• Competitor rules

    Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.

Poll: ** The AMD VEGA Thread **

On or off the hype train?

  • (off) Train has derailed

    Votes: 207 39.2%

  • (on) Overcrowding, standing room only

    Votes: 100 18.9%

  • (never ever got on) Chinese escalator

    Votes: 221 41.9%
  • Total voters
    528
Status
Not open for further replies.
zornyan said:
to be honest that doesn't say pro drivers, that says optimised, as in the 16gb memory is actually more optimised for deep learning for example because you need all the vram you can get.

much like how Intel doesn't advertise the titan as a gaming card and advertise is as a deep learning card that 'can' game.
Click to expand...
Okay then optimised drivers in addition to the gaming drivers. So pro specific?
 
Sargatanas2511 said:
Why are you so 100% certain of all of this based on charts you said yourself were unreliable? Why would you expect AMD to bench using CUDA as well?
Click to expand...


Which charts? The vega vs Titan chart is a joke because the correct drivers weren't used to benchmark that kind of software, and its has some serious erros such as getting the half-precision performance of the Titan completely wrong.


The Deep learning benchmark is also completely misleading because it didn't use the correct version of DeepBench for the P100. I work machine learning problems professionals, and sometimes dabble with deep learning so I follow this stuff closely and discussed such flaws in this benchmark elsewhere. A major giveaway of the flaw in AMD"s slides is they give the P100 different results on different slides, so they aren't even internally consistent. Deep learning is strongly influenced by the theoretical half-precision performance, Vega is about 15% faster in this regard. However, Nvidia have massively invested in a harmonious software-hardware solution to maximize performance for HPC and deep learning. CUDA is very mature, and extremely well optimized. CUDA is a low-level API that can extra the best eprformance out of NVidia hardware, a bit Like DX12 is a low level API. OpenCL which is used by AMD is a high level API and is more generic in scope with more over head and less performance. You can look at any number of benchmarks comparing a CUDA optimized HPC/DL system to the OpenCL and you see a ~30% performance improvement using CUDA.

It also helps that CUDA is the de-facto industry standard for this stuff, so a vast majority of software supports a CUDA path and in fact a lot of software will only support CUDA as the GPOU acceleration. I'm nt even sure if TensorFLow has an OpenCL code path yet, at least untul recently it was CUDA or CPU.

And all of this is going that the P100 is far more than just a deep learning accelerator, it is 1:2 FP 64 support while vega is at 1:16. That is a game changer with the HPC world.

Not even AMD are seriously trying to compete with nvidia in that market. What they hope to achieve is picking up a lot of sales form a more budget orientated group, university researchers or prosumers for example. Nvidi can sell their big P100 and V100 for 10,000K a chip and their target audience happily pays that kind of money. AMD isn't a player in that market at this point in time. Vega 20 is supposed to have 1:2 FP64 support and if Navi succeeds then they could be extremely competitive with NVidia' core HPC market.
 
AMD does not have to beat nvidias top dog by a landslide to make me a happy camper. Just 5% within the 1080ti is perfectly fine with me and of course it shouldnt cost more either.. My 3440x1440p monitor needs some power behind it.
 
five8five said:
I was meaning if it was genuine (as some of the results are actually genuine).
Click to expand...

If it was genuine, I would expect all specs to be from a single card as there would be no other logical way to do it. They don't even state "up to 13.1TFlops" so it's just shown as a flat number. I don't see it as unreasonable to expect that's from the card performing as shown.
 
Sargatanas2511 said:
If it was genuine, I would expect all specs to be from a single card as there would be no other logical way to do it. They don't even state "up to 13.1TFlops" so it's just shown as a flat number. I don't see it as unreasonable to expect that's from the card performing as shown.
Click to expand...
AMD have actually quoted approx. 13 TFLOPS. I'd assume that was the best case.
 
five8five said:
True, but can it be dismissed out of hand. Would a retailer fake specs? Isn't it more likely that it's real (including errors)?
Click to expand...

Happens all the time - they just want to get your pre-order money by being the first people taking orders. They don't have stock, prices, or specifications, but they have to put something on the webpage, so they grab whatever rumour or click-bait info is around and use that.
 
Status
Not open for further replies.
Back
Top Bottom