Adjusting their GeForce GTX 1050 Family info page over the weekend, NVIDIA quietly announced that the expansion of the GeForce 10-series of cards with another entry: the GeForce GTX 1050 3GB. Slotting between the GeForce GTX 1050 Ti and original GeForce GTX 1050 (2GB), the GTX 1050 3GB is, despite the name, not actually equal to the original GTX 1050 2GB. Instead it features 768 CUDA cores but on a smaller 96-bit memory bus.

If this story sounds familiar, it’s because it is. After launching the GTX 1060 in July 2016, NVIDIA followed up with the GTX 1060 3GB, which featured 1152 CUDA cores to the GTX 1060 6GB’s 1280.

NVIDIA GPU Specification Comparison
  GTX 1060 3GB GTX 1050 Ti GTX 1050 3GB GTX 1050 (2GB) GT 1030 GDDR5
CUDA Cores 1152 768 768 640 384
Texture Units 72 48 48 40 24
ROPs 48 32 24? 32 8
Core Clock 1506MHz 1290MHz 1392MHz 1354MHz 1227MHz
Boost Clock 1709MHz 1392MHz 1518MHz 1455MHz 1468MHz
Memory Clock 8Gbps GDDR5 7Gbps GDDR5 7Gbps GDDR5 7Gbps GDDR5 6Gbps GDDR5
Memory Bus Width 192-bit 128-bit 96-bit 128-bit 64-bit
VRAM 3GB 4GB 3GB 2GB 2GB
FP64 1/32 1/32 1/32 1/32 1/32
TDP 120W 75W 75W 75W 30W
GPU GP106 GP107 GP107 GP107 GP108
Transistor Count 4.4B 3.3B 3.3B 3.3B 1.02B
Manufacturing Process TSMC 16nm Samsung 14nm Samsung 14nm Samsung 14nm TSMC 16nm
Launch Date 08/18/2016 10/25/2016 5/2018 10/25/2016 5/2017
Launch Price $199 $139 - $109 $80

Going down to the silicon, the GeForce GTX 1050 3GB does indeed feature 3GB of VRAM, an extra GB over the original GTX 1050, which in turn only ever came in a 2GB version in the desktop space. But ultimately, the GTX 1050 3GB is another GPU configuration: the card has an extra enabled SM – now matching the GTX 1050 Ti at 6 SMs/768 CUDA cores – while also equipped with a cut-down 96-bit bus as opposed to the 128-bit bus of its other desktop GTX 1050 series siblings. In actual memory bandwidth numbers, this translates to 84 GB/s as opposed to 112 GB/s.

Otherwise, the raster engine and TMU counts are presumably identical to the GTX 1050 Ti, as the SM count requires the frontend to be a fully enabled GP107 part. As for clockspeeds, the GTX 1050 3GB features faster base and boost than both GTX 1050 Ti and GTX 1050 2GB.

Things get a bit trickier on the backend though. While NVIDIA's published specifications don't offer clarity on this, the company's designs normally have a 1:1 ratio between ROP partitions and memory controllers. Meaning that as this configuration has a 96-bit memory bus – and thus only 3 out of 4 memory controllers are enabled – then it has almost certainly lost a ROP partition as well. This also means that a quarter of the GPU's L2 cache (256KB) has also been shut off, leaving the chip with just 768KB of L2.

This is a prime example of why unusual memory configurations are more important than they may seem at first, as several parts of the GPU are strongly bound to the size of the memory bus. It also means that estimating the card's performance based on just clockspeeds and functional block counts is harder than it may seem, as now one needs to take into account memory bandwidth, ROP throughput, and cache hits.


Unofficial GTX 1050 3GB Block Diagram

All told, the existence of this card is not entirely unexpected, but it is odd. That NVIDIA is going with a 3 memory channel configuration stongly implies that the company has built up a stockpile of GP107 parts with one bad ROP/L2/MC partition, which is not unusual but GP107 is also a small enough die that the need to salvage to this level isn't quite as great. Otherwise, NVIDIA could have produced a 3GB card with a 128-bit memory bus and unbalanced memory configuration (2x1GB + 2x0.5GB), which brings its own tradeoffs but is generally the more straightforward option. At any rate, it's rare to see a salvage part like this enter the retail market this late in the cycle.

In total, NVIDIA cited the performance difference to PCWorld as averaging out to around 10% uplift of the GTX 1050 3GB over the 2GB. And so once again, we are now looking at multiple GPU configurations being sold at retail under a GTX 10x0 banner.

Contextually speaking, this new SKU comes in at a time of cryptocurrency mining demand, which has inflated prices and reduced availabilities for video cards all around. In that sense, where there may not have been room for an in-between GTX 1050 SKU, there is now. Or given diminishing cryptodemand, silicon that could be re-configured for consumer GeForce use. And in terms of competition, the GTX 1050 series match up with the AMD Radeon RX 560 (16 and 14 CU variants), a nominally single SKU which amusingly has had its own share of naming troubles. Lastly, recent NVIDIA GeForce launch cadence would suggest that the post GeForce GTX 10-series are on the horizon, a situation where Pascal-based lower-end product launches are easily compatible with.

And for NVIDIA’s part, a quick look at their own Linux driver documentation of supported GPUs by PCI ID shows the amount of variants and configurations for GP107 and GP106, with some device IDs implying GP104-based GTX 1060 cards. Though not all of the PCI IDs may reflect a currently-selling product, some might be regional or simply unpublicized; for the rumored GTX 1060 5GB, it had already found itself on NVIDIA’s driver support lists earlier this year. So the appearance of another configuration isn’t necessarily out of the blue. For the GTX 1050 3GB today, it may not be based on GP107, and given that most manufacturers rarely, if ever, comment on the lowest performing parts in their stack, future clarification is unlikely.

But while strange GPU configurations at the lower-half of performance brackets are not uncommon, as is the dearth of clear specification documentation, the obfuscating nomenclature is less palatable. As mentioned at the time of the GTX 1060 3GB launch, NVIDIA hadn’t had multiple GPU configurations selling under a single retail GTX model number for some years now, but now they’ve done it again – excluding the recent naming shenanigans with entry-level GT 1030 GDDR5/DDR4 as well as the MX150 variants. Keeping the different GPU configurations separate to the memory configuration like with the GTX 1060 is still unintuitive or misleading for most consumers, who may not research entry-level components or be unaware of how GPUs are configured. It appears that NVIDIA has no intention of changing from its ongoing nomenclature philosophy.

Going forward, however, this approach hardly educates the consumer. For one, all previous references to the “GTX 1050” would need to be updated; on the GTX 1050 Family page, the performance graphs still refer to the “GTX 1050,” something that undiscerning readers would not pick up on.

As noted to PCWorld in response to availability and pricing, the GeForce GTX 1050 3GB will be in the same price bracket as the GTX 1050 Ti and GTX 1050 2GB.

Source: NVIDIA

POST A COMMENT

49 Comments

View All Comments

  • bji - Tuesday, May 22, 2018 - link

    Well, this is definitely the age of drama. Everyone with any kind of complaint enjoys putting their complaint down in type and seeing it posted amongst similar complaints by all the other complainers of the world. Then there are those of us who are sick of drama and would rather look for rational explanations than insist that everything that everyone else does conform to our own personal perceptions of perfect clarity. Reply
  • kpb321 - Monday, May 21, 2018 - link

    Because it's not a clear and obvious name. Often times the amount of memory is included in the description of the card after the name even when the card doesn't come in multiple memory configurations. When the card is available in multiple memory configurations including the amount of memory is very common and in those cases it doesn't indicate any differences other than the amount of memory. Compare a 570 or 580 from AMD 4gb or 8gb only changes the amount of memory on the card nothing else. Look at past NVIDIA generations and that was generally true too. They could have just as easily called it the GTX 1050 T or 1050 Se or any number of other things to make it immediately obvious that this card isn't just a GTX 1050 with an extra GB of memory. Reply
  • bji - Tuesday, May 22, 2018 - link

    In any and all cases benchmarking would have to be done to detect and quantify the actual performance differences. Different 'classes' of GPU only roughly fit into performance profiles (sometimes the gaps are wide, sometimes narrow, in performance between different parts even with different names such as 1080ti and Titan Xp), so when the performance between parts is likely very very close, the vendor may not feel it is in the consumer's interest to create a different name that would seem to identify a larger difference in performance than is actually present.

    Besides, most 3rd party vendors overclock and add all kinds of tweaks to the cards which means that without very explicit benchmarking, you can only be roughly certain of a given card's likely performance. So someone choosing between a 1050 and a 1050 3 GB really is in no different position than anyone else choosing a card by different vendors. Even the same vendor may use similar names for parts with differences in performance similar to the 1050/1050 3 GB (which is to say, small differences likely not noticeably by most consumers for most workloads).
    Reply
  • 29a - Monday, May 21, 2018 - link

    I agree with you BJI, as far as I am concerned they have different names and I have no problems telling which is which. I'm not sure if these people are slow or just here to argue. Reply
  • 29a - Monday, May 21, 2018 - link

    Oh and I forgot to add, finish the fucking Ryzen article. Reply
  • DanNeely - Tuesday, May 22, 2018 - link

    I'm going to blame mining for this thing existing. Nvidia and AMD have both said they're making and selling cards as fast as their supply chain will let them; salvaging this odd config - even if they don't have a whole lot of them - is one way to stretch part of it farther. And as noted it's far from the first screwball salvaged die model ever released; although weird low volume configs have more often been released as low end quadro or OEM only models vs something available at retail. Reply
  • mapesdhs - Tuesday, May 22, 2018 - link

    Marketing/PR at NVIDIA have it seems learned nothing from their recent meddlings with specs and branding. It won't stop until consumers cease buying the products and tech sites loudly call them out on such nonsense, ditto AMD when they do it. Reply
  • 0siris - Tuesday, May 22, 2018 - link

    "With GPP, we asked our partners to brand their products in a way that would be crystal clear. The choice of GPU greatly defines a gaming platform. So, the GPU brand should be clearly transparent – no substitute GPUs hidden behind a pile of techno-jargon."
    -nvidia, less than three weeks ago

    "We hold everyone around us to standards we ourselves wipe our ass with regularly because we became too big for our boots and all we care about is money"
    Reply
  • vvume - Thursday, May 24, 2018 - link

    Designed for Netflix 4K? Regular 1050 couldn't run it because it only had 2GB of RAM (Netflix requires 3G oddly) and so 1050 Ti is the lowest card that PCs can run Netflix 4k. Reply

Log in

Don't have an account? Sign up now