The TU116 GPU and First Look at Cards from MSI and EVGA
Is this the true GTX 1060 successor?
NVIDIA is introducing the GTX 1660 Ti today, a card build from the ground up to take advantage of the new Turing architecture but without real-time ray tracing capabilities. It seems like the logical next step for NVIDIA as gamers eager for a current-generation replacement to the popular GTX 1060, and who may have been disappointed with the launch of the RTX 2060 because it was priced $100 above the 1060 6GB, now have something a lot closer to a true replacement in the GTX 1660 Ti.
There is more to the story of course, and we are still talking about a “Ti” part and not a vanilla GTX 1660, which presumably will be coming at some point down the road; but this new card should make an immediate impact. Is it fair to say that the GTX 1660 Ti the true successor to the GTX 1060 that we might have assumed the RTX 2060 to be? Perhaps. And is the $279 price tag a good value? We will endeavor to find out here.
RTX: Off
It has been a rocky start for RTX, and while some might say that releasing GTX cards after the fact represents back-peddling from NVIDIA, consider the possibility that the 2019 roadmap always had space for new GTX cards. Real-time ray tracing does not make sense below a certain performance threshold, and it was pretty clear with the launch of the RTX 2060 that DLSS was the only legitimate option for ray tracing at acceptable frame rates. DLSS itself has been maligned of late based on a questions about visual quality, which NVIDIA has now addressed in a recent blog post. There is clearly a lot invested in DLSS, and regardless of your stance on the technology NVIDIA is going to continue working on it and releasing updates to improve performance and visual quality in games.
As its “GTX” designation denotes, the GeForce GTX 1660 Ti does not include the RT and Tensor Cores that are found in GeForce RTX graphics cards. In order to deliver the Turing architecture to the sub-$300 graphics segment, we must be very thoughtful about the types and numbers of cores we use in the GPU: adding dedicated cores to accelerate Ray Tracing and AI doesn’t make sense unless you can first achieve a certain level of rendering performance. As a result, we chose to focus the GTX 1660 Ti’s cores exclusively on graphics rendering in order to achieve the best balance of performance, power, and cost.
If the RTX 2060 is the real-time ray tracing threshold, then it's pretty obvious that any card that NVIDIA released this year below that performance (and price) level would not carry RTX branding. And here we are with the next card, still based on the latest Turing architecture but with an all-new GPU that has no ray tracing support in hardware. There is nothing fused off here or disabled in software with TU116, and the considerable reduction in die size from the TU106 reflects this.
GTX 1660 Ti | RTX 2060 | RTX 2070 | GTX 1080 | GTX 1070 | GTX 1060 6GB | |
---|---|---|---|---|---|---|
GPU | TU116 | TU106 | TU106 | GP104 | GP104 | GP106 |
Architecture | Turing | Turing | Turing | Pascal | Pascal | Pascal |
SMs | 24 | 30 | 36 | 20 | 15 | 10 |
CUDA Cores | 1536 | 1920 | 2304 | 2560 | 1920 | 1280 |
Tensor Cores | N/A | 240 | 288 | N/A | N/A | N/A |
RT Cores | N/A | 30 | 36 | N/A | N/A | N/A |
Base Clock | 1500 MHz | 1365 MHz | 1410 MHz | 1607 MHz | 1506 MHz | 1506 MHz |
Boost Clock | 1770 MHz | 1680 MHz | 1620 MHz | 1733 MHz | 1683 MHz | 1708 MHz |
Texture Units | 96 | 120 | 144 | 160 | 120 | 80 |
ROPs | 48 | 48 | 64 | 64 | 64 | 48 |
Memory | 6GB GDDR6 | 6GB GDDR6 | 8GB GDDR6 | 8GB GDDR5X | 8GB GDDR5 | 6GB GDDR5 |
Memory Data Rate | 12 Gbps | 14 Gbps | 14 Gbps | 10 Gbps | 8 Gbps | 8 Gbps |
Memory Interface | 192-bit | 192-bit | 256-bit | 256-bit | 256-bit | 192-bit |
Memory Bandwidth | 288.1 GB/s | 336.1 GB/s | 448.0 GB/s | 320.3 GB/s | 256.3 GB/s | 192.2 GB/s |
Transistor Count | 6.6B | 10.8B | 10.8B | 7.2B | 7.2B | 4.4B |
Die Size | 284 mm2 | 445 mm2 | 445 mm2 | 314 mm2 | 314 mm2 | 200 mm2 |
Process Tech | 12 nm | 12 nm | 12 nm | 16 nm | 16 nm | 16 nm |
TDP | 120W | 160W | 175W | 180W | 150W | 120W |
Launch Price | $279 | $349 | $499 | $599 | $379 | $299 |
The New TU116 GPU
The GeForce GTX 1660 Ti is based on a brand new “TU116” Turing GPU that’s been carefully architected to balance performance, power, and cost. TU116 includes all of the new Turing Shader innovations that improve performance and efficiency, including support for Concurrent Floating Point and Integer Operations, a Unified Cache Architecture with larger L1 cache, and Adaptive Shading.
We aren't seeing a new graphics architecture here of course, and we have covered Turing in depth previously. TU116 is manufactored on the same 12 nm process as its larger counterparts, but this all-new part is just 284 mm2 (compared to 445 mm2 with TU106). It offers 1536 CUDA cores along with 96 TUs and 48 ROPs. And yes, we are getting the full TU116 with the GTX 1660 Ti, not a "cut down" version.
At the heart of the TU116 GPU lies a new SM. NVIDIA engineers have revamped the structure of the TU116 SM, outfitting this unit with dedicated cores for processing FP32 and integer operations simultaneously, FP16 processing at double the rate of FP32 operations, and have also updated the CUDA Cores to take advantage of the latest advancements in programmable shading. Like other Turing GPUs, the TU116 SM also features enhanced caches that are more configurable, offer more capacity, and deliver improved bandwidth. Ultimately, as a result of these changes, the TU116 SM offers a dramatic improvement over the prior generation GP106 Pascal SM used in GeForce GTX 1060.
And while it seemed that everything leaked out before this launch not every detail was correct, as NVIDIA was clear during our briefing that this part will only be offered with 6GB of GDDR6, and not in various configurations as rumored. The 192-bit memory bus width provides 288.1 GB/s of bandwidth, and this is lower than the same 6GB GDDR6 configuration from the RTX 2060 due to a lower 12 Gbps data rate (14 Gbps with the RTX 2060). However, it may just be possible that this memory rate can be bumped up to match the RTX cards. (We will cover that a little later…)
A Look at Partner Boards – Featuring MSI and EVGA
- EVGA NVIDIA GeForce GTX 1660 Ti XC Gaming: $279.99 (Amazon.com)
- MSI NVIDIA GeForce GTX 1660 Ti GAMING X: $309.99 (Amazon.com)
This is an add-in board (AIB) launch, meaning that NVIDIA will not be offering a Founders Edition, and that partners can release boards with their own power limits and cooling options. The idea that many of the cards at retail might end up with higher clock speeds that we saw in the above chart could make this an even more interesting product, and for our review we were sent cards by MSI (slightly overclocked) and EVGA (stock).
While only one card might be needed to assess performance from the new GPU these cards are quite different, and showcase the flexibility of implementation AIB partners will have with the GTX 1660 Ti.
The MSI GAMING X
This card is a good example of what partners will be able to offer with the GTX 1660 Ti, with the "highly overclockable" TU116 GPU likely to see many factory overclocks. This GAMING X version from MSI features a 10W increase in TDP over reference to allow for a boost clock of 1875 MHz out of the box, though memory speeds remain at defaults.
There should still be overclocking headroom with the card given its default 130W TDP and high performance Twin Frozr 7 cooler. The higher performance ceiling (and RGB lighting) comes at a modest $30 price premium with the card priced at $309.
The EVGA XC Black
This is an interesting alternative to standard dual-slot designs, providing stock 1660 Ti performance at the base $279 price point and featuring EVGA's short, triple-slot configuration. A single fan pushes air against a compact yet massive heatsink, and while the power limit does not exceed 100% in this particular version there is likely room for some light overclocking as well.
The EVGA card (right side) may look small from the front, but this is a full triple-slot design
On the next page we'll take a look at performance with the GTX 1660 Ti, beginning with synthetic tests and 1080p gaming benchmarks.
OMG! GTX1660Ti can’t even
OMG! GTX1660Ti can’t even keep up with PS5/XBOXx2.
True but those are will
True but those are will probably launch in holiday 2020 which is 22 months away.
thats not even the problem…
thats not even the problem… Check out the Metro PS4 vs PC vs XB1 comparison videos. And then check out the PC requirements for that game. I swear to me it look like the PS4 version (not even the pro version, stock ps4) looks the best. This issue has been going on forever… i thought with the consoles being glorified PC’s that this BS would stop but this trend continues to happen.
Fine, it doesn’t have to look the best on the PC but don’t actually require more processing power for no reason.
They are… out?
They are… out?
I think this is the first
I think this is the first generation of main-stream cards that are actually targeting reasonable 1440p performance or high frame rate 1080p. There are many gamers out there right now that are moving into those segments, and it makes sense to target them with an affordable card.
Yes, but then the question
Yes, but then the question becomes, “what is affordable?”
With this slow creep of increased cost, generation by generation, when will these cards no longer be “affordable?”
I’m going to pull two quotes
I’m going to pull two quotes from the article, and then explain my thoughts.
“the GTX 1060 6GB, which launched at $249 ($299 for the Founders Edition) way back in July 2016”
“$279 might be seen as the inflation-adjusted price for such a card after nearly two years, but that sort of talk will not win me any friends in the lovely world of internet comments”
We are seeing a card for approximately the same street price as the GTX 1060 was 2.5 years ago at launch, that gives ~50% more performance. As I said before, this level of performance is getting into solid 1440p and high FPS 1080p range, which is awesome for gamers. This is a great deal, no matter how you slice it. Yes, it would be better if it was cheaper. That can be said for every product, but in the real world, inflation is a thing, and nVidia is not immune to its effects. The other thing that cant be missed is that launch prices almost always drop, and we will probably see these cards hit $249 sometime this year.
IMO I look at it that it
IMO I look at it that it trades blows with the 1070/1070ti for $100 cheaper than the 1070 launched at. 1660Ti also has a lower TDP, so easier to cool and cheaper to power. Seems like a great value to me.
The performance/price ratio
The performance/price ratio looks great but in time of recession, despite economist lies, cash is king and any cent is worth to spare.
As usual, NV$ try to deplete gamer’s pockets with a higher price tag!
I’m still waiting to see any GTX 1660 under 249 USD… as a true GTX 1060 replacement.
Nice review Sebastian, could
Nice review Sebastian, could you please add a power consumption chart?
Thanks
Yes! Need to add that. I’d
Yes! Need to add that. I’d started power testing at 2560×1440 and didn’t want to mix with older 1080p/ultra results since 1440 has higher draw. Will have that up by this evening as I get through the other cards. Ran out of time last night (aka passed out at around 2am).
Any support for variable
Any support for variable frame rate for freesync?
Great question, it should
Great question, it should since the requirement is Pascal and up but I’ll test with a freesync monitor and the latest driver today.
The NoTX Turing Raster Gaming
The NoTX Turing Raster Gaming edition Card has arrived and the RX590 prices will have to fall along with the Vega 56. And the Vega 56 results are conspicuously absent for the benchmarking comparsion charts and that Vega 56 MSRP is reported to have been strategically lowered to $279 by retailers(1).
So there has to be some updates to the benchmarking to include the RX Vega 56’s stock and overclocked results in the testing results because the price delta between the GTX 1660Ti and Vega 56 amounts to $0.
What AMD needs to also begin offering is some lower binned Vega 20 die based variant that lower binned than the Radeon VII. Vega on a 7nm process node can still offer more competition against Nvidia for raster only gaming at the higher end for AMD.
Any price/performance comparsion charts going forward need to reflect the current Vega selling prices and not MSRP.
(1)
“AMD Partners Cut Pricing of Radeon RX Vega 56 to Preempt GeForce GTX 1660 Ti”
https://www.techpowerup.com/252900/amd-partners-cut-pricing-of-radeon-rx-vega-56-to-preempt-geforce-gtx-1660-ti
“…..There is nothing fused
“…..There is nothing fused off here or disabled in software with TU116….”
See for me I was SURE this was going to be a “binning” part, but I’m truly surprised they actively designed/manufactured/distributed this part as part of a planned out road map, very interesting.
Just because the TU116 base
Just because the TU116 base die tapeout’s top binned part is the GTX 1060Ti does not mean that there can not be a lower binned variant derived from some defective TU116 Die samples. And as always as the result of the imperfect diffusion process there will be Defective DIEs no matter the Base die tapeout that is used.
Nvidia will die harvest any TU116 parts for some lower price segement for any TU116 based GTX 1600 series non Ti/lower variants like a GTX 1650/1630 variants that may be used for OEM PCs and such.
Nvidia is smart to release a line of Raster Gaming Only focuesd GTX 1600 series parts for that market segement that can still rely on all gaming titles having alternative code paths in the games and graphics APIs(Vulkan, DX12/DXR, and such) that will also make some Ray Tracing/AI features available via a shader core accelerated software code path for all non RTX branded Nvidia, and AMD/Intel graphics SKUs.
So you can be damn sure that if Intel/AMD do not have any RTX/AI like enabled GPU IP in their hardware that they with work with both Microsoft and the Khronos Group(Vulkan) to make sure there are alternative code paths for Ray Tracing and AI accelerated workloads doen on the GPU’s shader cores on any GPU SKUs that do not have in hardware support with dedicated Ray Tracing and Tensor cores IP. Nvidia will as well just to continue selling Pascal/NoRTX enabled Turing offerings that can still make use of any alternative code paths Graphics/Compute APIs for GPU SKUs that are not RTX enabled.
Think about this! Intel has acquired the FPGA maker Altera so Intel could very well programm an FPGA to do the Ray Tracing/Bounding Volume Hierarchy (BVH) calcuations as well as FPGA implemented Tensor cores. Ditto for AMD working in partnership with Xilinx for some of the very same functionality. Xilinx and AMD have already worked on some Epyc Based HPC platforms that are paired with Alveo U250 accelerator cards(1).
AMD, via Microsoft’s and Sony’s deep pockets, could have the ability make use of some Xilinx FPGA related IP to be integrated along with next generation AMD Console APUs. And AMD and say Xilinx FPGA IP could be interfaced via AMD’s Infinity Fabric(xGMI) IP and that all can be done via AMD’s EESC division woking in partnership with Microsoft and/or Sony and Xilinx/others for FPGA’s programmed to do what AMDs Console APU hardware can not do currently except on the slower shader core/software code path route.
I really wish that AMD and Xilinx would cosy up even more to better compete with Intel’s massive IP portfolio that’s going to include Diecrete GPUs in the 2020/later time frame. AMD’s advantage of having both GPUs and x86 CPUs under its IP umbrella is going to be coming to an end after 2020. So there will be Intel with CPUs, Descrete GPUs/Integrated graphics, and FPGAs/Memory Other PC Technology IP all under Intel’s umbrella.
Nvidia’s CEO has got to be having some extra sleepless nights pondering Intel’s Descrete GPU market entry what with Intel having the FPGA IP to include with its Descrete GPUs to be programmed for tasks like Ray Tracing/BVH and Matrix Math(Tensor Cores) before Raja’s/Intel’s teams Can get a more ASIC like answer to Nvidia’s RTX IP.
Just remember that AMD’s Pro Pender Software/Plugins support simultaneous CPU cores/GPU Shader Cores Ray Tracing accaleration and that will be what the Khronos Group uses for Vulkan with that code base modified to work inside Vulkan’s API vai some AMD extentions or most likely some Vulkan cross platform inclusion in the non GPU maker limited section of the Vulkan graphics API standard. Some Vulkan extentions that were once made for a single maker’s GPU hardware get adopted as non extentions into the full Vulkan csoss platform specification once the rest of the market begins making use of that extention. And that adoption can be GPU hardware based or software code path based.
(1)
“30,000 Images/Second: Xilinx and AMD Claim AI Inferencing Record”
https://www.hpcwire.com/2018/10/03/30000-images-second-xilinx-and-amd-claim-ai-inferencing-record/
It would be interesting to
It would be interesting to know nVidia’s gross margin on these, because the die size of the 1660ti is quite a bit bigger than the 1060 despite the smaller process (16 to 12nm). GDDR6 is also more expensive than GDDR5 which the 1060’s launched with, yet the price is largely the same. At face value it seems like nVidia is accepting lower margins than the 1060 commanded, which is uncharacteristic of them.
Look to all the
Look to all the TU116/Turing/GTX has with improved Shader Core/SM/Cache/other tweaks and the extre shader cores at that. Turing has a different shader core to SM ratio than GP106/Pascal. The Cache subsystems on Turing are improved above Pascal’s cache subsystems. There is so much more new IP on Turing even with the RTX IP excuded and really how hard is it to bring up TechPowerUp’s GPU database in 2 browser tabs, one for the GTX 1060 and one for the GTX 1660Ti! Turing even without the RTX IP is still going to be larger because of the transistor count need to enable all the improvments to Turing’s Micro-Arch even witout any RTX IP included. And did I say more shader cores!
This is the mainstream GPU market where, unlike the Flagship GPU market, sales volume is where more revenues are to be had. It’s not unrealistic because Nvidia knows that the Raster Oriented gaming titles will continue to rule the roost for a few more years. And AMD, and soon, Intel, along with AMD, will be offering some stiff competition. With AMD come Navi time and after that whatever Intel will be offering.
And thus, Nvidia, who only has GPUs and not much else producing the lion’s share of its revenues will really have mainstream GPU competition from 2 other players and not just one. Nvidia has to retain its GTX Raster Gaming line for that mainstream market(Dominated By Raster Oreinted gaming Titles Currently and for some time to come) in addition to its more costly RTX high end branded RTX line of products that Nvidia is hoping(Betting Billions On) to become a newer gaming standard for games that make use of Ray Tracing and AI in the GPU hardware.
RTX(Ray Tracing/AI-Denoining and AI-AA/Upscaling) is going to be a do or die thing for Nvidia because that’s what Nvidia has and will use to differentiate it GPU offerings until AMD and Intel have the time to catch up. Nvidia will spend billions to get as many gaming titles as possible over the next few years RTX Enabled. So GTX/Turing is a stop gap measure for Nvidia to retain mainstream GPU market share that’s currently based on raster performance and not any RTX performance just yet or even within the next year or possibily two.
Nvidia has only GPUs for its bread and butter unlike AMD(CPUs, GPUs) currently or Intel(CPUs, Discrete GPUs/Incoming, and Optane/Memory/TB3) come 2020. I’d expect to see more GTX/Turing(TU116) lower binned SKUs arriving as the Pascal Variants supply channels run dry.
Here are the GTX 1060 specs:
Shading Units: 1280 TMUs: 80 ROPs: 48 SM Count: 10
GPU Name: GP106
GPU Variant: GP106-400-A1
Architecture: Pascal
Foundry: TSMC
Process Size: 16nm
Transistors: 4,400 million
Die Size: 200 mm²
Here are the GTX 1660Ti specs:
Shading Units: 1536 TMUs: 96 ROPs: 48 SM Count: 24
GPU Name: TU116
GPU Variant: TU116-400-A1
Architecture: Turing
Foundry: TSMC
Process Size: 12nm
Transistors: 6,600 million
Die Size: 284 mm²
The more curious thing for me
The more curious thing for me is if there will be a mobile version of this card? This looks very intriguing in a laptop.
The prices in the table are
The prices in the table are plainly wrong mixing Founders Edition with partner models.
How many stock options did nVidia give for this review?
I wish you would add fan
I wish you would add fan noise measurements. I’m explicitly getting a 1660ti because of the low Watts, and therefore hopefully quiet fan noise.
Oh boohoo! Why can’t Nvidia
Oh boohoo! Why can’t Nvidia make a GPU that can run 3 4K screens at 120hz and then give it away for free? They’re rich, they can afford it. But they won’t, and you know why? Because they are EVIL and GREEDY!
Wait a minute.. I understood
Wait a minute.. I understood that the ‘Turing AND ray tracing was only going to be available on the 2080 series..
When did NVIDIA change that stance from as little as 6 months ago? I use NVIDIA cards for creating content using Octane render is the reason I ask…. the 2080s and in ‘Ti’ being well over priced this series, imo, is why I ask.