Thursday, February 6th 2014

NVIDIA GM107 "Maxwell" Silicon Pictured

Here is the first picture of a couple of NVIDIA GM107 silicons in a tray, ahead of graphics card assembly. The packages appear to be as big as those of the GK106 from the previous generation, however, the die itself is estimated to be smaller, at roughly 156 mm², compared to the 221 mm² die of the GK106, and the 118 mm² of the GK107. The best part? All three chips are built on the same 28 nm silicon fab process. So what makes the GM107 die smaller than that of the GK106 despite having a similar feature-set? Narrower memory bus. The GM107 is said to feature a 128-bit wide GDDR5 memory interface, in comparison to the 192-bit wide interface of the GK106.

Apart from the 128-bit wide GDDR5 memory interface, the GM107 is said to feature a total of 960 CUDA cores, 80 TMUs, and 16 ROPs. The CUDA core count is identical to that of the GK106. The GM107 is built on NVIDIA's next-generation "Maxwell" GPU architecture. It will form the foundation of two SKUs, the GeForce GTX 750 Ti, and the GeForce GTX 750. The former features the full complement of 960 CUDA cores; while the latter is slightly cut down, and features just 768. The TDP of the GTX 750 Ti is approximated to be around 75 Watt. If true, the GTX 750 duo will set new standards on the performance-per-Watt metrics. NVIDIA is expected to launch both, later this month.


Source: VideoCardz
Add your own comment

29 Comments on NVIDIA GM107 "Maxwell" Silicon Pictured

#1
Xzibit
by: sergionography
Well remember this is Maxwell and not Kepler, and nvidia stated Maxwell is designed specifically for mobile and efficiency. If you look at the big picture nvidia started with Fermi all about compute but then back pedaled with Kepler and went all about efficiency and mobile. So each compute unit now has less compute resources and is geared more towards graphics unlike amd where in sea islands they pretty much only improved compute and did almost nothing to the graphics other than some Fine tuning for efficiency. So what do we have now? Bonaire and this gk107 both measuring around 160mm2 but with nvidia packing more cores on the same process. And with nvidia being about 20% faster than gcn per core for graphics intensive tasks, but then being Much behind in compute. It's obvious this is a direct competitor to Bonaire and performing about the same as GTX650boost but closer to a GTX660 when bandwidth is not as needed all with a smaller die meaning better efficiency. And to those who wonder why nvidia would release such a part that performs similar to the ones before? Because nvidia was competing with amds 160mm2 Bonaire with a 220mm2 go106 chip that had a few parts disabled which I bet still cost more.
You have to remember and take into account that the results are all on OC 750 Ti models and still not close to a 650 Ti Boost.

650 TI Boost = 768 / 64 / 24
750 Ti = 960 / 80 / 16
R7 260X = 896 / 56 / 16

R7 260X @ $119
R7 260X OC @ $139
GTX 750 @ $119
GTX 750 Ti @ $139-$149

It might be a competitor to the R7 260X after 5months of its release.
Posted on Reply
#2
HumanSmoke
by: sergionography
Well remember this is Maxwell and not Kepler, and nvidia stated Maxwell is designed specifically for mobile and efficiency. If you look at the big picture nvidia started with Fermi all about compute but then back pedaled with Kepler and went all about efficiency and mobile.
I wouldn't confuse an architecture with just the desktop variants of their implementation
by: sergionography
So each compute unit now has less compute resources and is geared more towards graphics unlike amd where in sea islands they pretty much only improved compute and did almost nothing to the graphics other than some Fine tuning for efficiency.
Likewise, Nvidia have prioritized floating point calc over integer since the G80, since the pro markets are geared more towards FP optimization. It also the reason that AMD's architectures excel at hashing (primarily an integer function that benefits greatly from AMD's integer shift implementation, and likely the compute unit-to-core ratio AMD has instituted- something that won't change if GM107 moves from 192 to 256 cores per SMX. Compute functionality covers a multitude of variables, and depending on workload and coding optimization can yield a varied array of "wins" and "losses" - further complicated by the fact that Nvidia deliberately cripples features in desktop SKUs to protect a lucrative pro market.
Posted on Reply
#3
Casecutter
by: sergionography
Well remember this is Maxwell and not Kepler, and nvidia stated Maxwell is designed specifically for mobile and efficiency. .
But are we hearing that there's a big change in the basic Cuda functions?

by: sergionography
Because nvidia was competing with amds 160mm2 Bonaire with a 220mm2 go106 chip that had a few parts disabled which I bet still cost more.
I believe you mean wasn't... I think the 192-Bit bus was overkill and draws to much for what it brought in performance. Getting off 192-Bit saves on the die while improvement the 128-Bit provides plenty of useable Bandwidth while more efficient.
Posted on Reply
#4
HumanSmoke
by: Xzibit
Nvidia could just paper launch a reference card that doesn't need a 6pin and let the partners add a 6pin. Nvidia can say it doesn't need one but the partners added.
That was what I was inferring earlier, although I'd say that the no external power will be the norm in most cases. I'm sure the overclocked versions will gain more publicity, but that is always the case on enthusiast sites, but the bread an butter cards will likely not need external power. I'd also think that a reference cooler might only be for OEM use- I'd expect most AIBs to have their own implementation.
Reference PCB shows option for PCI-E 6-pin, but not used in this case

by: Casecutter
But still being 960 Cuda part I can't see some 50% improvement on efficiency, all while higher clocks… on 20Nm perhaps. If they can find a 20% improvement for a 960 Cuda part they'll be doing great. While the 768 Cuda part on GK106 was 110W
I thought it had been established that the fully enabled part was a 640 core part, and the non-Ti 750 was to use a 512 core (1 SMX disabled) die
Posted on Reply