Ask GN 103: Multi-GPU Comeback? NVIDIA Limiting VRAM on Purpose?

This Ask GN is weeks overdue! We shot this one on October 3, but only just got around to finalizing it due to the crazy review cycles.

TIMESTAMPS

Multi-GPU Making a Comeback?

01:30 – TheDireLynx: “With Nvidia's new NVLink connector for SLI and AMD's rumoured counterpart (xGMI), as well as newer Nvidia sponsored titles working very well with multi-GPU setups, do you think multi-GPU will make a comeback? How do you feel about multi-GPU seemingly becoming more and more of a focus for both companies and do you think we'll see AMD try the same thing they did with the RX480 when Navi eventually comes (comparing higher end Nvidia cards to a multi-GPU setup)?”

Why Does Multi-GPU Scale Poorly?

06:10 – Soupa: “#askgn-questions Fundamentally, what causes such poor scaling in most games when using SLI (or is it NVlink)? Is it due to latency of the communication between the two or more GPUs?”

Why Not Use the PCIe Bus Instead of Bridges?

11:33 - Luke Matters: “Hi Steve, just question about sli/nvlink. I saw a video benching 2080 for mining and they used 2 cards and didn’t have the actual nv link.”

Which Game Graphics Settings Hit Perf Most?

15:45 - Frédéric Campagna: “What are the other settings you could suggest I should look for and maybe even explaining their impact on the system's ressources (video memory usage, GPU usage...)? Not asking specifically for this game but more on a general basis.”

RAM Overclocking?

21:50 - Grimm_XIII: “#askgn-questions @Steve Burke Hey Steve, wondering what exact kit of RAM you're using in your #RIPJAY stream is and if you have any recommendations for generally well-overclocking/performing sticks? I know G.Skill and their B-Die memes are highly regarded overall but am not sure if I should just try to tune a 3200 CL14 kit or intentionally buy like a 4000 kit for instance. Thanks.”

Is NVIDIA Purposefully Limiting VRAM?

24:00 - Andy Funderburg: “Do you feel Nvidia is purposefully limiting VRAM to keep the lifespan low so more sophisticated models are unable to run on older cards? If so, why are we not talking about expandable memory on GPUs. If we’re okay with a GPU taking up 2.75 slots, having multiple 8-pin connectors, and being priced over $1k USD, shouldn’t consumers have options to upgrade (GDDR DIMMs or GPU sockets)? Edit: I just signed up as a patron. Glad to be able to support the channel.”

Why is Game Clipping So Hard?

28:16 - Freedom4556: “#AskGN Regarding your gaming graphics request for questions: Why is it so hard for game developers to make sure nothing clips through other geometry in games? For example in Mass Effect having weapons on your back clip through character's model, or in Fallout 4 having the PIP boy clip through the female model's forearm during the walking animation. Is this just a case of there being too man permutations to test, or is it just not worth the effort to resolve?”

How Much Do You Work Per Day?

31:01 - Oreo Lamp: “#askgn How much do you actually work per day and what kind of work do you usually do?”

Top 3 Places You’ve Mountain Biked?

33:29 – YungViking: “#askgn-questions Hey Steve. I'm also a very avid mountain biker and I wanted to ask what you would rank as your top 3 places you've ever ridden. I know you've had some opportunities to ride different downhill parks like Whistler with the traveling you do. Also would there ever be a possibility of a Patreon meetup somewhere like Bailey Mountain Bike park? Not sure how many other Patrons are on the East Coast and also enjoy Mountain Biking but I think it could be a cool way for you to interact with your audience outside of conventions. I would love to hear your thoughts, thanks for your time!”

** Please like, comment, and subscribe for more! **

Links to Amazon and Newegg are typically monetized on our channel (affiliate links) and may return a commission of sales to us from the retailer. This is unrelated to the product manufacturer. Any advertisements or sponsorships are disclosed within the video ("this video is brought to you by") and above the fold in the description. We do not ever produce paid content or "sponsored content" (meaning that the content is our idea and is not funded externally aside from whatever ad placement is in the beginning) and we do not ever charge manufacturers for coverage.

Follow us in these locations for more gaming and hardware updates:

Host: Steve Burke
Video: Andrew Coleman, Josh Svoboda

[alpaolodecastro]

Tech jesus delivering content anytime of the day

[mashygreen]

It’s important to note that what SLI does is render a frame on each card sequentially and sends the rendered frame from the ‘slave’ card to the ‘master’ card. Only the ‘master’ card is connected to a monitor and the fully rendered frame has to pass from the ‘slave’ card to the ‘master’ card for output. That is what an SLI bridge mostly does - offers high bandwidth for passing a rendered frame.

Explicit mGPU does something very different where the actual rendering work itself is split between the cards. Each card rendered a portion of the frame and they get combined then outputted from the ‘master’ card. This way there is actually less bandwidth passed between the cards (at the start of a frame each card gets a portion of the frame, works on that independently, and just the portion from the ‘slave’ card has to be passed to the master). That’s why you can get away with it without a bridge.

Mining on the other hands is a completely independent workload. Each card works as if it is the only card. There is no output here and the cards don’t have to send each other any kind of data at all. Each card will have the exact same performance in a multi-GPU setup as a single card. It is similar to running two separate single threaded CPU programs rather than a single program that uses many cores (kinda).

If you want some more info on the compute side and NVLink specifics outside of SLI I’ll be happy to try and provide more info on that too. I personally only have access to NVLink via motherboards, not bridges (with Tesla cards on an HPC) but the idea is similar.

[droknron]

You keep calling Ashes of the Singularity a benchmark but I have like 100 hours in that game and play it constantly. For RTS lovers like myself (Command & Conquer, Supreme Commander, Planetary Annihilation etc) it's really quite good and enjoyable. It also runs really good with a lot going on at once on absolutely huge maps which really make it a lot of fun, I mostly play Skirmish.

I know that Stardock worked a lot to make the game easy to benchmark and utilising a lot of DX12 features like explicit multi-GPU rendering but that's for marketing purposes to keep the game in reviewers lineups for longer, it's a strategy to sell more copies of the game and it does work. If you want more evidence that it's an actual real playable game just look at all the DLC they've released for it, all playable content, new maps and so forth.

[mr_beezlebub]

A question for the next video: What do you think of Valve and Code Weaver's efforts with making Windows games compatible on Linux using Wine and DXVK in Valve's Steamplay initiative?

[Tarulia]

For the 6th question (socketable vram etc) it's also worth mentioning that like with CPUs those dies are basically manufacturing defects, which is also why the empty spot on the 2080 is different on each single card, because the controller where it's missing is just straight up not working. I think buildzoid actually mentioned that in the PCB breakdown.

[rysyndrome]

33:10 Why I am jealous but thankful for GN! So not easy to put in that time, mad respect and thanks!

[kkbqmbu]

No offense, but multi-gpu is being done away with because it doesn't sell more cards for the company; it sells used or old cards. If companies really wanted, they could create a GPU architecture that scaled within the GPU loop itself instead of with some software bandaid. Nvidia wants me to buy a 2080 Ti, instead of simply adding an extra 980 Ti to my rig, for example.

[swayingGrass]

#AskGN why do some game require you to reset/restart the game to change some/any graphic settings while some don't?

[renehoyvik]

12-14 hours a day? God damn! Since no one else is saying it. Thank you to you an your team for the hours of dedication to bring us great content! :)

[Thermophobe]

#AskGN given the significant improvement in streaming performance using NVENC in RTX series cards, can new PC builders who want to stream consider skipping higher core CPUs (over 6)?

[NightKev]

#AskGN - I guess this is the time to ask multi-GPU based questions; do "multi-die" (not sure if that's the right terminology here) GPUs like AMD's "R9 295 X2" solve any of the issues with technologies like SLI? Perhaps the Ryzen approach to CPUs could be applied to GPUs as well?

[RathOX]

in the uk atm it would cost me £1499 for a strix 2080Ti. i can SLI my 1080Ti poseidon's for less than that.

[Mandragras]

I was about to ask these days what happened to Ask GN. And there it is, with quite interesting questions.

[cheesewheele]

Speaking of mountain biking, have you ever made it over to Windrock OHV? They have a large downhill mountain bike course with shuttles to shuttle you back to the top running all day. It is only ~7 hours from where y'all are located so it is easily doable in a weekend.

[duaynenantau]

It is a shame you don't have more subs/views. You do exceptional work!

[aidanriches]

#AskGN How much does RAM speed effect Intel cpus? I know that Ryzen benefits greatly from higher speed memory (because of the way infinity fabric works right?) just curious as to what speeds will be the best value for my cpu and how the performance will change.

[johnlpz]

Watching your videos every day does that mean we are all nerds? Love your videos by the way. Keep it up

[xabiergranja]

This was really interesting. I literally saw NO difference in mesh quality between low/high (16:40). Good to know it has such an impact, as I can lower it from now on and not see the difference.

[CMDRSweeper]

Best way to explain why mining doesn't need any linkage between them is also because for mining, your data isn't time sensitive.
If you do not deliver the data quickly enough to the screen for a game, it results in low FPS or poor performance.
However, mining it is usually a "Done when it is done" instruction and you give different instructions to different GPUs.
That means that the GPU can sit there chugging on its workload or part of the job, without needing to worry about getting the answer off to the host as quickly as possible.

[kathmandu]

Interesting question about lapping heat spreaders and silicon. Lapping a heat spreader, in theory, produces a flatter surface for better heat transfer but all I've ever seen are people lapping the items cold. I have seen folks saying some cooler cold plates are specifically machined to compensate for heat spreaders changing shape as they warm up. The question is are we screwing up by lapping cold and not checking for flatness at normal running temperatures?