AMD 7nm GPU News and Rumours 2018/2019

[Satchfanuk]

It is a fair point. Looking at the flops, 7nm should be better than 12nm surely and big business isn't going to quibble over prices like that for time and efficiency I expect (not that I really know what big businesses do)

Yeah but for the same money as the RTX6000 (ie 4 x MI60), AMD are hitting almost 59 TFlops

Surely some big biz would like that

 

[bru]

RTX6000 $6300 active cooling workstation card
MI60 ~ $1500 passive cooling datacenter card.

Just out of interest, where did you find that price for the MI60, as I wasn't aware that AMD had released a price yet.

 

[Rroff]

Just out of interest, where did you find that price for the MI60, as I wasn't aware that AMD had released a price yet.

Given that the cost of the current AMD solution in that segment is approx. $5000-$11,000 depending on the installation and quantity I think he probably missed a zero off.

EDIT: Also something he failed to notice is that workstation cards usually have a fan for a reason - datacenter cards (and nVidia are no exception) don't tend to as the installation usually involves the card having a design that allows air flow right through a big fin array and an array of high performance fans in the chassis that push air through the case in a designed flow. If you look at nVidia's higher end datacenter cards like the Tesla V100 PCIe yup passive cooled.

 

[jigger]

RTX6000 $6300 active cooling workstation card
MI60 ~ $1500 passive cooling datacenter card.

Very interesting.

 

[LePhuronn]

We do not have any proof of this at all. its all speculation atm. We have Polaris and VEGA now why wouldn't we have a Navi replaces Polaris and a VEGA refresh on 7nm?

Because the stuff added to Vega 20 isn't for gaming and AMD don't have the money to respin Vega 10 onto 7nm to fill a 9 month gap until Navi arrives. Also, AMD have always said that Navi is a "mid-tier" product so it's a strong assumption that Navi 10 would be at or slightly above GTX 1080 performance, which is probably where a respun Vega 10 would end up anyway. So that either makes a respun RX Vega redundant or RX Navi redundant.

Furthermore, because AMD aren't swimming with cash, I can't see Polaris 30 existing purely as 1 single card as the RX 590. Leaked figures show the RX 590 beating the GTX 1060 so we potentially have a new budget gaming king. The existing 580 and 570 cards could be moved to Polaris 30 for a slight performance and efficiency boost and offer a solid presence in the entry level market.

So with Navi occupying the mid tier, Polaris 30 occupying the entry tier and AMD not having anything right now to occupy the top tier, where exactly would a respun Vega sit? Because however much we'd like it to happen, just shrinking Vega 10 to 7nm is not going to double its performance and take on the GTX 1080 Ti or RTX 2080 to become a top-tier card.

 

[Panos]

Just out of interest, where did you find that price for the MI60, as I wasn't aware that AMD had released a price yet.

Out of interest why you mix workstation and datacenter card? I thought you were better than this

As @Rroff correctly said, the MI60 going to be north of $5000, the equivalent NVidia solution is $9300




RTX6000 is not same market as the MI60. The Vega20 workstation card would be priced similarly to the Vega based WX9100 ($1300) card
That is why I wrote $1500, could be 1700-1800. But not $6000 the RTX6000 costs

 

[bru]

Out of interest why you mix workstation and datacenter card? I thought you were better than this

Wasn't aware of a Turing based Tesla card.

But yes you are correct the new, just announced 7nm MI60 card is just faster 14.7 TFlops than the year old last gen 12nm Tesla V100 at 14 TFlops, unless you look at the NVlink Variant of course.

 

[dragonslayer]

Why are people assuming that "mid range" navi is going to be 1080 level performance for all we know AMD could have been aiming at mid range for Nvidias Next gen(20xx series) since Vega64 already has the same peformance as a 1080 it would make navi rather pointless and not worthy of an upgrade for anyone

 

[Nasher]

But remember....the more you buy, the more you save.

 

[Gerard]

But remember....the more you buy, the more you save.

And the more you save the more you spend in postage for repeated rma's

 

[Lokken86]

But remember....the more you buy, the more you save.

It's just like, common sense you know.

 

[Panos]

Wasn't aware of a Turing based Tesla card.

But yes you are correct the new, just announced 7nm MI60 card is just faster 14.7 TFlops than the year old last gen 12nm Tesla V100 at 14 TFlops, unless you look at the NVlink Variant of course.

Last page I posted some matrix computing benchmarks. All the Nvidia Volta advertisement about 100Tflop at FP16 and dedicated Tensor cores, cannot be materialized and the reality is that AMD competes on the Datacenter section on even footing with the 14/12nm using the ancient ROCm 1.3. Now with the V20 is 6-10 months ahead (Financial analysts saying that), and yet we haven't seen the improvements of ROCm 2.0 brings on the table.

Workstation cards probably completely different matter, yet again, someone could buy 4 WX9100s (Vega 10) for a single RTX6000 atm. True someone could buy 4 RTX6000 also if money doesn't matter.

However consider this. I use my Vega 64 on Tensorflow (with ROCm 1.9.1), and is great card to work with.
Is just around 18-13% (depending usage) slower (undervolted and overclocked) than the $9000 TV100, for 1/20th the price. A bargain not only for me but many small/medium businesses and especially enthusiasts.
I know a company in York which is buying even today used V64s and is using them for work projects (it is using MS AI) because it is great on number crunching and dirty cheap compared to professional cards.

I do not remember if it was AdoredTV or the guy at Anandtech but around May time, he said that actually the majority of the Vega 64s was grabbed and set straight to the entertainment industry not on mining rigs.
You see can still use the professional tools on consumer cards, it as AMD doesn't have licences etc and all is open source.

That is why AMD was pointing again that there is no licencing, and everything is open source.

 

[Panos]

But remember....the more you buy, the more you save.

Some times I wonder if the guy wasn't drunk when said it, thinking is working for AMD...... Probably subconsciously in his nightmares....

 

[LePhuronn]

Why are people assuming that "mid range" navi is going to be 1080 level performance for all we know AMD could have been aiming at mid range for Nvidias Next gen(20xx series) since Vega64 already has the same peformance as a 1080 it would make navi rather pointless and not worthy of an upgrade for anyone

Why do people always assume it'll be purchased as an upgrade? I don't have a Vega, I don't have a Polaris. Hell, right now the only thing I have is a 6670 and an original GTX Titan in bits. So Navi would be a brand new card for me.

How would you define "mid range" then? Even if you think mid-range RTX series then you're looking at the RTX 2070 which, without Tensor and RT, is about GTX 1080 performance. Surprising that. Now yes, Vega 64 trades blows with the GTX 1080 so in terms of performance then you're right, Navi 10 won't see a performance increase. But Navi 10 is supposedly a much, much smaller die so it will be much more power efficient than Vega, it'll be much cheaper to produce than Vega and with a wee bit of luck it might even use GDDR6 which is cheaper to implement than HBM.

So performance that probably matches or slightly exceeds the RTX 2070, power draw on par with a GTX 1080 at around 2/3 the price? Yes please.

 

[ubersonic]

Some times I wonder if the guy wasn't drunk when said it

Unlikely, it actually makes complete sense in context.

The point was (IIRC) that savings are multiplied with sales, I.E if you buy two cards for SLI you will save twice as much.

Guys still a **** tho ^^

 

[Panos]

Unlikely, it actually makes complete sense in context.

The point was (IIRC) that savings are multiplied with sales, I.E if you buy two cards for SLI you will save twice as much.

Guys still a **** tho ^^

As someone wrote on youtube

You see that's why you should buy more of them. To make sure you have some workforce redundancy.

In case Joe RTX2080 decides not to come in to work on Monday, Bob RTX2080 can still handle enough of Joe's duties for the day.
And really you know, the more workers you have the more you save on unexpected absent employees.

 

[ubersonic]

As someone wrote on youtube

Now that guy looks drunk, lol.

 

[Panos]

Now that guy looks drunk, lol.

 

[Lokken86]

Jensen does talk some amount of rubbish haha.

 

[D.P.]

Just out of interest, where did you find that price for the MI60, as I wasn't aware that AMD had released a price yet.

AMD have not released a price because it is never going to be available for consumers, it is being sold directly to datacenters in bulk purchases.

It will probably cost somewhere in the $6-10,000 range.


But the real issue is AMD aren;t even comparing like for like. The datacenter version of the VoltaTV100 is the NVLink version, with higher raw performance than the MI60.