Will Nvidia be selling 7nm graphic cards in 2019?

[chroniclard]

Do I spot an AMD hype train slowly pulling out the depot?

 

[LePhuronn]

Do I spot an AMD hype train slowly pulling out the depot?

Not intentionally choo choo!

 

[KillBoY_UK]

I wonder if Zen and the mining boom have given AMD enough cash to spin Navi off into 2 memory directions? Keep the HBM version for the pro and datacenter cards, but branch off a GDDR6 variant for consumer gaming.

AMD already announced they upped there R&D budget 25% this year but its still less than what it was in 6 years ago around $350m per quarter, and way down from a Q4 2007 peak just over $500m

AMd and there priority have changed a lot since then and they might just double down on there semi custom designs for there customers (ie Navi if the rumours are true) as well as CPU's designs and discreet GPU's remain the 3rd priority

 

[Mauller]

I really hope you're right, the AMD high-end really needs a boost!

 

[Mauller]

At the very least would need redesigned memory controller (beyond just a die shrink) for it to work, not for performance reason but for power draw reasons.

We do badly need some competition

V20 is much more than a die shrink, It's size alone should tell you that. It would be as big as GV100 if it was on 12nm.

 

[D.P.]

V20 is much more than a die shrink, It's size alone should tell you that. It would be as big as GV100 if it was on 12nm.

They have added back in the 1:2 FP64 support, that takes up a lot of die space.

 

[Mauller]

They have added back in the 1:2 FP64 support, that takes up a lot of die space.

not as much as you think, they don't use dedicated cores like Nvidia. Their DP support is more like SMT in that it requires a few % more transistors per pair of ALU's.

but there is still more to the hardware for the die size compared to adding DP alone.

 

[KillBoY_UK]

V20 is much more than a die shrink, It's size alone should tell you that. It would be as big as GV100 if it was on 12nm.

I know, AMD said they doubled the bus width of a already power hungry memory controller. So if its Still "Vega" ie GCN then the memory controller is where I would focus.

 

[Mauller]

I know, AMD said they doubled the bus width of a already power hungry memory controller. So if its Still "Vega" ie GCN then the memory controller is where I would focus.

Power Hungry? What makes you think that?

 

[Nasher]

AMD cards have nothing built in which measures the power usage, so it's hard to get an accurate reading. The Vega was around the same as the 1080 afaik.

 

[D.P.]

not as much as you think, they don't use dedicated cores like Nvidia. Their DP support is more like SMT in that it requires a few % more transistors per pair of ALU's.

but there is still more to the hardware for the die size compared to adding DP alone.

It is more than a few percent, closer to 5-7%.

The real problem I'd if you issue DP ops then the total system throughout actually decreased. In mixed precision work, which basically everything is,then having dedicated DP units instead increased total OPS.And the DP units are trivially used as SP units. The issue is that extra 5% or more doesn't scale well. Once you have thousand of SP units that extra logic means you have.missed out on hundreds of dedicated units.

I wouldn't at all be surprised if AMD have moved in this direction. You gain a lot of efficiency at those cost of flexibility. I also wouldn't be surprised if AMD follow the logical next step and have their own dedicated mixed-precision tensor cores.
Vega 20 is aimed at HPC and DL, it won't compete well in the latter without an answer to Tensor cores

 

[D.P.]

AMD cards have nothing built in which measures the power usage, so it's hard to get an accurate reading. The Vega was around the same as the 1080 afaik.

You don't need to, you measure from the PSU in PCI port.

Vegas used much more power than even the 1080ti. It'd performance per watt is actually much closer to Maxwell on 28nm

 

[KillBoY_UK]

Power Hungry? What makes you think that?

The analyst by Gamers nexus and Buildzoid, we always knew AMD chips have been abit hungry but which bits where more power hungry than others I/we did'nt really know for a while

Speaking with Buildzoid, we know that Vega: Frontier Edition’s 16GB HBM2 pulls 20W max, using a DMM to determine this consumption. This ignores the voltage controller’s 3.3v draw, but we’re still at 20W memory, and no more than an additional 10W for the controller – that’s less than 30W for the entire memory system on Vega: Frontier Edition.

We also know that an RX 480 uses 40-50W for its 8GB, which is already a significant increase in power consumption per-GB over Vega: FE. The RX 480 also has a memory bandwidth of 256GB/s with 8GB GDDR5, versus Vega 64’s 484GB/s. The result is increased bandwidth, the same capacity, and lower power consumption, but at higher cost to build. In order for an RX 480 to hypothetically reach similar bandwidth, power consumption would increase significantly. Buildzoid calculates that a hypothetical 384-bit GDDR5 bus on Polaris architecture would push 60-75W, and an imaginary 512-bit bus would do 80-100W

 

[Rroff]

Vega never really had a chance in that respect - the GF 14nm process trailed behind even the Samsung version never mind the equivalent process elsewhere combined with the software tricks it would have used to significantly increase efficiency never coming to fruition for a range of reasons including AMD never putting the support behind them needed and/or not attractive to developers who'd have to add some extra setup functions to their engines.

 

[Mauller]

The analyst by Gamers nexus and Buildzoid, we always knew AMD chips have been abit hungry but which bits where more power hungry than others I/we did'nt really know for a while

Vega when the voltages are tweaked can be quite a bit more efficient. A friend has FE cards that can do 1600 @ 1V solid.

You also said the memory controller was power hungry then demonstrated the opposite.

 

[Mauller]

Vega never really had a chance in that respect - the GF 14nm process trailed behind even the Samsung version never mind the equivalent process elsewhere combined with the software tricks it would have used to significantly increase efficiency never coming to fruition for a range of reasons including AMD never putting the support behind them needed and/or not attractive to developers who'd have to add some extra setup functions to their engines.

The node didn't help, but the at hitecture is capable of 1800MHz + clocks. The reasons behind the other problems you mention are also not what people think.

 

[Rroff]

The node didn't help, but the at hitecture is capable of 1800MHz + clocks. The reasons behind the other problems you mention are also not what people think.

The architecture might be capable of it but the node is (was - not sure if it is still true) around 20% worse than the Samsung one for range of clocks at a given power level.

 

[Mauller]

The architecture might be capable of it but the node is (was - not sure if it is still true) around 20% worse than the Samsung one for range of clocks at a given power level.

12nm woul likely allow better variance so they could bin at lower voltages. It soul likely allow higher max clocks. But Vega on 7nm should be very interesting for a number of reasons.

 

[D.P.]

Vega when the voltages are tweaked can be quite a bit more efficient. A friend has FE cards that can do 1600 @ 1V solid.

You also said the memory controller was power hungry then demonstrated the opposite.

But you can also tweak Pascal clocks and voltages to get even more performance per watt.

Peoplr making this argument compare Vega's power with the 1080ti but the performance with the 1080.

In reality the 1080 used less power and I'd still5-8% faster. You can down clock and under volt the 1080 a little to get similsr performance to Vega at a fraction of the power

 

[Rroff]

But you can also tweak Pascal clocks and voltages to get even more performance per watt.

Peoplr making this argument compare Vega's power with the 1080ti but the performance with the 1080.

In reality the 1080 used less power and I'd still5-8% faster. You can down clock and under volt the 1080 a little to get similsr performance to Vega at a fraction of the power

If you slightly underclock - around 3-5% Vega you can drop the voltage loads on a lot of the cards (but not all) with a big drop in wattage - they seem to have (probably due to yields) put in a lot of overhead. The node seems to do really well at the lower end of the scale with ultra low voltage operation giving actually pretty decent clocks relative to the voltage - but it doesn't seem to hold up going the other way with massively more power used as you turn it up (I'm guessing the design focus was for ultra mobile stuff, etc.).