Geforce GTX 1060 Thread: faster than RX 480, 120W, $249

[Erenhardt]

The GTX 1070 block diagram and specifications?

Same ROPs as the 1080 with one GPC disabled. The ROPs are the dark blue blocks next to the L2 cache in the chip diagram for the GTX 1080 and GTX 1070.

That proved to be unreliable with cut down GM204 part

 

[AtenRa]

The GTX 1070 block diagram and specifications?

Same ROPs as the 1080 with one GPC disabled. The ROPs are the dark blue blocks next to the L2 cache in the chip diagram for the GTX 1080 and GTX 1070.

I can tell you with 99% certainty that ROPs are not connected to Memory Controllers but directly to L2 Cache.

 

[dark zero]

Maybe they learned from their mistake of overhyping the RX 480, and have now gone the complete opposite direction for the GTX 1060? Set themselves up for disappointment, only to be pleasantly surprised at launch!

I wish that were true, but we all know the real reason... :\

As I said... Pascal can't perfectly scale down... I fear that Pascal as Polaris didn't met the expected targets. Polaris can't scale up and Pascal on the opposite.

Also... Pascal has a disastrous FP64 (double precision) and still no Quadro counterpart is announced.

Still. Expecting the 6GB 1060 gives some more dynamism to the RX480 and starts to deliver even better drivers. Both might be OK for the budget market.

 

[Wall Street]

I can tell you with 99% certainty that ROPs are not connected to Memory Controllers but directly to L2 Cache.

http://wccftech.com/nvidia-geforce-gtx-970-memory-allocation-issue-returns-56-rops-64/

See, the ROPs, L2 and memory controllers form functional units similar to how the Dispatch units, shader cores, register files and L1 form an SM. There is no crossbar between the ROPs, L2 and the memory controllers, each ROP can only address its memory controller and each L2 segment only buffers data for its memory controller. You cannot add or remove memory controllers without adding or removing a corresponding number of ROPs if I understand nVidia Maxwell/Pascal architecture correctly.

 

[AtenRa]

http://wccftech.com/nvidia-geforce-gtx-970-memory-allocation-issue-returns-56-rops-64/

See, the ROPs, L2 and memory controllers form functional units similar to how the Dispatch units, shader cores, register files and L1 form an SM. There is no crossbar between the ROPs, L2 and the memory controllers, each ROP can only address its memory controller and each L2 segment only buffers data for its memory controller. You cannot add or remove memory controllers without adding or removing a corresponding number of ROPs if I understand nVidia Maxwell/Pascal architecture correctly.

Nice, have a look at the last memory controller (MC). Only the L2 cache is cut down including 8x ROPs that are directly connected to the L2 Cache, not the MC. There are still 265bit Memory controllers, just the last 32bit Memory Controller is not directly connected to L2 cache but through the 32bit MC next to it, and thus the lower bandwidth because they share the same interconnect and same 256KB of L2 Cache.

If ROPs were directly connected to MC then GTX 970 would have 64 ROPs, same as GTX 980. But, because they cut off the L2 Cache (256kB out of 2MB), 8x ROPs were also cut off and thus we have the GTX 970 with 56 ROPs instead of the 64 found in GTX 980 (full GM204 die)

 

[MrTeal]

Nice, have a look at the last memory controller (MC). Only the L2 cache is cut down including 8x ROPs that are directly connected to the L2 Cache, not the MC. There are still 265bit Memory controllers, just the last 32bit Memory Controller is not directly connected to L2 cache but through the 32bit MC next to it, and thus the lower bandwidth because they share the same interconnect and same 256KB of L2 Cache.

If ROPs were directly connected to MC then GTX 970 would have 64 ROPs, same as GTX 980. But, because they cut off the L2 Cache (256kB out of 2MB), 8x ROPs were also cut off and thus we have the GTX 970 with 56 ROPs instead of the 64 found in GTX 980 (full GM204 die)

http://international.download.nvidi...al/pdfs/GeForce_GTX_1080_Whitepaper_FINAL.pdf

Page 9 -

The GeForce GTX 1080 features eight 32-bit memory controllers (256-bit total). Tied to each 32-bit memory controller are eight ROP units and 256 KB of L2 cache. The full GP104 chip used in GTX 1080 ships with a total of 64 ROPs and 2048 KB of L2 cache.

 

[JoeRambo]

Makes perfect sense for Nvidia to copy-paste 6 Pascal blocks of 8 ROP/32 bit mem controllers. They were burnt by 970 memgate, so they are certainly keen to avoid any fancy setups...

120 TMUs, 48 ROPs and ~5TFlops @ 2Ghz OC clocks should beat 480 without any problems just by brute force pixel/texture pushing?

 

[Thala]

There are still 265bit Memory controllers, just the last 32bit Memory Controller is not directly connected to L2 cache but through the 32bit MC next to it, and thus the lower bandwidth because they share the same interconnect and same 256KB of L2 Cache.

Actually all MCs have the same bandwidth. Issue was, that if they were going with 1KB stride for all 8 MCs (like with GTX980) the 7th ROP/L2 interface would get double amount of requests in average with typical linear buffer allocation.
So instead of scratching the 8th MC completely, they assigned the upper 512kByte as special purpose pool while keeping the 1KB stride for the other 7 MCs for general purpose pools.
If you ask me, i would have scratched the 8th MC + DRAM.

 

[tweakboy]

Rather wait it out more. I play @ 1080p ,, if the 1080 can run 60fps at all times with ray tracing used. No one will get 60fps @ 4k resoulution for games is a disaster,, ... thx

 

[Sweepr]

Many pictures of the card here: www.chiphell.com/thread-1613381-1-1.html

Already in reviewers hands.

 

[sm625]

The lack of SLI support seems to me to be another indicator that this card could be priced lower than most people think, or faster than most people think. Since no one is going to be buying two of these instead of a single 1080, they would not need to worry about SLI 1060s cannibalizing 1080 sales.

 

[vissarix]

Makes perfect sense for Nvidia to copy-paste 6 Pascal blocks of 8 ROP/32 bit mem controllers. They were burnt by 970 memgate, so they are certainly keen to avoid any fancy setups...

120 TMUs, 48 ROPs and ~5TFlops @ 2Ghz OC clocks should beat 480 without any problems just by brute force pixel/texture pushing?

Im pretty sure the gtx1060 wont need 2ghz clocks to beat the rx480...it will come at 1.5 perhaps 1.6ghz and that will be enough to beat the rx480...

 

[tviceman]

The lack of SLI support seems to me to be another indicator that this card could be priced lower than most people think, or faster than most people think. Since no one is going to be buying two of these instead of a single 1080, they would not need to worry about SLI 1060s cannibalizing 1080 sales.

If there really is a 3gb without SLI support... 3gb is really pushing the minimum spec limit at this point in time, but I agree with the reason Nvidia wouldreove SLI support on a cheaper variant. Given that this is 1/2 the cores of a GTX 1080, but 75% of the ROP's and 60% of the bandwidth, GP106 in SLI should outperform a GTX 1080 when scaling is high.

 

[Glo.]

 

[tviceman]

Right at GTX 980 speed (within 2%):

Given that the 980 is ~11% faster than the RX 480 on average, and only losing in Hitman, my prediction was spot on and you were way off Glo.

 

[96Firebird]

For comparison...

GTX 1060 6GB Fire Strike Graphics score: 13,215
RX 480 8GB Fire Strike Graphics score: 12,195

GTX 1060 6GB Fire Strike Ultra Graphics score: 2,928
RX 480 8GB Fire Strike Ultra Graphics score: 2,682

Scores taken from images above and Guru3D's RX 480 8GB review.

System specs for GTX 1060 6GB unknown...

 

[tviceman]

I'd love to see a $250 price point for the 6gb, but who are we kidding here. It's going to be $279-299, with $299 being a bit of a slap in the face.

 

[sirmo]

If this card is faster than rx480 it will be $299.. that's my feeling on it.

 

[swilli89]

For comparison...

GTX 1060 6GB Fire Strike Graphics score: 13,215
RX 480 8GB Fire Strike Graphics score: 12,195

GTX 1060 6GB Fire Strike Ultra Graphics score: 2,928
RX 480 8GB Fire Strike Ultra Graphics score: 2,682

Scores taken from images above and Guru3D's RX 480 8GB review.

System specs for GTX 1060 6GB unknown...

So if these are the assumptions you would have to go 480, historical data tells us it will be about 30% faster 18 months from now. It would be going against the agreed trend to say that it will age well against the 480 while all the of the data from the last 3 years says the 480 will age much better.

 

[sirmo]

So if these are the assumptions you would have to go 480, historical data tells us it will be about 30% faster 18 months from now. It would be going against the agreed trend to say that it will age well against the 480 while all the of the data from the last 3 years says the 480 will age much better.

When I game with an rx480 the GPU utilization rarely ever crosses 60%. I tend to agree, I think there is some untapped potential there. Whether the future drivers can get more perf. out of it is a good question.

 

[96Firebird]

So if these are the assumptions you would have to go 480, historical data tells us it will be about 30% faster 18 months from now. It would be going against the agreed trend to say that it will age well against the 480 while all the of the data from the last 3 years says the 480 will age much better.

Performance now, or potential performance 3 years from now?

Not a real tough choice in my opinion, but I don't live with the motto "wait and see!"

 

[sirmo]

Performance now, or potential performance 3 years from now?

Not a real tough choice in my opinion, but I don't live with the motto "wait and see!"

rx480 will likely have better perf/dollar also 8gb of VRAM.. without even counting FreeSync.. if they are only 5-10% performance difference, rx480 is still a better buy.

 

[AtenRa]

Performance now, or potential performance 3 years from now?

Not a real tough choice in my opinion, but I don't live with the motto "wait and see!"

Im with you, better DX-12 performance today

 

[railven]

Performance now, or potential performance 3 years from now?

Not a real tough choice in my opinion, but I don't live with the motto "wait and see!"

When I had a similar discussion with another forum member here last year he rattled off a list of dx12 games that would make me question going 980 to over fury x.

Of those almost 10 games he listed so far about two have manifested with a win for AMD.

AMD has really sold the wait and see. Unfortunately I let my 7970s go before they got their day in the sun but that was well two years into owning them.

 

[Thala]

Given that the 980 is ~11% faster than the RX 480 on average, and only losing in Hitman

Here you are not up-to-date. If we are going by computerbase.de, who got the driver which fixed the PCIe bandwidth bug you get following results:

980 is 4% faster than RX480,
RX480 wins against GTX980 in:

- Ashes of Singularity
- Call of Duty Black Ops 3
- Hitman
- Just Cause 3
- The Division
- The Witcher 3

Out of these only Hitman and Ashes are DX12. It is safe to assume none of these games made it on NVidias list, which led to the 15% faster claim.

In any case, nice performance of GTX1060, inline with early expectations of around GTX980 performance.