Jump to content

Khvarrioiren

Member
  • Posts

    577
  • Joined

  • Last visited

Reputation Activity

  1. Informative
    Khvarrioiren reacted to patrickjp93 in 3D Mark Time Spy benchmark is not using async compute as we know it - new report suggests   
    What people have misunderstood about this benchmark is that it tests one specific way of doing asynchronous compute: pre-empting tasks that are not time-sensitive with ones that are. This is incredibly important in VR for a technique called "time warp" that has so many utilitarian uses that I'd rather send you out to google "VR Time Warp pre-emption" than fill this post about it.
     
    Second, this benchmark does also test parallelization across CUs. It just doesn't test interweaving tasks within a single Compute Unit/SMM. That is what AMD has been having developers use, specifically because it cripples Nvidia's performance. That doesn't make Nvidia's implementation wrong or this test invalid. You just have to know what it tests and understand how to interpret the data.
     
    I'd say the value in this test is the effect on VR. If we want to test gains vs. gains using different techniques, it would require a new benchmark.
     
    @-BirdiE- I could write one I suppose. I think it would take about 1200 lines, but it wouldn't look like anything. One thing that has annoyed me about graphics benchmarks is that they say they measure specific aspects of this API or that (Except for Firestrike which just is a super advanced, comprehensive graphics test indicative of real game performance in a way actual benchmarks aren't), but they don't. They actually render things and go through the whole pipeline instead of testing the aspect they say they do. Call me old fashioned, but if you see anything change on the screen, the benchmark is doing something other than gauging the time it takes to complete the task it's supposedly targeting,
     
    For example, if I wanted to test how long a computer took to calculate the 1 billionth digit of PI, and I printed each digit to the screen as it was calculated, I'd be spending thousands of instructions and cycles for each print which would drastically affect timing. Actually rendering something does exactly this. If I wanted to test the four primary aspects of AC/S I would rig up code that shoved a ton of triangles onto the rendering space and then begin timing as soon as I asked for shading and positioning to be queued up using n queues. I could test performance scaling with queue number and instill a limit of 128 (doubling until a performance loss occurred and then using binary search to find the optimal point between 2^n and 2^(n-1), and then I could run the same test under different modes where I could time context switches, gauge total frame count for 30 seconds using parallel CUs, gauge total frame count for 30 seconds using interleaved tasks, and then the same for parallel CUs and interleaved tasks.
     
    It's not like getting pure tests is difficult, but I've never understood why these benchmarks go through the whole pipeline. It just distorts the benchmark. As a rule in HPC: if 99% of your time isn't spent on doing what you're measuring, then your benchmark isn't valid.
  2. Agree
    Khvarrioiren reacted to patrickjp93 in 3D Mark Time Spy benchmark is not using async compute as we know it - new report suggests   
    That speculation isn't worth mentioning until someone proves having less parallelization is somehow crippling one or the other. A benchmark, to be canonical about it, should present a few different pieces of information: how well does a processor handle a certain operation, and how is that operation best used on a given processor. As long as 3DMark is not crippling either one, I couldn't give a flying rat's ass that it uses less parallelization than Doom. Has anyone proved Doom's level is totally optimal? No.
  3. Like
    Khvarrioiren got a reaction from 7850OC in Enter the RX 480 AIB Cards   
    I got three Sapphire Nitro+'s on order from a Finnish supplier, shipping date 20.07.2016
  4. Agree
    Khvarrioiren reacted to zMeul in FutureMark quietly releases DX12 benchmark, TimeSpy - avalable trough STEAM right now   
    thing is I paid like 9EU for the Advanced version and they want me to pony up 5 for DLC for DLC  ... nope! 
  5. Informative
    Khvarrioiren reacted to patrickjp93 in PCPer Tests 3DMark Timespy: Pascal Does have Async Compute   
    For AMD it is, but Nvidia gains what, 10% at most in Doom and in 3DMark. That means there's room for the driver to be able to optimize the pipeline more.
     
    What AMD has done is essentially hyper threading for CPU core clusters (CUs for AMD, SMMs for Nvidia). All Hyperthreading does is interleave two threads, two lines of execution, across one core. The only time you gain performance from that is if all of the following are true.
     
    1) There are unused cycles in at least one of the n pipelines (SMT can be up to 8 way in Power 8 and Sparc M7 for example)
     
    2) The n threads don't use the same resources at the same time.
     
    3) The hyper threading engine doesn't magically line up the threads to violate #2 or have their unused cycles line up, still leaving holes in the pipeline.
     
    There is a big pitfall in this if two threads don't quite line up optimally. Solution? Redo both threads as a single thread at the instruction level and don't drop cycles. It's difficult, but it's the level Facebook works at to squeeze out more performance these days. Nvidia simply does this same thing via drivers to make its GPU pipeline remain as full as possible for as long as possible in as many cases as possible.
     
    AMD's hardware scheduler does this same thing, but at the hardware level instead of software. AMD's gains via asynchronous-compute-enabled software are larger because its pipelines aren't well-fed via its driver. Yet, its driver overhead is still higher than Nvidia's. By proof of contradiction, I've now done a closed logical proof that Asynchronous Compute is currently unnecessary for Nvidia, as Nvidia has the superior performance using less hardware while still having lower CPU overhead.
     
    In conclusion, games do not suddenly need 8 pipelines and 64 queues. We've made it this far on 1 and 1. And clearly Nvidia's driver team has proven it can keep going a while more.
  6. Funny
    Khvarrioiren reacted to themaniac in PCPer Tests 3DMark Timespy: Pascal Does have Async Compute   
    damn, ran time spy and my result is 
  7. Agree
    Khvarrioiren reacted to Briggsy in PCPer Tests 3DMark Timespy: Pascal Does have Async Compute   
    That could have had something to do with the xbox 360 using dx9 at the time, I recall a number of PC ports being dx9 with a dx11 option tacked on for tessellation and other effects. For some reason I recall The Secret World as being one of the first PC games to have both options, but it wasn't even a console port.... so there goes that reasoning out the window. It definitely did happen for a year or two with a number of AAA titles. I recall when Skyrim came out, there was petitions, uproars etc. over the developers not wanting to add a DX11 option (with good reason, the game was in development forever and was a console port) because it seemed every other game at that time was doing the DX11 option.
  8. Informative
    Khvarrioiren reacted to patrickjp93 in 3D Mark Time Spy benchmark is not using async compute as we know it - new report suggests   
    The ACEs are hardware schedulers, each capable of controlling a few CUs. That said, each CU can interleave tasks. Nvidia's SMMs cannot do this.
     
    Fermi was the last with a hardware scheduler.
  9. Informative
    Khvarrioiren reacted to patrickjp93 in 3D Mark Time Spy benchmark is not using async compute as we know it - new report suggests   
    That's why Nvidia invests so much code (iirc 50 million lines as of 2015) into its drivers. Nvidia has its drivers optimize (and FIX) shader code on the fly. Some of that is done by deterministic optimizers, and some of it is done by swapping out binary blobs for blobs Nvidia engineers rewrote in-house (without giving them to the studios b/c defend your IP at all costs yada yada). That's also why Nvidia's gains under asynchronous compute for Pascal are smaller than they are for AMD. Nvidia's pipelines are already much better fed out of the box because of the driver investment.
  10. Agree
    Khvarrioiren reacted to zMeul in 3D Mark Time Spy benchmark is not using async compute as we know it - new report suggests   
    nVidia's guidielines on DX12 has a similar line
    it's a heavy switch that shouldn't be utilized too often: https://developer.nvidia.com/dx12-dos-and-donts
     
  11. Informative
    Khvarrioiren got a reaction from Shakaza in Someone Got Their GTX 1060 Early   
    It seems that some AIB RX480's can achieve upwards of 1500 Mhz clock speeds, but that's a big if and a statement best taken with salt. We'll just have to wait and see.
  12. Like
    Khvarrioiren got a reaction from Shakaza in Someone Got Their GTX 1060 Early   
    Same here. Mostly just waiting for either to come out, since I've got customer builds piling up waiting for the new mid-range GPUs.
  13. Like
    Khvarrioiren reacted to Tech Inquisition in Vulkan API for DOOM is here   
    I'm having a weird deja-vu feeling of seeing a troll named Zenbolt on WCCF that was also "DBZ" obsessed... hope he didn't find his way here as Sammael...
  14. Funny
    Khvarrioiren got a reaction from AlTech in RX 480 performs ~30% faster with Vulkan!...   
    There's now as much Vulkan releases in 2016 as there are DX11 releases :v
  15. Agree
    Khvarrioiren reacted to Prysin in RX 480 performs ~30% faster with Vulkan!...   
    Summing up the above posters opinions in four lines of text:
    Nvidia does poorly, AMD does good -> must be poorly optimized
    Nvidia scores good. AMD sucks -> Game is properly optimized, AMD just sucks.
    AMD does good, Nvidia does good -> Game is poorly optimized for GPUs.
  16. Agree
    Khvarrioiren reacted to iamdarkyoshi in NVIDIA GeForce GTX 1060 Reviewer’s Guide Leaked   
    The images. Wtf?
  17. Agree
    Khvarrioiren reacted to 19_blackie_73 in NVIDIA GeForce GTX 1060 Reviewer’s Guide Leaked   
    I wouldn't take the 480 results seriously, but i think the 1060 numbers could be actual reality
  18. Agree
    Khvarrioiren reacted to Energycore in NVIDIA GeForce GTX 1060 Reviewer’s Guide Leaked   
    So Kyle wasn't joking then
  19. Agree
    Khvarrioiren got a reaction from Energycore in NVIDIA GeForce GTX 1060 Reviewer’s Guide Leaked   
    Damn, that was really well done.
  20. Agree
    Khvarrioiren reacted to FatalityDiablo in NVIDIA GeForce GTX 1060 Reviewer’s Guide Leaked   
    So gtx980 performance was just a filthy rumor after all..
    edit: I like how their test rx480 performs like a potato compared to consumer benchmarks ^^
  21. Agree
    Khvarrioiren reacted to Marinatall_Ironside in Vulkan API for DOOM is here   
    Yeah probably...
     
    But still, my point still stands that he is an elitist and making people who bought an Nvidia GPU (like me) feel like a pile of shit by making those rhetoric claims without also considering the other facts at play as well (like electricity costs and PPW metrics). Saying x company's product(s) is better than y company(s) product without any statistics and relying on experience is, indeed, elitism. He was also using the Appeal to Pity fallacy by saying that because AMD is doing terribly business-wise (which is true), their products are god-like and leaves Nvidia and Intel in the dust. Sorry, I won't buy into that claim straight-up because from what I've seen in the performance metrics from the dozens of professional reviewers and publications, AMD isn't even near yet. But I agree that they've made some damn good progress so far. I would love to see AMD compete again and I would probably go AMD full-time, if they're products kicks ass that is
  22. Informative
    Khvarrioiren got a reaction from AlTech in RX 480 performs ~30% faster with Vulkan!...   
    30-40% seems to be the ballpark.
  23. Agree
    Khvarrioiren reacted to patrickjp93 in Future of Radeon   
    Please go learn microeconomics and how monopolistic competition and monopoly markets work. No one can just set prices in the sky. There's an optimal point on the supply and demand curves where they cross and maximum total profit may be achieved. All lack of competition does is shift the demand curve.
     
    And as for macroeconomics and behavioral economics, most people won't wait more than a month once the Titan P and 1080TI are released. If Vega is late, it won't matter what price it's at. The sales window will have closed.
  24. Agree
    Khvarrioiren got a reaction from AlTech in AMD Lists The Radeon RX 490 Flagship – Polaris based Dual GPU Graphics Card For 4K Ready Gaming   
    I'm very doubtful that it'd be a dual GPU card. They might have kept something up their sleeve for reasons unknown, but a dual Polaris card just seems absolutely idiotic.
  25. Like
    Khvarrioiren got a reaction from Dabombinable in AMD Lists The Radeon RX 490 Flagship – Polaris based Dual GPU Graphics Card For 4K Ready Gaming   
    About 200-300€ more. I'm kinda mad that he ordered the build before I started doing the client builds on my own. Would've made some nice bank instead of the shitty 16€/hr.
×