Why is AMD losing the next gen race to Nvidia?

[silent_guy]

With the same amount of SMs and way more logic, I don't think a GP100 can be made much faster than a GP102 for gaming workloads, assuming the same power budget.

We've already seen a P100 PCIe version spec'ed at 250W, with much lower clocks at 1.3GHz.

There's no way you can compensate a 15% lower core clock with a 50% increase in BW.

 

[spworley]

I don't think a GP100 can be made much faster than a GP102 for gaming workloads, assuming the same power budget.

I have long suspected the P100 doesn't even have the rasterizers needed for graphics and it's truly compute-only. The saved die space was used for those extra P100 features (fp16x2, flexible preemption, system wide memory addressing) and especially for the doubled number of registers per SM. NVIDIA is making a PCIE P100 Tesla, but if there was such graphics functionality on the P100 die it wouldn't NVIDIA also make a P7000 Quadro version too? Quadros sell for an even higher markup than Tesla boards, and all the PCIE PCB and cooler design work is already done. There's also precedent for a compute only chip: GK210 (used in the K80 Tesla) was a doubled-register chip and was never used for any graphics card, GeForce or Quadro.
The significant objection to this theory is that the P100 whitepaper does specifially mention graphics in its discussion of preemption.

 

[Ryan Smith]

I have long suspected the P100 doesn't even have the rasterizers needed for graphics and it's truly compute-only.

GP100 is fully graphics capable. I asked NVIDIA about that the moment I saw P100, and they were very clear in their answer.

 

[CarstenS]

Ethereum Mining is definitely something really ... irregular. Polaris seems really really good at it in it's current state. With some little tweaking, I'm getting the same or even higher hash rates out of a 480 as a R9 Nano was able to deliver (no, not power constrained). I'm leaning toward the memory and latency issue here as well (Fiji vs. Polaris, 1080 vs. 1070).

Someone on a mining forum suggested it had to do with the DAG-file size, which, after a certain size (IIRC ~2 GiB) was passed, caused the fall off of a performance cliff for Nvidia cards (references were made to TLB sizes???). Some of the more recent reports told success stories from new mining client updates combined with newer drivers (WDM 2.0 was mentioned) and the W10 anniversary update.

edit: Seemingly, something in W10 Anniv. Update really fixed it. Hash rate on a Geforce through the roof (compared to before, still not as many MH/TFLOP as on a Polaris). Need to re-check with a 1070 -> GDDR5

edit2: Fun fact: Access stride seems to play an enormous role here at least compared to TFLOPS or raw bandwidth. Results with Claymore 6.4, ETH SOLO.
1060/G5@8.8 GT/s -> ~22 MH/s
1070/G5 @8.8 GT/s -> ~31 MH/s
1080/G5X@ 11 GT/s -> ~26 MH/s
Titan X/G5X @11 GT/s -> ~38 MH/s
Not double checked, just from quick glance, so take with a healthy dose of salt.

FWIW, before the anniversary update, the Titan X was at 1.7 MH/s. Yep, one-point-seven.

 

[homerdog]

So AMD 'bigger' GPU Vega is 1H17, not even 1Q17, that is so disappointing....not even the big one with HBM..
Can we say they have lost this round already?

Just what is wrong with their 14nm GCN3.0? The process or the architecture? Who wants to put in some speculation with Polaris as a clue..?

This round is definitely over and is playing out even worse than Maxwell vs Rebrandeon. AMD's new cards barely hold up to Maxwell in perf/W! That's a 2 node jump..!! P10 shouldn't even breathe the same air as GM204 but hell GM204 is faster!!! It's like NVIDIA FX vs Radeon 9000 series in reverse.

 

[Deleted member 13524]

This round is definitely over and is playing out even worse than Maxwell vs Rebrandeon.

Is this solely based on performance/watt on DX11 titles for desktop graphics cards or do you have future sales data to back it up?

 

[function]

With AMD paying GF large sums of money to not use their fabs - presumably because the grass is significantly greener elsewhere - couldn't they have put (some of) that money into using larger dies and increased performance and lowered power that way?

A slightly larger Polaris 10 with more ROPs could have surpassed the 390 and the 1060, while offering more flexibility in terms of redundancy and a greater chance to differentiate between the 'rather close' 470 and 480.

Hindsight is a wonderful thing, but even without knowing where the 1060 would land, I can't help feeling that having the 480 come in 5% behind the old-as-heck 390 was not ideal targeting.

 

[homerdog]

Is this solely based on performance/watt on DX11 titles for desktop graphics cards or do you have future sales data to back it up?

Not just DX11 but yea current titles, which is what people buy graphics cards for. I have no insider knowledge or anything like that, but it seems obvious enough to me that P10 and P11 don't change much. R480/470 are good cards in their own right but they don't change AMD's positioning against NVIDIA from what it was last gen. Maybe Vega will be different? At least I hope so or GP102 will remain forever out of reach. But even if Vega is competitive we're back in the position of NV not needing HBM to compete with AMD, so AMD margins will suffer accordingly in the best case scenario, unless you think Big Vega will be competitive with superbeast GP100

 

[liolio]

Hindsight is a wonderful thing, but even without knowing where the 1060 would land, I can't help feeling that having the 480 come in 5% behind the old-as-heck 390 was not ideal targeting.

As the dust settles and the 1050 is to land it is even truer for the Rx460, GP107 (1050?) should kill it.

 

[Razor1]

Not just DX11 but yea current titles, which is what people buy graphics cards for. I have no insider knowledge or anything like that, but it seems obvious enough to me that P10 and P11 don't change much. R480/470 are good cards in their own right but they don't change AMD's positioning against NVIDIA from what it was last gen. Maybe Vega will be different? At least I hope so or GP102 will remain forever out of reach. But even if Vega is competitive we're back in the position of NV not needing HBM to compete with AMD, so AMD margins will suffer accordingly in the best case scenario, unless you think Big Vega will be competitive with superbeast GP100

The P10 and P11 didn't change the dynamics of the perf/watt to Pascal, from r9 and Fiji to Maxwell 2, but actually its worse.

 

[Anarchist4000]

Considering the recent change to the agreement with GF, I'd have to imagine there is a process problem. The amendment was literally a payout close to what I'd expect GF's profits to be over the course of several years. That being the case, 480 and 1060 might not be that far apart on perf/watt. Polaris should have had a mechanism to reduce power usage, but it seems like they just maxed the voltages to get a reliable supply. Also a bit surprised there isn't a dual GPU floating around. One of the partners even designed a cooler for one and I can't imagine dual Fiji was the target. It's possible the lack of supply we're seeing is a shift to Samsung, as the process IP should be identical, but who knows?

 

[Razor1]

Considering the recent change to the agreement with GF, I'd have to imagine there is a process problem. The amendment was literally a payout close to what I'd expect GF's profits to be over the course of several years. That being the case, 480 and 1060 might not be that far apart on perf/watt. Polaris should have had a mechanism to reduce power usage, but it seems like they just maxed the voltages to get a reliable supply. Also a bit surprised there isn't a dual GPU floating around. One of the partners even designed a cooler for one and I can't imagine dual Fiji was the target. It's possible the lack of supply we're seeing is a shift to Samsung, as the process IP should be identical, but who knows?

Well to match the 1060, they would need to address a 20%+ or - depending on the vram disadvantage (we don't have exact figures for the GPU alone so its somewhere in that neighborhood), even with a process problem, that is a lot to address. And we aren't even talking about the best perf/watt chip in the Pascal nor have we see what the performance cards would have been like from AMD, or will be like.

 

[jacozz]

Maybe Vega will be different? At least I hope so or GP102 will remain forever out of reach.

This statement is why AMD don't have a real chance to compete. A lot of folks only hope for competition so that the can buy Nvidia cards for cheap. If Vega is good, why not just buy one?

 

[Razor1]

This statement is why AMD don't have a real chance to compete. A lot of folks only hope for competition so that the can buy Nvidia cards for cheap. If Vega is good, why not just buy one?

well the contention is getting to pascal's performance level seems unlike the hope is to at least get to them, which will help competition, getting past it, is more than hope, that would be wishful thinking or praying at this point......

 

[Entropy]

well the contention is getting to pascal's performance level seems unlike the hope is to at least get to them, which will help competition, getting past it, is more than hope, that would be wishful thinking or praying at this point......

As I see it, there are a couple of problems with with this assertion.
The major one is that it is based on performance as the only metric for competitiveness. Even at the higher end of the performance scale, prospectove customers may judge competitiveness with somewhat more complex viewpoints.
Another is the assumption that AMD will use the same process, power/performance and architecture as they did for Polaris, none of which is certain. At this point, we just don't know.
Predictions are pretty useless beyond the trivial one - AMD will not release a product positioned not to sell. It will be competitive somehow, even if just in price/performance. And that is actually enough in many respects.

 

[jacozz]

Here's a thought.
If I was Lisa su I would:

• Ditch the traditional desktop market.
It´s seems to be a hopeless case anyway. It really doesn't matter if Zen, or Vega is up to snuff, (or at least close). Market will still be with Intel/Nvidia.

• Instead, do what the competition can't do.
Smack a Zen + Vega on a interposer with plenty of HBM. Skip the rest. No need for ATX, MicroATX or even MiniITX. Instead you have the smallest high performance package on earth. Now that would be sexy and be a lot easier to sell.

 

[Razor1]

As I see it, there are a couple of problems with with this assertion.
The major one is that it is based on performance as the only metric for competitiveness. Even at the higher end of the performance scale, prospectove customers may judge competitiveness with somewhat more complex viewpoints.
Another is the assumption that AMD will use the same process, power/performance and architecture as they did for Polaris, none of which is certain. At this point, we just don't know.
Predictions are pretty useless beyond the trivial one - AMD will not release a product positioned not to sell. It will be competitive somehow, even if just in price/performance. And that is actually enough in many respects.

Can't predict from the unknown though, we don't know if its a process problem that prevented Polaris from being competitive, and looking at a 20% deficient it can't be process alone. We don't know the IP changes entail for Vega either. Looking at it from the outside, though, the changes necessary for AMD to take to catch up to nV's Pascal, there doesn't seem to be enough time for them to do such things. Polaris showed us that in spades, another 6 months going to make that much of a difference? I think not. The changes nV made to Maxwell 2 from Keplar are things they learned form a few years of work with Tegra, those are things AMD has to learn too, and after selling off their mobile division are things they have to reinvest in, while we saw a decrease in R&D things like that most likely didn't make it into their portfolio again, specially since I think it was 2 Q's ago, AMD stated in their financial call, they want to reinvest in mobile technology, so it kinda means they haven't focused on those things at the point, which goes back to the time needed to further advancements......

Also added to this, I'm not talking about performance alone, all metrics it looks like AMD is behind, perf/watt, power consumption, perf/mm, perf/transistor, at least for gaming. But all of these are inter related.

 

[Alexko]

Here's a thought.
If I was Lisa su I would:

• Ditch the traditional desktop market.
It´s seems to be a hopeless case anyway. It really doesn't matter if Zen, or Vega is up to snuff, (or at least close). Market will still be with Intel/Nvidia.

• Instead, do what the competition can't do.
Smack a Zen + Vega on a interposer with plenty of HBM. Skip the rest. No need for ATX, MicroATX or even MiniITX. Instead you have the smallest high performance package on earth. Now that would be sexy and be a lot easier to sell.

AMD is doing just that, it's their HPC APU. It might do well, or it might not. Betting the company on this would be insane.

 

[eastmen]

Here's a thought.
If I was Lisa su I would:

• Ditch the traditional desktop market.
It´s seems to be a hopeless case anyway. It really doesn't matter if Zen, or Vega is up to snuff, (or at least close). Market will still be with Intel/Nvidia.

• Instead, do what the competition can't do.
Smack a Zen + Vega on a interposer with plenty of HBM. Skip the rest. No need for ATX, MicroATX or even MiniITX. Instead you have the smallest high performance package on earth. Now that would be sexy and be a lot easier to sell.

The server market is still growing and big. Making chips for the server market and you already have what you need for traditional desktop / laptop market. So it makes sense for them to continue to try.

 

[Ethatron]

In the end a company needs to do what it wants, and only bother with the competition, if they inhibit that possibility. Game producing companies do this all the time, some also deconstruct after, because the people involved want to do somethings else.
AMD is doing that to some degree, but also sucked into the marketing-battle with Nvidia, which is a big problem because they can not focus on what they want to focus on. They fight to stay at least where they are, because sadly, much smaller doesn't work in this type of market.
If you think about it, ARM found a very cool way to be relevant, but entirely disconnected from the actual 'realization' aspects, by largely only doing the prints, but not the silicon.