Not only that, Intel's Skylake-SP cache and interconnect changes appear to be hurting its performance in gaming workloads.

In many games, Skylake-X (which uses the same dies as Skylake-SP) actually performs worse than mainstream Skylake-S (or Kaby Lake-S, which is microarchitecturally identical but manufactured on a slightly-improved process) at the same clocks.

At 10:27 in the above video, Adored notes that in an unexpected change of pace, NVIDIA has started to recommend AMD CPUs!

@Bob iirc it's also os dependant (since it needs to support two drivers) and needs on die integrated graphics.

The fact that enthusiasts are having serious thermal problems with Skylake-X, along with the pricing and performance potential of Threadripper, does not bode well for the future of the Intel HEDT platform.

@JourneymanGeek I think virtu isn't a thing anymore :/

Nope. I don't have anything newer to play with. It's possible it just works.

@Bob it was barely even a thing when I had an X58 motherboard that supported it; it was ridiculously broken, with only certain OS version + Intel driver version + AMD/Nvidia driver versions supported, and it took months for them to update after the vendors released a WHQL driver

Pre Sandy integrated graphics on Intel switched between integrated and discrete on the same pcie lanes

and that was at the best of times... they just stopped updating at some point, then they tried to charge for new versions of the software for hardware you had on your motherboard

a horribly run company (engineering-wise) by all estimates; they had such an opportunity to capture a market for combining IGP and discrete performance into one thing by working closely with Intel/AMD/Nvidia, but they let it slip away

So... pcgamer.com/…

@allquixotic and on the desktop, they were solving a non existent problem, especially as discrete GPUs got more efficient at idle

The i9-7900X hits 4.0 GHz on all cores with full Turbo Boost 2.0.

@JourneymanGeek to a certain extent, yes, but people who have marginal discrete GPUs like the GTX 1050 might actually see meaningful performance gains by adding at least a fraction of the performance from, say, Kaby Lake GT2 or GT3 onto what the discrete can already do.

laptops moreso than desktops, obviously

@allquixotic but that didn't add on. It just let you switch what cards you used to render.

The i9-7920X is probably going to do something in the ballpark of 3.6 GHz, but given that the i9-7900X can already overheat at full Turbo speeds without overclocking, I'm not sure if they can actually hit these sorts of clocks without similar thermal problems.

the combined horsepower of a Kaby Lake GT2 IGP + a mobile 1050 is probably about as good as a 1050 Ti, maybe more depending on how well it's optimized

@JourneymanGeek wrong; Virtu MVP was designed specifically to add on!

So vaguely Optimus like

it wasn't just another hybrid graphics switcher like Optimus

it'd actually offload rendering calls onto your IGP if your discrete was pegged

while rendering on whichever framebuffer your monitor was plugged into

@allquixotic eh. Was that the later paid addon?

I'm actually more concerned about how the 16C and 18C chips will perform.

@JourneymanGeek nope, when I first got my X58 I got it to work with several games

Ah

Odd. Never installed or played with that for some reason

they did something silly like hook the D3D9_xx.dll, intercept the calls, and broker out DX9 render calls to an offscreen surface on your IGP, then combine "whatever it did" back into the main scene on your discrete's framebuffer

Oo

Well, the problem is that the gap between integrated graphics and dedicated graphics has widened dramatically over the years and this sort of acceleration (by using both GPUs concurrently) is no longer viable.

obviously it couldn't work with every possible draw call or sequence of calls

Yup

@bwDraco I dunno; the Radeon Pro 560 in the 15" MBP, a bona fide discrete card with 4GB of VRAM, is only 3x faster than the Kaby Lake Iris graphics in the 13" MBP

in an ideal world with a perfect driver / software ecosystem optimized for it, your total perf would be (4/3) * perf_of(Pro 560) in a system with both Iris IGP and a Pro 560

Hm. In theory you could offload specific tasks by design to another card. Probably a pain to write :p

...and I'm getting sick and tired of waiting on renovations.

in a non-ideal world it'd be something greater than the base perf of the 560 and something less than 4/3*

For the same reason SLI/CrossFire scaling is less than linear.

@JourneymanGeek aren't some games starting to use OpenCL or CUDA for physics? they could certainly just offload some or all of that to your IGP and only use your dGPU for graphics rendering

there might also be machine learning or "AI" algorithms running on your GPU in upcoming games

and many of those algorithms are so embarrassingly parallel that you could benchmark the user's IGP and dGPU and assign a proportionate amount of the workload to each

if you could do, just to throw out a number, 5,000 operations of a certain type on the IGP per second and 25,000 on your dGPU, as you're going along you could assign 1/6 of the compute work to the IGP

AMD saw the future in GPU acceleration. They were too far ahead of their time, but now it's becoming a big factor.

@allquixotic or throw in an older graphics card

No matter how you look at it, Intel HEDT is a huge mess.

@JourneymanGeek yeah; there are programs for altcoin mining (e.g. Ethereum's ethminer) that already support multi-vendor, multi-platform simultaneously mining

on my Alienware at home, I can mine Ethereum on the GTX 1060 via CUDA and the external (TB3) RX 580 via OpenCL simultaneously from the same process

@allquixotic That's easier because the GPUs don't have to work in lockstep. With gaming, you need consistent frame delivery.

and because the problemspace is so embarrassingly parallel, my hashrate is purely additive

20 MH/s on one, 25 MH/s on the other, and when combined I get a straight 45 MH/s

Lol. Is mining etherium on Nvidia any efficient?

For context on allq's Alienware machine:

@JourneymanGeek it's actually more efficient on Pascal the lower-end your card is, oddly enough

Oo

and the desktop RX 580 gets 20% more than either of them

Ah

Also. Ugh. Where is my tbolt flavoured beast :p

(budget external dock)

dunno about budget but the Apple eGPU developer kit is pretty reasonably priced, and more importantly, in stock (a term rarely applied to the Polaris 10 GPU family, thanks Ethereum!)

Case in point: GT 1030 (GP108) uses 30W. GTX 1050 Ti (GP107) uses 75W (150% more power) but delivers only 90% more performance, based on TFLOPS at the boost clock speed.

@bwDraco Oh, did @allquixotic keep this one? :P

@Bob I did!

it's my primary gaming Googling box while in my slacking/procrastinating room office

@allquixotic the current models are sub100 USD...

Well, the laptop I'm currently on, a low-end Lenovo 2-in-1, was a bit of an impulse buy...

So they will be rediculously cheap compared to anything on the market

@bwDraco Hello! Everything I own (even my car) was completely an impulse buy. :D

I'm basically @NotDog with a job