GP106 speculations

[JDG1980]

There are rumored die shots of the GP106, but no hard data of any kind yet. In the Polaris 10 threads, we've seen this upcoming chip discussed as Nvidia's likeliest direct competitor. So, what can we expect from GP106?

The basic building block of Pascal chips appears to be the GPC cluster. This consists of 10 shader modules with 64 CUDA cores each. That means each GPC has 640 CUDA cores, and we can expect all Pascal chips to have an even multiple of this. Although this was never officially specified anywhere, it's likely there are also 16 ROPs per GPC, which would match what we know about the GP104 (64 ROPs total with four GPCs).

GTX 1070 uses a GP104 chip that has 25% of its shaders disabled (1920 active CUDA cores). Thus, we can assume that GP106 will be even smaller than that. The only logical assumption is that GP106 will have two GPCs, for 1280 CUDA cores and 32 ROPs. This would match what was done with Maxwell, where GM206 was almost exactly half of a GM204.

What does this mean in terms of performance? The GM206-based GTX 960 ran at about the same clock rate as the GM204-based GTX 980, and provides about 53% as much performance at 1080p. (TPU's latest ranking puts the 960 at 34% of a GTX 1080, and the 980 at 64%. 34 / 64 ~= 0.53.) This is very close to what we'd expect, given that GM206 has almost exactly half the resources of GM204. If we go with the assumptions outlined above, we could place a hypothetical GTX 1060 at 53% (with GTX 1080 at 100%). That's a bit lower than the GTX 970 (55%) and R9 390 (56%). But it should still be just powerful enough to qualify as "VR Ready" based on Steam's benchmarks, if Nvidia thinks that is a useful marketing point.

How about performance per watt? The GM206 actually had the worst perf/watt of any Maxwell chip. Peak gaming power was 118W for the GTX 960, compared to 184W for the GTX 980 (and GTX 1080). In other words, GTX 960 used 64% as much power as GTX 980 to get only 53% the performance. That is a roughly 17% reduction in perf/watt. Will we see the same thing for GP106? It depends. Some rumors say that GP106 has a 192-bit bus, and if that's the case, it will hurt perf/watt a bit. (They would get similar bandwidth and better perf/watt with a 128-bit GDDR5X bus, but that's likely not yet cost-effective.) Another question is whether Nvidia (and their customers) will be satisfied with only 970/390-level performance. Assuming the memory bus isn't a bottleneck, they could do better by upping the core clocks to 2000 MHz. That would probably outcompete an overclocked GTX 980, but at the cost of a substantial perf/watt hit. It might be difficult to keep it from throttling at a 150W power limit.

Overall, I expect GP106 to fall slightly behind Polaris in terms of both raw performance and perf/watt. It won't be a total blowout, but I think that it is going to be hard to justify the purchase of a GTX 1060 over a RX 480 (or whatever other SKUs that AMD comes out with). Of course, this won't stop Nvidia from selling a ton of them anyway. Sales of the GTX 960 were huge, even though for much of its life it was nearly as expensive as the far more powerful R9 290. (And surely not all of these sales were to people with crappy PSUs, 4K TVs with HDMI 2.0 only, or people who needed HEVC for whatever reason.)

Thoughts?

 

[renderstate]

We don't know if RX480 has any new VR oriented feature but we do know Pascal supports single pass stereo rendering and lens matched shading. Once applications start making use of these features I'd expect GP106 to be significantly faster than RX480 in VR (again, assuming RX480 doesn't have some yet undisclosed VR ace up its sleeve). On the other hand all AMD talk about is enabling VR with RX480 so it'd quite ironic if they get beaten at that.

 

[maddie]

If it's 192 bit memory bus width, we can reasonably assume 3GB and 6GB variants. Could this cause some 2nd thoughts especially if power and price are considered?

Will be interesting to see how some respond after the flak the 4GB RX480 has gotten.

 

[tviceman]

If GP106 uses the same arrangement of clusters per GPC then it will be 1280 cuda cores.

However GM107 did not have the same arrangement and Nvidia may opt to do a different combination of shader modules per GPC cluster with GP106 and GP107 because otherwise GP107 would only have 25% of the raw computing power of GP104 ending up with a very small jump over GM107.

 

[parvadomus]

If nvidia follows the same route as with maxwell I expect GP106 (basically GP104/2) to get annihilated by Polaris 10. No matter what VR magic pascal may have.

 

[maddie]

If GP106 uses the same arrangement of clusters per GPC then it will be 1280 cuda cores.

However GM107 did not have the same arrangement and Nvidia may opt to do a different combination of shader modules per GPC cluster with GP106 and GP107 because otherwise GP107 would only have 25% of the raw computing power of GP104 ending up with a very small jump over GM107.

But GM107 was a sort of first pass at Maxwell, then refined for successive designs. This time, all should be consistent.

 

[tviceman]

But GM107 was a sort of first pass at Maxwell, then refined for successive designs. This time, all should be consistent.

GM107 was just as efficient in perf/w as GM204. It didn't have some of the more advanced features, which is probably the "pass" you're talking about. GM107 was 5 SMM modules, while GM204 was 16 SMM modules. If GP107 is 5 GPC clusters, it would be significantly smaller in size and lesser in performance vs. GP104 in comparison to the Maxwell family. It stands to reason that there is just as much likelihood GP107 (and perhaps GP106) could have different GPC cluster arrangements (say 8 clusters of 64 cores, instead of 10 clusters). This would allow GP106 to be 60% that of GP104 with 3 GPC's, and GP107 to be 40% that of GP104.

All I'm saying is they did it with Maxwell and having a GP107 being so under-powered with 1/4 GP104 power, the possibility of it happening again is real.

 

[tviceman]

If nvidia follows the same route as with maxwell I expect GP106 (basically GP104/2) to get annihilated by Polaris 10. No matter what VR magic pascal may have.

You've got a point. Why even release it at all then? Nvidia should just throw the towel in now amirite?

 

[MrTeal]

If GP106 is pretty much G104/2, it wouldn't be the first time that there's no direct overlap in the two lineups. If that was the case, GP106 would probably be a die that's about halfway between the estimates for P11 and P10. I wouldn't expect nVidia to price the 1060 much lower than the announced $200 price for a (4GB) 480, so the comparison will be an interesting one.

Now, if GP106 is GP104/2, would you expect a 128-bit bus or the proposed 192-bit bus?

 

[kagui]

We don't know if RX480 has any new VR oriented feature but we do know Pascal supports single pass stereo rendering and lens matched shading. Once applications start making use of these features I'd expect GP106 to be significantly faster than RX480 in VR (again, assuming RX480 doesn't have some yet undisclosed VR ace up its sleeve). On the other hand all AMD talk about is enabling VR with RX480 so it'd quite ironic if they get beaten at that.

so far NVIDIA havent showed any advantages on VR the simultaneous multi-projection, seems like a tech that doesnt decrease lag betwen frames, so you are going to get dizzy faster than whit AMD implementation

 

[JDG1980]

Now, if GP106 is GP104/2, would you expect a 128-bit bus or the proposed 192-bit bus?

It's hard to say, but I'm leaning towards the 192-bit rumors being true. First of all, the 128-bit bus on GM206 was an aberration (it surprised me at the time); GK106 had a 192-bit bus. Also, keep in mind that GP104 (at least the top SKU) has GDDR5X now. To get half the performance of GTX 1080, about half the memory bandwidth will probably be required. GTX 1080 has 320 GB/sec of bandwidth, so we're looking at a target of no less than 160 GB/sec for GTX 1060. Even with a RAM clock of 2000 MHz, standard GDDR5 (top line 8Gbps modules) in a 128-bit bus would provide only 128 GB/sec of bandwidth. With a 192-bit bus, they could use cheaper 7Gbps modules and still get 168 GB/sec, which is more than enough.

GTX 960 debuted at $199. But I don't think Nvidia wants to directly compete with AMD on price. I wouldn't be surprised to see the base GTX 1060 3GB model launch at $249, with a ~$30 premium for going up to 6GB. Nvidia's brand recognition guarantees they'd get plenty of sales even if performance is slightly inferior to AMD's entries. They can always drop the price later if they're not moving enough product.

 

[renderstate]

so far NVIDIA havent showed any advantages on VR the simultaneous multi-projection, seems like a tech that doesnt decrease lag betwen frames, so you are going to get dizzy faster than whit AMD implementation

Pascal has the most fine grained graphics preemption available on any GPU, which will help cutting lag. The other VR features decrease the rendering time and will help maintaining a stable frame rate even with the most challenging content. Alternatively they can be used to render the scene at higher resolution to get rid of aliasing artifacts.

 

[renderstate]

If lens matched shading is effective at reducing the number of shaded pixels and if developers adopt it..even supposedly smaller SKUs based on GP106/7 might pull some VR miracle

 

[mohit9206]

Here's what i think.They will do the same thing they did with GTX960.Performance far away from 1070 so that it does not cannabilize 1070 sales but delivering equal price/perf to 1070.Now you do the math.If priced at $249 i expect perf to be similar to 970.Either way perf/$ is without a doubt going to be lower than 480.But obviously that won't stop 1060 from outselling RX480 4:1.

 

[FORTHEWIND]

The basic building block of Pascal chips appears to be the GPC cluster. This consists of 10 shader modules with 64 CUDA cores each. That means each GPC has 640 CUDA cores, and we can expect all Pascal chips to have an even multiple of this. Although this was never officially specified anywhere, it's likely there are also 16 ROPs per GPC, which would match what we know about the GP104 (64 ROPs total with four GPCs).

GTX 1070 uses a GP104 chip that has 25% of its shaders disabled (1920 active CUDA cores). Thus, we can assume that GP106 will be even smaller than that. The only logical assumption is that GP106 will have two GPCs, for 1280 CUDA cores and 32 ROPs. This would match what was done with Maxwell, where GM206 was almost exactly half of a GM204.

I don't think GP106 could be using 64 CUDA Cores per SM a-la GP100. I could be wrong though. GP104 used the same SM config as GM204 (128 CUDA Cores per SM) meaning 32+32+32+32 CUDA Cores blocks in a SM(A TPC house a single SM plus PolyMorph). If GP106 really does use GP100 64 CUDA Cores block (32+32) in a SM config (A TPC with 2 SM inside it equaling 128 CUDA Cores) then GP106 may have a chance with Polaris (Both have same stream processing width so GP106 get same treatment with GCN). You can check both GP100 and GP104 block diagram if you don't believe me. GPU cores count would still be half a GP104 but throughput may be different. Again I could be wrong.

 

[Qwertilot]

We'll see

If the 480 does fall out 25-30% slower than the 1070 then it would surely be slightly odd if NV didn't have something at least moderately close to it?

At least the base model. If there is a 480 model pushing up to say 10-15% slower with big clocks, then I slightly doubt if people will try matching that by pushing the 1060. Just upsell to a 1070 instead.

There's an actual rumor on page 150 of the Pascal thread: ~70% 1070 performance, 192gb bus/6 GB ram, 90-100w power and $249-279.

May very well just be speculation of course, but it seems semi plausible. That would I think probably sell quite well, and maybe deserve to if it was that much more efficient than the 480.

Of course a huge battleground for these chips is going to be laptops.

 

[antihelten]

The basic building block of Pascal chips appears to be the GPC cluster. This consists of 10 shader modules with 64 CUDA cores each. That means each GPC has 640 CUDA cores, and we can expect all Pascal chips to have an even multiple of this. Although this was never officially specified anywhere, it's likely there are also 16 ROPs per GPC, which would match what we know about the GP104 (64 ROPs total with four GPCs).

The 64 cores per SM is only true for GP100, GP104 has the same 128 cores as Maxwell did, and GP106 (and GP107) will likely also have 128 cores per SM. By extension this of course also means that these GPUs only have 5 SMs per GPC not 10.

GM107 was just as efficient in perf/w as GM204. It didn't have some of the more advanced features, which is probably the "pass" you're talking about. GM107 was 5 SMM modules, while GM204 was 16 SMM modules. If GP107 is 5 GPC clusters, it would be significantly smaller in size and lesser in performance vs. GP104 in comparison to the Maxwell family. It stands to reason that there is just as much likelihood GP107 (and perhaps GP106) could have different GPC cluster arrangements (say 8 clusters of 64 cores, instead of 10 clusters). This would allow GP106 to be 60% that of GP104 with 3 GPC's, and GP107 to be 40% that of GP104.

All I'm saying is they did it with Maxwell and having a GP107 being so under-powered with 1/4 GP104 power, the possibility of it happening again is real.

A slight correction, but you're messing up the terms slightly here. A Pascal GPC cluster is not the same as an Maxwell SMM module. In Pascal each GPC cluster contains 5 SMs (10 in the case of GP100), and each SM contains 128 CUDA cores (64 in the case of GP100). In Maxwell each GPC had 5 SMs in the case of Maxwell 1 (GM107), and 4 in the case of Maxwell 2 (GM206, GM204 and GM200), and again each SM has 128 CUDA cores.

As such there are really only two realistic possibilities for GP107. Either it has 1 GPC cluster with 5 SMs, just like GP104 (and GP106 most likely), or Nvidia adds an extra SM like the did with GM107, and it has 640 CUDA cores.

If lens matched shading is effective at reducing the number of shaded pixels and if developers adopt it..even supposedly smaller SKUs based on GP106/7 might pull some VR miracle

It's worth noting that if GP106 sees a similar improvement over GM206, as what GP104 saw over GM204, then GP106 would pretty much match a 970 in performance, which is already the recommended minimum for VR. So even without lens matched shading and single pass stereo, GP106 should be a decent entry level VR GPU.

In Nvidias demonstration of the aforementioned features they saw a speedup of roughly 33%. That would make GP106 roughly equal to an R9 Fury. Coincidently AMD has claimed that RX 480 would match the VR performance of a $500 GPU, which has been widely interpreted as being the R9 Nano, which has roughly the same performance as the R9 Fury. So based on these rather vague numbers, it would appear that a GP106 and an RX 480 will have roughly the same performance in VR.

As for GP107, if this is indeed half of a GP106, then I doubt it will be a particularly good VR GPU even with the aforementioned features. A 33% speedup would still fall well short of the minimum level (970 performance).

 

[ShintaiDK]

The current GP106 leak from sources that is relatively trustful is, that GP106 is 6GB, 192bit, 90-100W, ~970 and ~980 performance in its 2 bins. 249-279$ for top bin. In other words, somewhat equal to Polaris 10. Just much better perf/watt and multi projection for VR.

 

[Krteq]

Except price of course. And GCN also can render to multiple viewports in single pass (aka marketing "simultaneous multi-projection"), even Maxwll can

 

[sontin]

It's worth noting that if GP106 sees a similar improvement over GM206, as what GP104 saw over GM204, then GP106 would pretty much match a 970 in performance, which is already the recommended minimum for VR. So even without lens matched shading and single pass stereo, GP106 should be a decent entry level VR GPU.

GP106 should be much faster than a GTX970:
GP106: 1280C x 1750Mhz x 2 = 4.5 TFLOPs
GM204: 1664C x 1200Mhz x 2 = 3.99 TFLOPs

That is alone a 15% increase of the compute performance and nearly on par with the GTX980. Combine this with 1.33x (Single-Stereo-Pass) and 1.33x (Lens Matched Shading) GP106 should be roughly 2x faster than the GTX970 with VR.

 

[ShintaiDK]

Except price of course. And GCN also can render to multiple viewports in single pass (aka marketing "simultaneous multi-projection"), even Maxwll can

Then why dont they do it?

 

[Flapdrol1337]

GP106 should be much faster than a GTX970:
GP106: 1280C x 1750Mhz x 2 = 4.5 TFLOPs
GM204: 1664C x 1200Mhz x 2 = 3.99 TFLOPs

The way I see it it should be a little over (because of the relatively higher powerlimit) half a 1080. Looking at the graphs at techpowerup I'd say about on par with a 970, which is 55% of a 1080.

a 960 is also a few % over half of a 980.

 

[Krteq]

Combine this with 1.33x (Single-Stereo-Pass) and 1.33x (Lens Matched Shading) GP106 should be roughly 2x faster than the GTX970 with VR.

Both "technologies" are already present in maxwell.

 

[ShintaiDK]

Both "technologies" are already present in maxwell.

Why haven't we seen it then? Same for your GCN claim.

Perhaps because its not there?

 

[DownTheSky]

The current GP106 leak from sources that is relatively trustful is, that GP106 is 6GB, 192bit, 90-100W, ~970 and ~980 performance in its 2 bins. 249-279$ for top bin. In other words, somewhat equal to Polaris 10. Just much better perf/watt and multi projection for VR.

You have no idea of RX 480 power consumption. 150W is max board power (75W from PCIe + 75W from 6 pin). Most likely will draw a lot less.

Edit: Interesting point.
https://youtu.be/PJ5wYVu-tM8?t=562