Nvidia Maxwell GM200 pictured

[Carfax83]

Maxwell cores are individually more efficient than GCN cores so core count isn't a good indicator, and NVidia also has ways of mitigating their bandwidth disadvantage, like increasing the size of the L2 cache.

It's actually quite amazing that GM204 can compete so favorably against Hawaii at 4K, considering the gap in overall bandwidth. I expect GM200 to further refine bandwidth efficiency, making it competitive with AMD's HBM..

 

[Silverforce11]

Well its not a good idea to compare different architectures directly on shaders, TMUs, ROPs or bandwidth. What we can compare is improvements over previous GPUs of the same or similar arch.

If R390X is ~70% faster than R290X, it will be very competitive with GM200 which could end up ~40% faster than 980 due to its specs.

One advantage NV has is extra TDP room to run at higher default vcore & boost clocks.

 

[RussianSensation]

It's actually quite amazing that GM204 can compete so favorably against Hawaii at 4K, considering the gap in overall bandwidth. I expect GM200 to further refine bandwidth efficiency, making it competitive with AMD's HBM..

That's because the memory bandwidth for GCN 1.0 and 1.1 is very inefficient. The theoretical memory bandwidth is very high but the GPU can't take advantage of probably 40%+ of it. My benchmark scores don't go up that much with 336GB/sec memory bandwidth on a 7970, up from a stock value of 264GB/sec. Now that Tonga came out, I understand why because my 336GB/sec 7970 memory bandwidth is at best as good as 240GB/sec on a GCN 1.2 chip.

The gap should widen instead since Maxwell already incorporates a 33% or so increase in memory bandwidth efficiency but AMD is going to bring 40%+ memory bandwidth efficiency and HBM on top of GCN 1.1 290X. In 3DMark Vantage, a good theoretical memory bandwidth test, 285 is 70% faster than a 7950B, despite the latter having 36% more memory bandwidth (240GB/sec vs. Tonga's 176GB/sec).

There is more to it than that since Tonga doubled the rasterized triangles/clock (to match Hawaii's) against GCN 1.0 7950/7970 parts. Still, AMD claims a 40% memory bandwidth efficiency with GCN 1.2 over 1.1, but I think AMD is way too conservative. The improvement is more like double.

285 with a 176GB/sec memory bandwidth is beating an R9 290 with 320GB/sec (mind-blown) in a memory bandwidth limited scenario. You can see 280X under 770, which again goes back to my point that its effective memory bandwidth should be divided by 1.4X at least.

This means 512GB/sec on a 390X would be "equivalent" to over 700GB/sec on a 290X, or alternatively 290X's 320GB/sec is only "equivalent" to 228GB/sec on GCN 1.2 if we take AMD's conservative 40% memory bandwidth improvement as a ballpark. Thus, the 264GB/sec and 320GB/sec memory bandwidth of Tahiti XT and Hawaii are just fluffy theoretical numbers vs. highly efficient memory bandwidth of Maxwell/Tonga. That should change once AMD incorporates Tonga's memory bandwidth efficiencies into a 390X.

It's kinda pointless to compare theoreticals between AMD and NV or even among brands as we've seen with HD6970 vs. 7970 (50% better ROP performance with nearly identical ROP paper specs) or 680 vs. 7970 (the latter is nowhere near as fast as the difference in SPs and memory bandwidth would have implied) or Kepler vs. Maxwell (the latter is faster despite lower functional units and memory bandwidth), but in isolation you can see GCN 1.0 and 1.1 are very memory bandwidth inefficient, despite packing lots of its on paper.

While 980 SLI does compete well against 295X2/290Xs, we have to take it into context what that is. 980 is a next gen mid-range Maxwell chip and chances are AMD's next gen 380X mid-range will be close to a 290X. 980 barely beating 295X2 at 4K isn't impressive in that sense.

You can see 295X2/290Xs are 50%+ faster at 4K than a 7990/690. Chances are dual 390Xs could also be 40-50% faster than 290Xs! All of a sudden 980 SLI looks slow and mid-range level only for a next generation GPU (let's call next gen from Sept 2014 when GM204 launched to Nov 2016 when Pascal is supposed to come out). 980s look good on paper since they are competing against GCN 1.1 with at least a 40% memory bandwidth penalty baked in vs. Tonga, and they use low amount of power but look at 4K performance increase from 290/290X vs. 970/980, it's MEDIOCRE!

If all AMD did was just take 290X, and swap in Tonga's memory bandwidth efficiency in there, a 290X would already beat a 980 at 4K. However, since GPU performance is bound not only by memory bandwidth, it's not as if 390X will scale linearly with the extra bandwidth. Still, if 390X has 512GB/sec HBM, it's memory bandwidth is effectively more than double of a 290X. :awe:

 

[Gloomy]

If all AMD did was just take 290X, and swap in Tonga's memory bandwidth efficiency in there, a 290X would already beat a 980 at 4K. However, since GPU performance is bound not only by memory bandwidth, it's not as if 390X will scale linearly with the extra bandwidth. :awe:

If it didn't scale with the extra bandwidth, AMD wouldn't put the extra bandwidth on it.

Unless they're just testing HBM @ 1.25GHz I don't see the point of clocking it that high. It's a waste of power, if the GPU can't use it.

 

[RussianSensation]

If it didn't scale with the extra bandwidth, AMD wouldn't put the extra bandwidth on it.

Unless they're just testing HBM @ 1.25GHz I don't see the point of clocking it that high. It's a waste of power, if the GPU can't use it.

It could be HBM at 1Ghz, still 3x the BW/watt improvement over 512-bit @ 8Gbps GDDR5!

NV's GM200 will try to counter this with L2 cache and more efficient CUDA cores. We've seen that NV needs way less CUDA cores to match the Steam Processors from AMD.

7970Ghz with 2048 vs. 1536 for 680 (33% difference), but NV has slightly higher real world clocks since it boosted beyond 1058mhz in games.

290X with 2816 vs. 2048 with a 980 (37.5% difference) but more than a 20% advantage in clocks on the GM204 in games. ***Note that in this match up, Maxwell already incorporates a 33% increase in memory bandwidth efficiency.

The 390X's rumoured 4096 SPs is a similar 33% differential with GM200's rumoured 3072 CUDA cores. But, this time it is AMD that gets at least a 40% increase in memory bandwidth efficiency while GM200 is probably going to be very similar to GM204, meaning the gap will widen in AMD's favour, unless GM200 brings another 20-30% efficiency increase over GM204 or NV widens the L2 cache even more (which is what I expect them to do)!

 

[Silverforce11]

On this topic of HBM and the huge decrease in latency..

One of the glaring things we've observed over many generations now is how close AMD's 2nd tier GPU come to the top-tier when the clocks are run at the same speed. The 5850 vs 5870 situation is repeated all the way to the R290 vs R290X. There's ~5% performance deficit for a big decrease in shader count.

This suggests that many of the extra shaders are under utilized, being idled (but without power gating, still consumes power).

Will lower vram latency along with massively increased bandwidth contribute to ensuring higher uptime on each shader? If so, it'll directly lead to better efficiency or performance per shader.

On NV's side, the 2nd tier card at the same clocks never come close. The gap is maintained @ ~whatever the shader cut is. Clearly NV is managing to maximize uptime on each shader.

 

[Carfax83]

One advantage NV has is extra TDP room to run at higher default vcore & boost clocks.

That's a good point. I'm hoping that the GM200 is close to GTX 970 SLI speeds, because I'm looking to ditch SLI and go back to single cards.

 

[RussianSensation]

On NV's side, the 2nd tier card at the same clocks never come close. The gap is maintained @ ~whatever the shader cut is. Clearly NV is managing to maximize uptime on each shader.

That's half the story. You forgot the other half, which is more important -- AMD cuts less of the other key functional units out of its 2nd tier cards, maintains the back-end and doesn't cut VRAM either. NV can cut VRAM, front or back-end or both!

1. HD5870 vs. HD5850 1GB
1600 SPs vs. 1440 SPs (11%)
80 TMUs vs. 72 TMUs (11%)
32 ROPs vs. 32 ROPs
256-bit vs. 256-bit
http://www.anandtech.com/show/2841

2. HD6970 vs. HD6950 2GB
1536 SPs vs. 1408 (9%)
96 TMUs vs. 88 TMUs (9%)
32 ROPs vs. 32 ROPs
256-bit vs. 256-bit

3. HD7970 vs. 7950 3GB
2048 SPs vs. 1792 SPs (14%)
128 TMUs vs. 112 TMUs (14%)
32 ROPs vs. 32 ROPs
384-bit vs. 384-bit
http://www.anandtech.com/show/5476/amd-radeon-7950-review

4. R9 290X vs. 290 4GB
2816 SPs vs. 2560 (10%)
176 TMUs vs. 160 TMUs (10%)
64 ROPs vs. 64 ROPs
512-bit vs. 512-bit
http://www.anandtech.com/show/7481/the-amd-radeon-r9-290-review

That's why with overclocking, the 2nd tier AMD card gets you 90%+ of the performance of their flagship. Over the years NV has cut down the 2nd tier card more and more.

1. During GT200 series, NV reduced 260 216's memory bandwidth and ROP count by 14% over the 280.

2. NV neutered both the GTX470 and GTX570 by 20% on the ROP, memory bandwidth side! compared to the 480/580.

3. While interestingly enough NV maintained the back-end of the 780Ti, they neutered 780's CUDA cores and TMUs by a whopping 25% against the 780Ti.

4. With GM204, NV once again chopped off 23% off the CUDA/TMU count of a 970.

That's why I've been so impressed with AMD's 2nd tier cards for so many years. If it wasn't for mining where the extra SPs were really utilized, for gaming only I would have went with 7950s.

If this continues, 390 should retain the same HBM width, and just be slightly cut down on the SP/TMU side, making it a high-end value buy. I could see a 390 with 3584 SPs / 224 TMUs vs. 4096 SPs / 256 TMUs on the 390X. Based on how NV has cut 20-25% with 470/570/780/970, but AMD have cut 9-14% chances are 2nd tier GM200 will be more cut-down than a 2nd tier 390 card. Now if only AMD brought back 6950 --> 6970 bios flashing unlocking with the 390-->390X, that would be EPIC. Imagine an unlocked $499 390 vs. a $699 GTX1080Ti?

 

[Carfax83]

That's because the memory bandwidth for GCN 1.0 and 1.1 is very inefficient. The theoretical memory bandwidth is very high but the GPU can't take advantage of probably 40%+ of it. My benchmark scores don't go up that much with 336GB/sec memory bandwidth on a 7970, up from a stock value of 264GB/sec. Now that Tonga came out, I understand why because my 336GB/sec 7970 memory bandwidth is at best as good as 240GB/sec on a GCN 1.2 chip.

Saying that the memory bandwidth for GCN 1.0 and 1.1 is very inefficient is quite a statement. AMD's engineers aren't fools, so I don't think they would add a 512 bit memory bus to Hawaii without serious consideration, as that would increase the power consumption significantly.

Also, just because your benchmark scores don't go up that much from overclocking your memory, doesn't mean the architecture is inefficient with memory bandwidth. It could also mean that the computational units aren't bandwidth limited. Some GPUs aren't bandwidth limited at all, even at stock memory speeds.

When I had my GTX 770s, overclocking the memory yielded a larger performance increase than overclocking the core, so it was definitely bandwidth starved to a large degree.

While 980 SLI does compete well against 295X2/290Xs, we have to take it into context what that is. 980 is a next gen mid-range Maxwell chip and chances are AMD's next gen 380X mid-range will be close to a 290X. 980 barely beating 295X2 at 4K isn't impressive in that sense.

It's impressive, because bandwidth is crucial at higher resolutions and the GTX 980 has much less raw bandwidth than Hawaii.

If all AMD did was just take 290X, and swap in Tonga's memory bandwidth efficiency in there, a 290X would already beat a 980 at 4K.

This is kind of a contradiction to what you said earlier. You claimed that Tahiti and Hawaii were inefficient with their bandwidth. So if thats the case, how would increasing their bandwidth even more make it beat the 980 at 4K if it can't already do so with 43% greater bandwidth compared to the GTX 980?

Thing is, a GPU's bandwidth usage is tied to it's computational ability. So increasing bandwidth will not increase performance if the GCN cores and ROPS aren't starved for bandwidth.

However, since GPU performance is bound not only by memory bandwidth, it's not as if 390X will scale linearly with the extra bandwidth. Still, if 390X has 512GB/sec HBM, it's memory bandwidth is effectively more than double of a 290X. :awe:

Yes, those numbers are impressive, but like I said above, unless the R9 390x's computational units can utilize the bandwidth, then it won't make a difference.

We have to assume that AMD's engineers wouldn't increase bandwidth beyond what the GPU is capable of using....but we'll have to see.

 

[Enigmoid]

Maxwell is much more than 33% more efficient from a bandwidth perspective.

850m DDR3 is more than 75% more powerful than the 750m DDR3 when both have the same bandwidth (750m DDR3 is BW limited).

http://www.tomshardware.com/reviews/graphics-performance-myths-debunked,3739-4.html

Now, we didn't pick these two cards at random. We chose them because they come equipped with the same 128-bit interface at 1350MHz, delivering up to 86.4GB/s. At an equal throughput (FPS), their bandwidth utilization should be directly comparable. Their frame rates aren't the same, though. The GeForce GTX 750 Ti achieves higher performance, shown by the frame rate chart. Thus, we normalize the metric for the GM107-based board using the GeForce GTX 650 Ti's performance. That is the third blue line you see in the chart.

The results are impressive. A Maxwell-based GPU appears to deliver 25% more FPS than a Kepler GPU in the same price range, while at the same time reducing its memory bandwidth utilization by 33%. Stated differently, on a per-frame basis, the GeForce GTX 750 Ti needs half of the memory bandwidth in the Metro: Last Light benchmark.

As I have said before I strongly expect the 960 with a 128 bit interface at 7 ghz to keep up with the 770 (within 10% performance) at the same memory clocks with a mild core overclock. The 770 is strongly bandwidth bound so this would equate to a rough doubling of memory efficiency.

The 980 has no problem with +20% over the 780 Ti with 2/3 the memory bandwidth. Or 70% over the bandwidth bound 770 with the same bandwidth. Maxwell is a good deal more than 33% more efficient.

 

[RussianSensation]

Saying that the memory bandwidth for GCN 1.0 and 1.1 is very inefficient is quite a statement.

It's not really an opinion, it's a fact.

Based on AMD's own comments, Tonga has at least 40% more memory bandwidth efficiency vs. GCN 1.0/1.1. However, 3dMark Vantage results show better than that, close to double.

In my previous post, AT's testing shows an R9 285 providing 20% higher 3dMark Vantage fill-rate but it only has 176GB/sec memory bandwidth vs. 320GB/sec of the R9 290. Both Tonga and Hawaii have the same rasterization rate of 4 triangles/clock. TechReport shows similar numbers with Tonga having 70% more performance over the R9 280/7950B despite 176GB/sec vs. 240GB/sec of the 7950!

If Tonga has 40-100% more efficient memory bandwidth, that means GCN 1.0 7970 and GCN 1.1 290X are inefficient.

AMD's engineers aren't fools, so I don't think they would add a 512 bit memory bus to Hawaii without serious consideration, as that would increase the power consumption significantly.

They aren't fools, but they went with 512-bit memory controller not for the reason one might think. They were targeting memory bandwidth/mm2 efficiency to save die space. They did not target a 512-bit bus in order to increase the memory bandwidth throughput efficiency.

The main benefits of the 512-bit bus were 20% less die area than a 384-bit Tahiti XT controller, while increasing the total available memory bandwidth by 20%. Yet, both Hawaii and Tahiti still incur at least a 40%+ penalty for their total memory bandwidth.

That means 7970's 264GB/sec is really "Tonga's/GCN 1.2 equivalent" of 189GB/sec, while R9 290X's 320GB/sec is at best 229GB/sec of GCN 1.2. That's why while 7970/290X/s memory bandwidth looks great on paper, in the real world, it's more like marketing. Like a 700 HP car (Challenger Hellcat) that can't put it's power to the ground as well as a 500 HP all-wheel drive sports car (Nissan GT-R).

Also, just because your benchmark scores don't go up that much from overclocking your memory, doesn't mean the architecture is inefficient with memory bandwidth. It could also mean that the computational units aren't bandwidth limited. Some GPUs aren't bandwidth limited at all, even at stock memory speeds.

I get that but my point was piling on extra memory bandwidth on cards like 7970/R9 280X/290X does very little to boost their performance and probably like you said because they don't have enough functional units. Now with 390X (or 380X?), what if your goal is to increase performance 50%, suddenly add Tonga's 40% increase in efficiency and add HBM to reduce power usage.

When I had my GTX 770s, overclocking the memory yielded a larger performance increase than overclocking the core, so it was definitely bandwidth starved to a large degree.

I never owned 770s or tested them but AT's testing shows that the card was more GPU than memory bandwidth limited. With GPU clocks up 9% and memory clocks up 14%, the average net gain was 9-12%.
http://www.anandtech.com/show/6994/nvidia-geforce-gtx-770-review/17

Either way, since my point was GCN 1.0/1.1 inefficiency, how 770 scales with extra bandwidth is not related.

It's impressive, because bandwidth is crucial at higher resolutions and the GTX 980 has much less raw bandwidth than Hawaii.

We don't know that. What if 980 and Tonga have similarly efficient memory bandwidth? That means in "980/Tonga" terms, R9 290X's 320GB/sec is only as high as 229GB/sec!!! What if 980 is even more memory bandwidth efficient than Tonga itself? That's why comparing memory bandwidth between AMD and NV is a waste of time most of the time. But we can make more interesting comparisons of what major improvements 300 series will have based on what AMD has already done with Tonga.

See, since 980 barely beats a 290X at 4K, 380X will have the benefit of at least 40% more efficient memory bandwidth AND piles more of it due to HBM.

This is kind of a contradiction to what you said earlier. You claimed that Tahiti and Hawaii were inefficient with their bandwidth. So if thats the case, how would increasing their bandwidth even more make it beat the 980 at 4K if it can't already do so with 43% greater bandwidth compared to the GTX 980?

That's only greater on paper. While I probably should backtrack that improving R9 290X's memory bandwidth would automatically let it beats a 980 since it could still be more ROP/TMU/SP limited, I don't think your comparison of Maxwell's memory bandwidth to GCN 1.1 is relevant at all. The only way that would work is if both Maxwell and GCN 1.1 had identical memory bandwidth architecture efficiency, but we don't know that.

Thing is, a GPU's bandwidth usage is tied to it's computational ability. So increasing bandwidth will not increase performance if the GCN cores and ROPS aren't starved for bandwidth.

You are right.

Yes, those numbers are impressive, but like I said above, unless the R9 390x's computational units can utilize the bandwidth, then it won't make a difference.

We have to assume that AMD's engineers wouldn't increase bandwidth beyond what the GPU is capable of using....but we'll have to see.

I would not assume that one. AMD's engineers piled way too much memory bandwidth for an HD4890 and 7950/7970.

HD4870 = 115GB/sec
HD4890 = 129GB/sec
HD5850 = destroys them both with only 128GB/sec, meaning HD4870/4890 had too much wasted memory bandwidth available.

HD7970 = 264GB/sec
HD7970Ghz = 288GB/sec
R9 290X = 320GB/sec but the performance is 30-35% faster, despite only 21% and 11% greater memory bandwidth, meaning HD7970 had too much wasted memory bandwidth.

Tonga though shows that both 7970 and R9 290X are wasting memory bandwidth. :sneaky:

If 390X has 512-640GB/sec, it might also be 'too much' of a good thing but if HBM starts at 1GB/sec, you have your minimum speed there and well for marketing reasons it sounds nice.

 

[Carfax83]

Maxwell is much more than 33% more efficient from a bandwidth perspective.

Yes, I agree with this. The "up to 33%" came solely from the new color compression, but the increased size and improved function of the L2 cache also increased bandwidth efficiency itself as well.

NVidia has been rather coy about how advantageous Maxwell's L2 cache is when it comes to increasing bandwidth efficiency I've noticed..

 

[HurleyBird]

You people have this funny notion that a GPU is always core limited when a core overclock contributes more than a memory one (and vice versa), and is somehow a "balanced design" when the two contribute equally.

Ignoring the fact that bandwidth utilization varies wildly depending on the usage scenario, such an opinion assumes that the core and memory systems take up equal die space and produce equal heat. If the memory subsystem uses less than half the power and takes up less than half the die space then having the GPU respond equally to core and memory overclocks does not point to a balanced design -- it's actually the exact opposite.

Anandtech forum goers, please stop over-simplifying this aspect of GPU design. You guys are smart enough to realize that there are a lot more factors than just "memory clock scaling" vs. "core clock scaling." When the 380/390X comes out and responds poorly to memory overclocking, actually look at die space and power consumption of the memory system before screaming "This is an imbalanced design, AMD put too much emphasis on the memory system!" from the rooftops.

 

[RussianSensation]

If the memory subsystem uses less than half the power and takes up less than half the die space then having the GPU respond equally to core and memory over.

We were talking about memory bandwidth efficiency that the GPU can utilize, not memory bandwidth from a bandwidth/watt or bandwidth/mm2 die size efficiency. I already addressed this in my posts above. Simply stated, cards like 4890, 7970/7970Ghz had high memory bandwidth on paper but either it was too much for the GPU which itself limited the performance much more or the GPU simply couldn't utilize all of that memory bandwidth due to inefficiencies Tonga addressed. I also realize that some applications or games can be memory bandwidth limited.

I think the key takeaway from this, and what I agree with Carfax83 on, is that we can't apply R9 390X's on paper memory bandwidth "advantage" against GM200 is because there are plenty of examples that prove we can't directly compare on-paper memory bandwidth:

1. HD4890 vs. 5850
2. GTX285 vs. GTX470
3. GTX480/580 vs. 670/680
4. GTX780Ti vs 980
5. HD7950/7970 vs. 290X
6. HD7950/7970 vs. 670/680
7. R9 290/290X vs. 970/980

Whether we compare NV to NV, AMD to AMD or NV to AMD, on paper memory bandwidth can't tell us which part will be faster, unless the application is heavily memory bandwidth limited.

We know that memory bandwidth efficiency matters more. NV's 960 should beat a 7950 after-all.

Perhaps, the extra memory bandwidth will come a lot more handy in professional and compute applications. HBM could also be one of those advanced technologies where the main benefits for early generation GPUs will be power efficiency, latency and PCB complexity reduction, but as GPUs get faster, HBM is NV/AMD's future-proof tech that will ensure future GPU architectures aren't bottlenecked primarily by memory bandwidth.

If I were to make an educated guess if an R9 390X needs 512-640GB/sec memory bandwidth for games, my guess is it doesn't need that much. Why would AMD increase memory bandwidth 70-100%, add Tonga's 40% memory bandwidth efficiency but their chip will only grow 50% in functional units (4096/256 TMU vs. 2816/176)? Seems like AMD's primary motivators here are to save on PCB complexity, latency reduction, power consumption savings on the GPU's memory controller and BW/watt that HBM brings over GDDR5, not necessarily because R9 390X would somehow be bottlenecked with a 400GB/sec memory bandwidth with Tonga's efficiencies.

Side note: I've never owned any GPU where memory overclocking provided me a bigger gain in performance over GPU overclocking. That's why most overclockers and AIBs focus on GPU overclocking a lot more.

 

[Flapdrol1337]

If I were to make an educated guess if an R9 390X needs 512-640GB/sec memory bandwidth for games, my guess is it doesn't need that much.

But it doesn't hurt.

 

[HurleyBird]

I think the key takeaway from this, and what I agree with Carfax83 on, is that we can't apply R9 390X's on paper memory bandwidth "advantage" against GM200 is because there are plenty of examples that prove we can't directly compare on-paper memory bandwidth:

1. HD4890 vs. 5850
2. GTX285 vs. GTX470
3. GTX480/580 vs. 670/680
4. GTX780Ti vs 980
5. HD7950/7970 vs. 290X
6. HD7950/7970 vs. 670/680
7. R9 290/290X vs. 970/980

Here's the issue with that line of argument.

1) Products on new nodes are often pad limited, so are often unable to greatly increase main memory bandwidth. Of course the new parts perform better, but it's insanity to use that as an argument that the older part had too much bandwidth.
2) Cache size typically increases, or at least what is available gets faster and better utilized, with new nodes and architectures.
3) More advanced architectures tend to spend more effort in making sure available bandwidth is used efficiently.

You can spout "memory overclock vs. core overclock" or "last generation product had as much bandwidth as much higher performance next generation product" all you want, but these kinds gross oversimplifications aren't any more useful than complete fabrications.