[Videocardz] 1650 details

Page 6 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

Yotsugi

Golden Member
Oct 16, 2017
1,029
487
106
Yes, which is more than compensated for by vastly superior performance and efficiency
Well, no.
Turing burned a lot of area for very little in terms of perf and some decent-ish power gains.
AMD being able to support sizable R&D budgets due to their fantastic execution on the CPU side
Except that part when the design costs are skyrocketing, so GPUs will always be the second-class citizen for AMD.
The money hasn't been there to develop NAVI
uArch work isn't about money.
# of dies is.
Turing is 5 dies already.
Navi will be lucky to be 2 by EOY.
Nvidia's current Turing architecture is literally the same one from their GTX 200 series
Basically nothing of Tesla remains since Kepler.
But Fermi, yes, nV is still using the same design principles since Fermi, with separated geometry processing and setup pipes.
Most of Nvidia's gains were achieved by increasing the die size and using GDDR6.
Well no, Turing is ~client Volta, and Volta had plenty of uArch changes.
 

ozzy702

Golden Member
Nov 1, 2011
1,151
530
136
WTF are you talking about? Nvidia's current Turing architecture is literally the same one from their GTX 200 series, which was their foray into unified shaders. Ever since then its been iterations of the same architecture. Turing it literally GTX 200 iterated over 7 generations.

Whether AMD uses GCN or not, whatever they call it, doesn't matter. It's all about finding the most efficiency that you can while keeping the product profitable. We'll have to wait and see what Navi brings to the table, but I think it's a fully fledged series that will compete with the RTX 2080ti at a certain degree. Maybe they'll fall a bit short off of it, but they are probably going to sell it much cheaper for it.

I think AMD's performance leap with Navi over Polaris/Vega is going to be bigger than Nvidia's leap from Pascal to Turing. Most of Nvidia's gains were achieved by increasing the die size and using GDDR6.


I literally said that Turing is the result of iteration after iteration so I'm not sure where your confusion comes from. I'd be thrilled to see a high end offering from AMD, but I wouldn't be surprised to see 2070 and under performance targeted.

Pascal to Turing was a decent bump in efficiency when you factor out the added die size dedicated to RTX but yeah, it wasn't huge. NAVI is going to need a much larger bump just to catch up with Turing, let alone "leapfrog". Hopefully my pessimism and skepticism are unfounded and NAVI is a solid performer. It would be nice to see some competition and balance restored to the space.

How boring is the DOA 1650? So boring that none of us want to waste time talking about it! :D
 

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
Pascal to Turing was a decent bump in efficiency when you factor out the added die size dedicated to RTX but yeah, it wasn't huge.
The low-end Turing parts lack RTX components yet they are still massive compared to Pascal. For example GTX 1660Ti is based on a 12nm 284mm2 chip with 6,600 million transistors. By comparison GTX 1080 is based on a 16nm 314mm2 chip with 7,200 million transistors. If we were to normalize performance and account for the 12nm process gains, Turing may actually come up short in terms of efficiency.
 
  • Like
Reactions: SirDinadan

ozzy702

Golden Member
Nov 1, 2011
1,151
530
136
The low-end Turing parts lack RTX components yet they are still massive compared to Pascal. For example GTX 1660Ti is based on a 12nm 284mm2 chip with 6,600 million transistors. By comparison GTX 1080 is based on a 16nm 314mm2 chip with 7,200 million transistors. If we were to normalize performance and account for the 12nm process gains, Turing may actually come up short in terms of efficiency.

As I mentioned before, efficiency per watt is the metric that matters because it opens up so many options in the design. It's easy to get tunnel vision when looking at just die size when we're talking efficiency, but it's just one part of the design. A power efficient design needs less robust power delivery, less cooling, and in NVIDIA's case, a smaller buss because of their memory compression. All this results in a sizable offset to the potentially higher die cost.

Turing is more efficient per watt than Pascal and that's not just because of the minor node shrink and it will become significantly faster in the future with modern games because of it's ability concurrently execution of floating point and integer instructions (see Wolfenstein II for an example which up until Turing heavily favored AMD). Yes there's a slight die size increase in price, but time and time again NVIDIA has more than compensated for this with performance and the pricing of their cards. This is where AMD lags behind greatly, and it's held them back repeatedly.

performance-per-watt_1920-1080.png
wolf 2.png
 

Attachments

  • wolf.png
    wolf.png
    28.1 KB · Views: 6
Last edited:

AtenRa

Lifer
Feb 2, 2009
14,000
3,357
136
Since we are talking about the Chips efficiency (that means perf/watt @ iso perf or iso watt) and not the Graphics Cards efficiency, the above chart is irrelevant.

If you want to see the Chip efficiency then you need to take chips with close the same size and transistor count and measure perf/watt at iso perf or iso watt.
 
  • Like
Reactions: coercitiv

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
As I mentioned before, efficiency per watt is the metric that matters because it opens up so many options in the design.
Agreed, this is why we should take a closer look at the data you quoted from Techpowerup: GTX 1080 is 22% faster than Zotac 1660Ti but also 15% less efficient at stock clocks.

What do you reckon would happen to GTX 1080 efficiency if we were to drop clocks enough as to match 1660Ti in performance? Based on the data above there's room for almost 20% drop in clocks which also means a 10-15% drop in voltage, which in turn translates to 40+% drop in chip power usage. And since power dropped like a brick at chip level, we can redesign the board for a significantly more efficient VRM setup and maybe clock the memory a bit lower too. You won't get a 40% drop in board TDP, but you just might hit 30 lower power, meaning 125-130W TDP. Sounds familiar?

Concurrent execution of FP and INT is a nice touch for Turing, but even that doesn't save it from the realization that most of the efficiency gains seems to have come at the cost of die size increase across the entire lineup. When 1660Ti launched I kept saying the chip must be heavily cut down, as 1660Ti sports 1536 cores while GTX 1080 comes in at 2560 cores. I kept looking at 1660Ti PCB images, seeing room for more memory chips and thinking there might be a 256bit bus in there maybe a 2000+ cores full chip waiting for launch.

But then 1650 came in based on a 200mm2 chip, virtually the same size as GP106 in GTX 1060. Turing chips are big, there's no going around that, and we should take care when comparing them against Pascal chips which were clocked for a different perf/mm2 profile.
 
  • Like
Reactions: Ranulf and Elfear

coercitiv

Diamond Member
Jan 24, 2014
6,151
11,686
136
Are there any?
nV uses a custom node, and the details on it were never reported.
GP106 is 200mm2 and 4,400 million transistors
TU117 is 200mm2 and 4,700 million transistors

I don't know what made the density gain possible. What we do know is TSMC claimed 12nm was originally planned as a fourth variant of 16nm with "lower leakage and better cost characteristics".
 
Last edited: