Gadgetory

and these high bandwidth cache controller protocol is one of the Keystone's to the Vega architecture marked by RTG lead Raja Kaduri as a personal favorite feature of Vega and highlighted in previous marketing materials as offering a potential 50% performance uplift in average FPS when in vram constrained scenarios and you can see up to 50 ports on average FPS improvement and actually more importantly almost 2 X min FPS with a few driver revisions now behind us we're revisiting our Vega 56 hybrid card to benchmark HPCC in a B fashion testing and memory constraint scenarios to determine advocacy and real gaming workloads before we get into that this content is brought to you by the Thermaltake flow RGB closed-loop liquid cooler which is a 360 millimetre radiator plus 3 120 fans that are RGB illuminated the if-then we'll take it rain fans at that this is a 4.5 gen edge detect pump which is one of the faster pumps you can learn more at the link in the description below HPCC is the controller for Andy's high bandwidth cache what the company has functionally renamed its vram 2 so there's no hard threshold as to what governs the high bandwidth cache naming you could have for example a Vega GPU with gddr5 memory and if that hypothetical card existed it would still be considered high bandwidth cache it doesn't need to be HP m in order to be HPC under Andy's new naming scheme and the HPCC protocol so this is a look at the HPCC controller which is toggleable through the AMD drivers that is disabled by default and we're using 17 10.1 but there is a new 17.10.2011 so the results are based on that as for the cache controller when enabled the controller effectively converts vram into a last level cache and will tap into system memory as a sort of expanded video memory in this scenario if the applications page out of the on card 8 gigabytes of HP on to a trade-off between latency and capacity occurs and the GPU taps into system memory to grab its needed pages if you're storing 4k textures in HP m2 and exceed that 8 gigabyte capacity for example and then maybe need another one gigabyte for other assets those items can be pushed onto system memory and then pulled via the PCIe bus this is less effective than increasing on card memory but it's significantly cheaper than doing so and it's not always possible to increase the on card memory due to either physical space limitations or memory controller limitations things like that this is even in spite of DDR pricing right now and expanding its a system memory isn't a new thing for really any type of technology when looking at a GPS and ap use but it does enable Vega to potentially expand the working area for data when it's dealing with larger data sets something that Vega has really targeted since the beginning latency is introduced to by means of traveling across the PCIe interface through the CPU and down the memory bus and then back again but it's still faster than having to dump memory and swap data locally hypothetically this technology would permit large working data sets to exist on system memory and give the GPUs memory controllers somewhat direct access to that data set although it's worse for latency technically speaking the additional capacity should outweigh that latency deficit in the right scenarios so it's going to take proper implementations to ever realize the gains and that implementation must happen at both the software developer and AMD levels of deployment and the first showcase this technology in a tech demo in March around when Rison was launching and they showed HPCC on vs. off with a hypothetical vega card that didn't come to exist and probably won't or at least if not anytime soon that card was a 4 gigabyte vega GPU with an unspecified shader count so it is a 4 gigabyte card as opposed to the modern 8 on both 64 and 56 and because it was 4 gigabytes it was far more likely to become memory constrained than something with 8 it's pretty hard to fill an 8 gigabyte frame buffer with video games or even synthetic benchmarks so it makes sense that in order to showcase the difference you would do something like use a 4 gigabyte card unfortunately we don't presently have a good way of constraining our cards to four gigabytes like an attack demo but we have some ideas potentially for the future we'll see if they work or not though but either way when they showcase this Deus Ex mankind divided showed the differences of about 50% in average FPS according to Andy and they 100% improvement in minimums so and these showing averages got a 50% increase in performance by enabling HB CC which expanded the effective vram we'll call it that Vega could tap into when working with the games data set that would include things like textures for example now this is disabled by default it's been tested in the past around launch we figured we'd revisit it with some newer drivers and see if anything's changed we are using all the components to find in the article below and the testing methods are there as well along with all the components used but for the most important components 3200 megahertz of CL 16 memory is being used we have 32 gigabytes of it available to the system and of that depending on the scenario we're expanding and the Vega 56v Ram allocation to an effective 18 Giga bytes with HPCC on except in shadow of war where we only expand it to 12 gigabytes because we wanted to make sure there was enough memory there to work with the 4k special texture pack because there can be issues if if not it just won't load them and you may not even notice it so let's get started we're looking at not just the theoretical but the practical performance from HB CC and that means we're looking primarily at gaming and at synthetic workloads to really drill down into differences and build our accuracy and confidence we first started with fire striking scripted execution of 20 test passes with HB CC on and HPCC off using another in-house script to export and analyze all that data there's some variants but this many test passes will average it out we plotted a standard deviation of 20 7.48 points with HB CC off and 35.8 points with HB c is he on as for the scoring the HPCC off tests were 19400 18 point 1 3 points where HPCC on with 18 gigabytes allocation scored 19,000 683 this marks HPCC enablement as providing a 1.37 performance improvement over HPCC off which manifests itself as a boost in FPS which we've charted of 93.8 FBS average versus 95 point 2 FPS average for GT 1 and seventy six point eight versus seventy-seven point eight for GT two again this is over 20 test passes for each benchmark so we can be pretty certain that within the parameters of our test system and Windows this is a repeatable difference and is outside of variance of just the firestrike application times pi underwent the same treatment using the same script for its repeated execution we found much closer scores for this one with the average graphics score at sixty sixty-two point three for HPCC disabled or sixty ninety four point three seven for HBC C enabled that grants HPCC an advantage of 0.53% repeatably and translates into a scoring of 41 0.7 FPS average versus forty one point nine for gt1 and favoring HPCC enabled it's also thirty three point two FPS versus thirty three point three nine for GT two again favoring enabled superposition is our next synthetic test offering a look at performance scaling from the eyes of Unigine rather than future mark superposition wasn't repeated as many times but we do have five test passes for each configuration with minor standard deviations of two point nine points for HPCC on and one point four points for HPCC off and this is on a scale of many thousands of points so a deviation of one is really pretty good this application is impressively consistent in scoring HPCC off results in a score of 30 for 84 points in superposition as opposed to 3624 with HPCC enabled and that's a 1080p extreme the performance difference amounts to a four percent improvement which is the biggest we've measured thus far of course we're so bound by other elements of the card like the core frequency streaming processors Rob's so forth that a four percent improvement is only represented by a one FPS increase because we're just choking everywhere else and can't really realize that gain in a meaningful way as for frequency using stock settings enabling HPCC plotted a stable frequency of about 13 10 to 13 25 megahertz for the first half of the test while disabling HPCC plotted a more variable but higher frequency of 1315 to 1352 megahertz this is repeatable and happened effectively every time and we weren't really trying to control for that frequency change because we were interested in how the card behaved when HPCC was on versus off and this was part of that behavior and observation of that behavior despite the slight frequency deficit on the HPCC enabled test it was a bit more consistent and a card ultimately managed to improve performance by a few percentage points overall looking like about 4 in this case Sniper Elite 4 is our first game benchmark at 4k with high settings DirectX 12 and async compute we found performance to be functionally equal between the HPCC a/b tests with HPCC on blotting 52.7 FPS averaged 46 points to under sound lows and 45 is 0.7 lows disabling HPCC landed us at 52 point 2 FPS average of 0.5 FPS difference forty five point nine one percent lows and forty four point eight FPS 0.1% lows not only is there no discernible difference there's no difference outside of usual deviation and error margins so we can't actually say whether there's an improvement here it just kind of looks all the same moving into ashes of the singularity with DirectX 12 4k resolution and high settings we mostly saw the same these results had us at sixty six point six FPS average to sixty six point nine FPS average with forty four point three to forty three point five one percent lows and 41 verses forty 0.1% lows for the FPS once again these are within test a test deviation Ghost Recon wildlands at 4k plots the HPCC tests at 36 FPS average with HPCC on and off with lows largely tied at 33 and thirty two point five were 32 and 31 point eight F yes so all pretty much the same numbers there's no appreciable difference here and we are once again within variance we also saw no difference at 1080p that that's less significant given the diminished focus on memory at this but pretty much the same results if you were curious about 1080p for honor at 4k and using extreme settings makes for one of the more memory intensive real-world to games that we work with they go with HB CC enabled bought at a 43 FPS average versus 42.7 when disabled marking these scores functionally equivalent and with invariance the lows are forty point seven FPS versus forty point three and thirty nine point three zero point one percent lows versus 38 once again all this is within test deviation and variance for this particular title at 1080p for honor has us at 125 FPS average for both HPCC enabled and disabled we would expect less likelihood for differences to emerge have these lower resolutions given the vram users reduction we configure to help lead to 4k with very high settings providing one of the more taxing games for testing a vago 56 card under both HPCC conditions performance was exceptionally consistent and plotted about the same numbers between HPCC on and off once again for held late at 4k testing shadow of war with the normal texture pack 4k resolution and high settings we were at 42 to 43 FPS average on both iterations of the HPCC testing marking equal performance once again we later enabled the 4k texture pack and ran ultra settings with HPCC reconfigured to a 12 gigabyte frame buffer to preserve ram for the pack just in case though settings gave us once again about the same performance between our HPCC on and off tests consulting the work of friend of the site rob williams over at tech age it looks like they found pretty much the same results when the card first shipped or close to it anyway their results for the most part featured basically the same plus or minus 1 FPS within error types of differences and they did also test mankind divided which we didn't do for this one and tech gauge didn't measure any meaningful change with the 8 gigabyte Vega cards but like we were saying at the beginning of this content the article and the video it looks like perhaps there would be a bigger difference if you could restrict Vega to 4 gigabytes so this HPCC control the toggle for the controller may come into play more if either a 4 gigabyte vega cart comes into existence which would be an awfully interesting combination the price of HBM at four gigabytes i don't know we'll see if it works out if it ever actually exists but that would be a scenario where you would potentially see a difference and you certainly showed that in their tech demo we can't recreate it because we only have the cards that exist on the market which are 8 gigabytes and the other option would be potentially if somehow games just started storing tons and tons of data into vram in the immediate future or maybe if you resurrect these cards and several years when probably they are end-of-life for various other reasons maybe at that point because of technology changes and shifts in game development you might see implementation of or actually utilization rather of that expanded memory or effective memory on the video card what we haven't tested is production level applications I'm not sure if tech gauge has tested that yet or if they plan to but production applications are a potential area where there might also be gains Andy has its own SSG solution for that space particularly with things like live 8k scrubbing and premiere with the SSG Radeon cards so there may be some kind of impacts in production level workloads but for gaming we're not really seeing anything right now in the games we've tested it does look like superposition plots a somewhat statistically significant change in that we're seeing a four percent uplift with HPCC the only problem there is that by the time we're tapping into that advantage the extra four percent advantage we're so limited elsewhere in the pipeline that it's just as a player if you could play the game superposition if it were a game at the point that the advantage is coming into play you're already at such a low FPS that it's unplayable anyway so you'd actually be better off just lowering your settings sacrificing the advantage and running with something that's a bit easier for the card to work with the exception would be if you had a workload that had a very memory intensive set of assets like maybe 8k text or something we haven't tried for example modding fallout or Skyrim if any of you have let us know what your results were but those would be scenarios where you could definitely pump the memory usage on the video card with user created assets that are one less optimized and two don't have the same considerations that game developers have in terms of capacity console compatibility things like that so they can go heavier on the memory consumption on the video card so those would be all the scenarios where you might see an uplift for the stuff we've tested thus far we're not seeing one so don't worry about that toggle right now it's not really worth turning it off you don't really get a performance hit unless you run into issues with system memory limitations so from what we've seen that you could probably turn it on and maybe every now and then you'll get a 0.5 to 4 percent performance uplift it's not really a bad thing but if you do encounter stability issues obviously it's also not a bad thing to just turn it off and live without it so one of those things where the technology is there and it's kind of questionable when it works but it seems to work in some instances particularly synthetics so those might be enough to encourage turning it on but it's certainly not a silver bullet for anything so that's all for this one we'll revisit this as more driver updates and games and applications seem like they're using it Wolfenstein 2 will be a potential option for looking into this because Bethesda has sort of cryptically stated that they will be leveraging Vega to greater levels than other games do currently probably just because of olkin but we'll see as always you can go to patreon.com/scishow and actions to helps that directly or stored our gamers nexus dotnet to pick my shirt like this one thank you for watching and subscribe for more we'll see you all next time but for the most important components 3200 megahertz of a it's very loud thunder