Requesting GPU Reviews!!! 128 gamers 1 PC?!?!

[Jrhumphrey89]

I would like to request that LTT please do a review of nvidia M10, M6, and M60 graphics cards and show what they are all about. Maybe something like using 4 M60 cards to do 128 gamers 1 PC? Then to test it you can have fans of LTT run a remote desktop and game on it with you live in twitch during Wan show? Everyone run a benchmark together and put their results in some google doc???

 

I loved when Linus did the Multiple people on one pc video. Now I might be in the minority here, but I love getting reviews of high end enterprise hardware, cause honestly that's where the big money and performance is. I was really hoping after Linus started down the path of multi gamers 1 cpu that we would get into a lot of virtualization and server gpus, but it never happened. I am super curious about the scaling of how many people you can have running off of a single system using new enterprise software\gpus and what the experience is like.

 

Just something I would like to see. Like the post if you agree.

[Jurrunio]

Compute cards are powerful, but I dont think they can do the job of a CPU.

[DocSwag]

I think the bigger problem is getting enough CPU cores to make 128 gamers viable.

[Hogger]

9 minutes ago, Jrhumphrey89 said:

I would like to request that LTT please do a review of nvidia M10, M6, and M60 graphics cards and show what they are all about. Maybe something like using 4 M60 cards to do 128 gamers 1 PC? Then to test it you can have fans of LTT run a remote desktop and game on it with you live in twitch during Wan show? Everyone run a benchmark together and put their results in some google doc???

 

I loved when Linus did the Multiple people on one pc video. Now I might be in the minority here, but I love getting reviews of high end enterprise hardware, cause honestly that's where the big money and performance is. I was really hoping after Linus started down the path of multi gamers 1 cpu that we would get into a lot of virtualization and server gpus, but it never happened. I am super curious about the scaling of how many people you can have running off of a single system using new enterprise software\gpus and what the experience is like.

 

Just something I would like to see. Like the post if you agree.

You couldn't get quite to 128 gamers, the reason that the multiple gamers worked was because he split the processor power up, and through this software, he could give each person a GPU. There isn't a processor out there that can provide enough cpus for 128 people. Even if you get a dual socket board, impossible. You COULD get that 4 socket board from that one company Linus visited, but then you'd only have 1 or 2 PCI-e slots, not enough space for graphics cards, especially, the bandwidth would be a bottle neck as well.

[DocSwag]

1 minute ago, Hogger said:

You couldn't get quite to 128 gamers, the reason that the multiple gamers worked was because he split the processor power up, and through this software, he could give each person a GPU. There isn't a processor out there that can provide enough cpus for 128 people. Even if you get a dual socket board, impossible. You COULD get that 4 socket board from that one company Linus visited, but then you'd only have 1 or 2 PCI-e slots, not enough space for graphics cards, especially, the bandwidth would be a bottle neck as well.

The new Xeon platinums support 8 socket configs, which at the top would give you 8*28 or 224 cores. That's still only enough for some pcs to only have one core though, plus you can't possibly connect them to 128 GPUs.

[Ashiella]

1 minute ago, DocSwag said:

The new Xeon platinums support 8 socket configs, which at the top would give you 8*28 or 224 cores. That's still only enough for some pcs to only have one core though, plus you can't possibly connect them to 128 GPUs.

X4 top of the line e7 xeons get to 192t, soooo, almost?

[Jrhumphrey89]

10 minutes ago, RandomGuy13 said:

What makes you think that you can get 128 gamers from 4 of these cards.

 

Correct me if I am wrong, but the way that Linus has done the 8 gamers 1 CPU is to pass each card into a single Windows VM, and use the display outs on that graphics card to connect a monitor. Tesla cards don't have display outputs because they are compute cards. So I don't see a way of getting a single display out of them, let alone the 32 you are suggesting.

 

I guess you maybe could do it through steam inhome streaming, but I don't see how you can achieve multiple VMs on the same card, while having the direct passthrough that games require. Maby if you did an older game.

 

I haven't got any experience with Tesla cards so there might be a technology that I don't know about that might let a VM have direct access to a GPU.

I would however love to see Teslas being shown on the channel, I dont think they really have a use for them though.

I would like to direct you to nvidias grid datasheet. http://images.nvidia.com/content/grid/pdf/188270-NVIDIA-GRID-Datasheet-NV-US-FNL-Web.pdf

Using Grid virtualization GPU passthrough you can get can get a Tesla M60 to a max  of vGPU instances 32 users for "Performance optimized". This is what the cards are designed to do.

[DocSwag]

1 minute ago, Jrhumphrey89 said:

I would like to direct you to nvidias grid datasheet. http://images.nvidia.com/content/grid/pdf/188270-NVIDIA-GRID-Datasheet-NV-US-FNL-Web.pdf

Using Grid virtualization GPU passthrough you can get can get a Tesla M60 to a max  of vGPU instances 32 users for "Performance optimized". This is what the cards are designed to do.

The problem is, how do you get display outputs off those

[JoostinOnline]

1 minute ago, DocSwag said:

The problem is, how do you get display outputs off those

Splitscreen the hell out of it.

[DocSwag]

Just now, JoostinOnline said:

Splitscreen the hell out of it.

Teslas don't have display outputs tho

[Jrhumphrey89]

I understand you arguments about CPU however with multi threaded games and a quad socket montherboard I think it is possible. Even if it doesn't have mind boggling performance I think it would be worth seeing what the performance is. It is recommend for load balancing and multi threading that you assign 4+ virtual cores per thread core. So a quad socket 20 core server should be able to host 200 people depending on their workload easily.

 

As for connecting each to a gpu I refer you to nividas grid datasheet. http://images.nvidia.com/content/grid/pdf/188270-NVIDIA-GRID-Datasheet-NV-US-FNL-Web.pdf

Using Grid virtualization GPU passthrough you can get can get 4 Tesla M60 to a max  of vGPU instances 128 users for "Performance optimized". This is what the cards are designed to do.

[Jrhumphrey89]

6 minutes ago, DocSwag said:

The problem is, how do you get display outputs off those

Through the power of the inter web. This are not meant to be plugged into the back of PCs and monitors, they don't even have video output!!! You connect to a virtual session that runs the vm. This is exactly how nvidia does their shield game streaming.  

[Eduard the weeb]

you know how linus one time was like it is actually 2 cpus well it actually is 12 cpus in 3 separate mobos or one of those AMD super computer that each cpu can handle 1 terabyte of ram also linus will finally have 100 percent ram using on a 128 gig kit lol

[Eduard the weeb]

interesting thought no matter how unpractical.

[Mira Yurizaki]

While I appreciate the probable "For Science!" shpiel, I think the biggest stumbling block is I don't even think Linus can just ask NVIDIA for a bunch of high-end compute cards. Especially ones they don't really sell to consumers.

[Jrhumphrey89]

2 minutes ago, M.Yurizaki said:

While I appreciate the probable "For Science!" shpiel, I think the biggest stumbling block is I don't even think Linus can just ask NVIDIA for a bunch of high-end compute cards. Especially ones they don't really sell to consumers.

I think if anyone can ask for expensive computer hardware for no practical purpose its linus

[Eduard the weeb]

2 minutes ago, M.Yurizaki said:

While I appreciate the probable "For Science!" shpiel, I think the biggest stumbling block is I don't even think Linus can just ask NVIDIA for a bunch of high-end compute cards. Especially ones they don't really sell to consumers.

this project would require a already built super computer like either of the one Linus took a tour of and the linus helped build for AMD

[Jrhumphrey89]

19 minutes ago, Eduard the weeb said:

this project would require a already built super computer like either of the one Linus took a tour of and the linus helped build for AMD

I would say 8 a socket motherboard with: The Intel® Xeon® Platinum 8180 Processor

 

So 448 threads total, 1 TB of ram, and 4+ M60 tesla gpus would do it

 

You could probably get away with half of that and have 2 cores (256 threads), 4 gigs of ram (512 gigs), and .5 gigs of vram (4 m60 gpus) for entry level games on med graphics 1080p.

 

The thing about this is you can assign 4 virtual cores per thread. When you run a game on 4 threads you don't see 100% cpu usage on all 4 threads. I think with 128 people assigned to a server you will see load balancing of vm loads across under utilized threads 

[Mira Yurizaki]

44 minutes ago, Jrhumphrey89 said:

I think if anyone can ask for expensive computer hardware for no practical purpose its linus

I think the difference here is if a company gives Linus hardware, it's because they're sponsoring a project for basically free advertising of their product. i.e., A company will give Linus something in hopes a consumer will buy it. Even if the consumer doesn't intend to use it in the way Linus does, it's still advertising. So getting a bunch of high-end Xeons and Teslas together doesn't make sense to me for a company to sponsor because Linus's target audience aren't those whom said parts are marketed for.

 

The only evidence to support this theory is the curious lack of professional grade hardware in their setups. But that could also just be due to them not really having reliability issues using consumer grade hardware.

[DocSwag]

1 hour ago, Jrhumphrey89 said:

Through the power of the inter web. This are not meant to be plugged into the back of PCs and monitors, they don't even have video output!!! You connect to a virtual session that runs the vm. This is exactly how nvidia does their shield game streaming.  

I see. You still have the issue of getting CPUs that can power it though.

[Hogger]

128 windows of mine sweeper

[Shreyas1]

3 hours ago, DocSwag said:

I think the bigger problem is getting enough CPU cores to make 128 gamers viable.

The biggest problem is getting 128 people

[Jrhumphrey89]

11 minutes ago, Shreyas1 said:

The biggest problem is getting 128 people

Live Wan show test?