Gamers Nexus: HBM2 Costs Estimated... Is AMD Profiting much on Vega?

[cj09beira]

1 minute ago, DocSwag said:

Something that can't be denied though is Pascal as higher perf/mm^2 and can clock higher than Vega.

 

I don't doubt that if you remove all limitations Pascal could get high in power draw but compared to Vega it's superior.

 

I thought that removing the voltage limit at least could be achieved just by soldering the vsense pin on the vrm controller to ground, it's not that bad?... Or is there more to it if you want to get voltage control?

the problem is that there are parts in the die it self changing the voltage, and limiting the gpu many many times per second, 

[DocSwag]

1 minute ago, cj09beira said:

the problem is that there are parts in the die it self changing the voltage, and limiting the gpu many many times per second, 

Isn't pretty much all of the voltage work done by the vrm controller though? Correct me if wrong...

[cj09beira]

Just now, DocSwag said:

Isn't pretty much all of the voltage work done by the vrm controller though? Correct me if wrong...

we dont know how they do it, but there is something being done in the gpu itself, for example if you raise voltage too much performance gets worse and worse,

buildzoid talked about this in one of his livestreams when he was preparing to take a 1070 into ln2

[Dylanc1500]

5 minutes ago, cj09beira said:

imagine ryzen done in intel's labs hmm delicious 

I was thinking about Pascal on Intel's current 14nm node. I mean Intel could make some serious cash if they leased their fabs.

[DocSwag]

1 minute ago, cj09beira said:

we dont know how they do it, but there is something being done in the gpu itself, for example if you raise voltage too much performance gets worse and worse,

buildzoid talked about this in one of his livestreams when he was preparing to take a 1070 into ln2

Ah, now I know what you're talkina about, I remember watching that.

 

So basically then, you can mod Pascal a bit, but not that much.

[cj09beira]

Just now, DocSwag said:

Ah, now I know what you're talkina about, I remember watching that.

 

So basically then, you can mod Pascal a bit, but not that much.

ya, but if i remember correctly he was able to disable more of it i just don't remember how

[cj09beira]

3 minutes ago, Dylanc1500 said:

I was thinking about Pascal on Intel's current 14nm node. I mean Intel could make some serious cash if they leased their fabs.

intel would never help nvidea, they are too much of a treat to their server monopoly (and starting now so is amd )

[Drak3]

30 minutes ago, cj09beira said:

they wont go back to gddr, its less efficient, needs more die space on gpu to control it,

amd is trying to increase volume of hbm so that it becomes cheaper 

As it stands, AMD has no GOOD reason to exclusively use HBM for NAVI. If they manage to get HBM production up, NAVI's successor can become an HBM exclusive stack without everyonr feeling a bit of pain.

[Dylanc1500]

Just now, DocSwag said:

Ah, now I know what you're talkina about, I remember watching that.

 

So basically then, you can mod Pascal a bit, but not that much.

It is built within the GPU itself, Volta is the same way as can actually intercommunicate with other GPUs to set a max threshold with power draw for each GPU and it will govern itself to that set TDP target.

[ltguy]

I think this video while very informative misses the elephant in the room. If vega, compared to the rx580/480, uses so much power, is so memory bandwidth starved, and isnt really hitting the performance numbers they are looking for, then why are they using vega to begin with. Sounds like vega is the problem. RX580 had good performance, good power usage, could use gddr5, sounds like thats a better starting point than vega. What is the benefit of vega to amd? sounds like it required more fixes and work arounds then benefits gained.

[Dylanc1500]

2 minutes ago, cj09beira said:

intel would never help nvidea, they are too much of a treat to their server monopoly (and starting now so is amd )

The only thing they are competing with is the Xeon Phi. All of those Tesla systems are using Xeons or Power chips.

[XenosTech]

14 minutes ago, cj09beira said:

imagine ryzen done in intel's labs hmm delicious 

lol that would be a tiny bit insane

14 minutes ago, DocSwag said:

I think it wouldn't be as good as Pascal still but it'd be somewhat better. As demonstrated by the 1050/1050 ti glofo/Samsung 14nm does not clock as high as tsmc, tsmc clocks about 10% ish higher.

Probably won't be as good as pascal or the golden nugget that is polaris but I could imaging what that would do for the power consumption since the chips probably won't need as much voltage pumped into them to get their desired clocks then we end up with chips that are basically tapped out from AMD from the get go

[Agosto]

24 minutes ago, DocSwag said:

I think it wouldn't be as good as Pascal still but it'd be somewhat better. As demonstrated by the 1050/1050 ti glofo/Samsung 14nm does not clock as high as tsmc, tsmc clocks about 10% ish higher.

1050/Ti also have higher load stock voltage than every other pascal card (1.1V vs 1.06V) and higher idle voltage/clock

[DocSwag]

21 minutes ago, cj09beira said:

intel would never help nvidea, they are too much of a treat to their server monopoly (and starting now so is amd )

I don't think it's as much server (that's more of power being the threat) as much as hpc ATM.

 

Nvidia is doing really well when it comes to hpc and deep learning and whatnot, enough so to threaten Intel.

[System Error Message]

AMD's vega is also good at AI too. Just look at fp 16 and int 8 performance as well. pascal is good at AI because of int 8.

 

Its not that nvidia is a threat to intel, intel also wants to jump on the AI bandwagon. However intel's IGP if made into an actualy GPU can be a contender in the consumer compute market. It may not be the fastest at graphics but as far as compute goes it beats nvidia and AMD hands down shader and core.

[leadeater]

2 hours ago, System Error Message said:

AMD's vega is also good at AI too. Just look at fp 16 and int 8 performance as well. pascal is good at AI because of int 8.

 

Its not that nvidia is a threat to intel, intel also wants to jump on the AI bandwagon. However intel's IGP if made into an actualy GPU can be a contender in the consumer compute market. It may not be the fastest at graphics but as far as compute goes it beats nvidia and AMD hands down shader and core.

What? Intel IGP's are very low performance across the board. If you're going Intel for HPC that is Xeon Phi.

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/17

http://www.sisoftware.eu/2017/04/12/intel-graphics-gpgpu-performance/

 

The above tests don't show stellar performance at all and for Anandtech's article it's comparing against mobile GPUs which is the first sign that performance is not going to be great.

 

Be rather wary of Intel's performance claims when it comes to HPC, Nvidia pointed out how there benchmarks are BS multiple times.

https://blogs.nvidia.com/blog/2016/08/16/correcting-some-mistakes/

http://www.nvidia.com/object/justthefacts.html

[cj09beira]

4 hours ago, DocSwag said:

I don't think it's as much server (that's more of power being the threat) as much as hpc ATM.

 

Nvidia is doing really well when it comes to hpc and deep learning and whatnot, enough so to threaten Intel.

went i said server i meant server space, including ai, hpc etc

[Dabombinable]

Still finding it funny that AMD's divisions have done a reversal. Now its the graphics cards which don't compete well at the high end (excluding compute).....

[System Error Message]

2 hours ago, leadeater said:

What? Intel IGP's are very low performance across the board. If you're going Intel for HPC that is Xeon Phi.

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/17

http://www.sisoftware.eu/2017/04/12/intel-graphics-gpgpu-performance/

 

The above tests don't show stellar performance at all and for Anandtech's article it's comparing against mobile GPUs which is the first sign that performance is not going to be great.

 

Be rather wary of Intel's performance claims when it comes to HPC, Nvidia pointed out how there benchmarks are BS multiple times.

https://blogs.nvidia.com/blog/2016/08/16/correcting-some-mistakes/

http://www.nvidia.com/object/justthefacts.html

give your intel IGP a workload like luxmark. Compare the benchmark score to a GPU from nvidia and AMD and also compare the cores, shaders and clocks. Clearly intel has the IPC advantage here, each of their GPU core simply does more than an nvidia or AMD GPU core given the same clocks and same core count. If intel ever made a new dedicated GPU not just the xeon phi that already exists it could be really fast especially when paired with lots of memory bandwidth.

 

A cluster of CPUs with iris pro 580 is also decent too.

[leadeater]

15 minutes ago, System Error Message said:

If intel ever made a new dedicated GPU not just the xeon phi that already exists it could be really fast especially when paired with lots of memory bandwidth.

They did, it sucked. You can't always scale something up and the IGP has a slight advantage of being on package with the CPU so has much lower latency than a GPU on the PCIe bus, even when the lanes are directly on the CPU and not the chipset.

 

The architecture design is also totally different, it's not as simple as going Iris Pro has 40 core and scores x so if it were x10 that it'll score y.

 

Quote

The Crystalwell die measures 7mm x 12mm (84mm^2), while the quad-core Haswell + GT3 die is a whopping 264mm^2 (16.2mm x 16.3mm). Working backwards from the official data Intel provided (177mm^2 for quad-core GT2), I came up with an 87mm^2 adder for the extra hardware in Haswell GT3 vs. GT2. Doubling that 87mm^2 we get a rough idea of how big the full 40 EU Haswell GPU might be: 174mm^2. If my math is right, this means that in a quad-core Haswell GT3 die, around 65% of the die area is GPU. This is contrary to the ~33% in a quad-core Haswell GT2. I suspect a dual-core + GT3 design is at least half GPU.

http://www.anandtech.com/show/6993/intel-iris-pro-5200-graphics-review-core-i74950hq-tested/4

 

Performance per mm2 is lower than that of Nvidia and AMD for IGP vs GPU, Intel's scores as just bigger so 40 can do more but it doesn't make them better.

[System Error Message]

Just now, leadeater said:

 

The architecture design is also totally different, it's not as simple as going Iris Pro has 40 core and scores x so if it were x10 that it'll score y.

 

actually they have 72 with the iris pro 580.

[leadeater]

3 minutes ago, System Error Message said:

actually they have 72 with the iris pro 580.

Yea, had to go with what as available data/review wise. Haven't seen any done after GT3e. There's game benchmarks of GT4e just not compute as far as I've found, not good ones by independent reviewers that is.

 

GT4e is around 220mm2 of die space delivering 1.152 TFLOP of performance for compute tasks, an Nvidia 1080Ti is 471mm2 delivering 11.3 TFLOP of performance for compute tasks. The GTX 750 is roughly equivalent compute performance wise and is 148mm2. 

 

Iris Pro/ Intel IGP has certainly gotten better and gives very respectable performance but they aren't high performance designs.

 

User based results so no idea how trustworthy they are.

Spoiler

https://www.notebookcheck.net/Intel-Iris-Pro-Graphics-580.160664.0.html

 

[System Error Message]

40 minutes ago, leadeater said:

Yea, had to go with what as available data/review wise. Haven't seen any done after GT3e. There's game benchmarks of GT4e just not compute as far as I've found, not good ones by independent reviewers that is.

 

GT4e is around 220mm2 of die space delivering 1.152 TFLOP of performance for compute tasks, an Nvidia 1080Ti is 471mm2 delivering 11.3 TFLOP of performance for compute tasks. The GTX 750 is roughly equivalent compute performance wise and is 148mm2. 

 

Iris Pro/ Intel IGP has certainly gotten better and gives very respectable performance but they aren't high performance designs.

 

User based results so no idea how trustworthy they are.

  Hide contents

 

https://www.notebookcheck.net/Intel-Iris-Pro-Graphics-580.160664.0.html

 

ive runned luxmark on it, the numbers are accurate that for compute the 1080 ti is less than 10x faster but if it had 720 shaders it would have the same flops as the 1080 ti which has around 3500 shaders. i heard intel's graphics is based on their own x86 having the same core units but different pipeline, decoders and so on. is the 220mm2 just the graphics alone or including the cpu?

 

it can scale up, i think memory bandwidth is important when scaling up.

[leadeater]

2 hours ago, System Error Message said:

is the 220mm2 just the graphics alone or including the cpu?

Just the IGP, it's around 60%-70% of the total die space according the the review I was reading.

 

At the current performance density Intel would be hitting die size limits before achieving compute performance of a 1080Ti or Vega. As compute SDKs get more and more general and the industry gets more standardized I do expect Intel to get more competitive though. GPUs are getting less and less pure rendering engines and are just becoming very powerful math processors.

 

Volta is also signaling a turning point as well with the inclusion of Tensor cores. Depending on how things turn out in the next say 5 years Tesla cards might not actually have a GPU in them at all.

[AshleyAshes]

7 minutes ago, leadeater said:

GPUs are getting less and less pure rendering engines and are just becoming very powerful math processors.

The irony there is that that's basically a return to the start of GPUs from the 70's through 90's.  Obviously a bajillion times more powerful but they all started as not much more than math co-processors that highly specialized in floating point math.