119.99!!

It's only a matter of time before Nvidia steps into the ring, and then it's off to the races so to speak, dunno yet what part Maxwell will play in this game... :(

I'm all for gaming and computing inspired innovation, and I really hope Nvidia is focusing on that with Maxwell, but the Pandora box has been opened, and fact is there's a big market in the mining business right now, and sooner than later we'll have two contenders in this arena.

I'm looking forward particularly to what Maxwell will bring in terms of power
efficiency, not very exited about the desktop parts, but mostly about the mobile parts!
Posted on Reply
#17
BiggieShady
by: 15th Warlock
You know it's coming... Newegg is already inflating prices of AMD cards, and hoarding their most sought after cards, artificially restricting supply so they can create little gems like this beauty here:


Mining Kit UKO-K140K: AMD Sempron 2.8GHz Single Core, 6 X Radeon R9 290X 4GB, 4GB RAM, 2 X 1500W 80 PLUS Gold Power Supply
All this can be yours for only: $4,119.99!!

It's only a matter of time before Nvidia steps into the ring, and then it's off to the races so to speak, dunno yet what part Maxwell will play in this game... :(

I'm all for gaming and computing inspiring innovation, and I really hope Nvidia is focusing on that with Maxwell, but the Pandora box has been opened, and fact is there's a big market in the mining business right now, and sooner than later we'll have two contenders in this arena.

I'm looking forward particularly to what Maxwell will bring in terms of power
efficiency, not very exited about the desktop parts, but mostly about the mobile parts!
Those bastards, they did it ... save $250 LOL
At least they could throw in 6 PCIE risers and a piece of a PC gamer's soul :cry:
Posted on Reply
#18
xorbe
by: Xzibit
Benchmark leaks from Coolaler.com
If true, that's a big gap from 750 Ti to 760 yeah?
Posted on Reply
#19
theoneandonlymrk
Is this looking like the other side of the tegra k1 binning tree to anyone else, ???.
Like poor core bin but max gpu array bin
Posted on Reply
Add your own comment