Place Your Bets Now About The Power Efficiency Of The Radeon RX 480 On Linux

[juno]

It is hard to predict how the market will react to APUs with HBM because this particular combination of technologies will be a new one. The question is whether people will be willing to pay ~300 EUR for an APU equipped with 4GB HBM for example.

No, most certainly they won't.
But I could see an APU with 1-2 HBM stacks for a lower price, and there where it makes sense. On the desktop, you have plenty space and power and there are not much applications yet that benefit from HSA, unified memory or higher bandwidth between CPU/GPU. So you'd rather take a decent CPU+GPU combination.

On a notebook, however things are different. There are plenty notebooks that don't use SO-DIMMs anymore but the RAM soldered directly onto the mainboard, so you can't upgrade either. It is smaller and easier for the OEM to build with to just one central die to implement and cool instead of adding a MXM module. Intel uses eDRAM for their Iris Pro parts, which is expensive, too. But that's more like a L4$, just 128 MiB max. And they have no graphics solution that could compare to something like 12+ GCN CUs (I think +50% for 14 compared to 28 nm seem legit).
An APU with 2-4 Zen Cores, 12 CUs and HBM would be premium. Intel has recommended customer prices of 490-1200 USD for their recent notebook CPUs with Iris Pro and they sell. Yes, I know no OEM is paying those prices, though.

It doesn't even have to be the full system memory with 16 GiB (however possible with two 8Hi-stacks), 1 stack would be enough as a start, compared to Intel's eDRAM.



I don't think one HBM stack and the small interposer would make this unaffordable. IMHO, "Premium OEMs" could be interested in something like this. Apple could even use that and push HSA, If they liked to...

[juno]

There are no stronger APUs than Bristol Ridge yet. Console SoCs are semi-custom, designed together with and funded by the customers/partners.
Raven Ridge should bring both Zen cores and GCN Gen.4 together next year, if we are lucky maybe even with HBM, but I think that's more wishful thinking and out of profitability

[artivision]

Just curious, what is this based on ?



Yeah, as long as consoles plug into the wall and laptops run on batteries there's going to be pressure to put smaller HW into laptops. That said you see big-ass gaming laptops today and I don't see any sign of that changing.



Don't understand this - you're talking about Line Pitch aren't you ? What is the connection between lowest LP lithography and clocks ? Laptops tend to get low clocks simply because wide low-clocked hardware tends to use less power than narrow high-clocked hardware for the same performance.

There is a connection of Lithography with Frequency and Consumption. If you cut your laptop processor with a High version then you can get 3.7Ghz or more with an X consumption. If you cut with a Low method you cannot get more than 3-3.3Ghz and you will have 0.5X consumption even if the difference on Frequency is not 1.5-2X.

Also for strong Laptop APUs, I mean if they will exist at all, not if they require bigger Laptops.

[bridgman]

I am afraid from the bad naming(M480 for an equivalent of the RX460), that the M480 price will be to high.

Just curious, what is this based on ?

I am also afraid that consoles will get again a stronger APU(4.5-6Tflops) versus laptops(2-2.5).

Yeah, as long as consoles plug into the wall and laptops run on batteries there's going to be pressure to put smaller HW into laptops. That said you see big-ass gaming laptops today and I don't see any sign of that changing.

I also don't understand why laptops don't get the lowest LP lithography (3-3.3ghz max for the cpu, 0.9-1ghz max for the gpu).

Don't understand this - you're talking about Line Pitch aren't you ? What is the connection between lowest LP lithography and clocks ? Laptops tend to get low clocks simply because wide low-clocked hardware tends to use less power than narrow high-clocked hardware for the same performance.

[artivision]

Not sure about intended launch sequence; we always talked about them internally as a pair so I imagine the plan was to launch at more or less the same time. My impression was that per-CU PG was for the smaller chip only since that was the one that would see heavy use in laptops.

Golden settings are the accumulation of all the "set register ABC to XYZ" directions we get from the hardware team, turned into a sequenced data table (sequence can be really important) and set up at the start of HW initialization (or presumably on restore).

I am afraid from the bad naming(M480 for an equivalent of the RX460), that the M480 price will be to high. I am also afraid that consoles will get again a stronger APU(4.5-6Tflops) versus laptops(2-2.5). I also don't understand why laptops don't get the lowest LP lithography (3-3.3ghz max for the cpu, 0.9-1ghz max for the gpu).

[bridgman]

Not sure about intended launch sequence; we always talked about them internally as a pair so I imagine the plan was to launch at more or less the same time. My impression was that per-CU PG was for the smaller chip only since that was the one that would see heavy use in laptops.

Golden settings are the accumulation of all the "set register ABC to XYZ" directions we get from the hardware team, turned into a sequenced data table (sequence can be really important) and set up at the start of HW initialization (or presumably on restore).

[juno]

Just a caution - IIRC the per-CU power gating is for Polaris 11 only and the initial test results will be for Polaris 10.

Oh really? It is not supported by hardware?
That's sad, even CZ was capable a year ago (while I don't know if supported by any production driver)...

https://cgit.freedesktop.org/~agd5f/...t-4.8-wip#n410
https://cgit.freedesktop.org/~agd5f/...-4.8-wip#n5368
https://cgit.freedesktop.org/~agd5f/...t-4.8-wip#n162

Seems you're right, however

Polaris11 only for now

Does raise hopes... Still a pity it couldn't make it into the release drivers, however I'm not sure about the actual effect, of course...
Still a bit confusing, was P11 planned to be released before P10 initially?

One more day to go, anyone knows the actual time of NDA expiration?

BTW: What are "golden settings"? I wonder all the time...

[drSeehas]

... IIRC the per-CU power gating is for Polaris 11 only and the initial test results will be for Polaris 10.

So tomorrow we won't see test results for RX 460 (Polaris 11)? Only RX 480? :-(

[bridgman]

Just a caution - IIRC the per-CU power gating is for Polaris 11 only and the initial test results will be for Polaris 10.

[juno]

I would not place a bet but looking at some of the open code contributions, polaris might have the potential to improve efficiency on Linux a lot. Of course I don't know how mature the driver code is, but when I read things like "per CU power gating", it could save a lot of power in those exact situations where it is underperforming and not fully occupied. "Just" completely power-gate those CUs that are not working, while keeping the clocks up and not throttle other parts of the chip down -> rather big efficiency boost!(?) we'll see...