UPDATED* AMD announces the Radeon VII - but it's $699 | Nvidia calls it "Lousy"

[Taf the Ghost]

1 minute ago, Humbug said:

LOL If AMD doesn't Nvidia will by then just launch the Volta replacement high end at like $1500 maintaining the same price/perf as a the previous gen and calling it a new generation.

Volta is HBM2-based. We don't know if the Volta replacement will be HBM2 as well. Turing took a lot from Volta, but it's not quite the same. Nvidia, several generations ago, switched to a Compute First approach. (Please ignore they should have had Async several generations ago...) So whatever is coming in 7nm gaming cards is going to be seen this year in their next Compute card. Which, if we do see it in Retail, is going to be more expensive than the RTX Titan.

 

The name is probably Ampere, but we don't know if that's the gaming product or the compute one. At least publicly. (I have a suspicion that Turing/Ampere were the same project at one point, which Nvidia branching when they had a better appreciation for the 7nm node situation back in 2016. It's part of why there was a dual rumor mill around the post-Pascal product for a while.)

[Tieox]

Can the cards fans rotate with a 1/4 pounder cheeseburger on them?

 

 

IMHO, as I have a lot of catching up to do on tech more competition is brilliant.  Best of luck AMD

[porina]

I'll pass.

 

[mr moose]

38 minutes ago, porina said:

I'll pass.

 

So forgive my ignorance (I really haven't looked closely at the vega as it doesn't interest me), but does this mean when people were trying to argue this was a better deal than a 2080 (because RT etc features were as yet unrealized), were doing so on the assumption that this new vega had better performance than it does?

[riklaunim]

1 minute ago, mr moose said:

So forgive my ignorance (I really haven't looked closely at the vega as it doesn't interest me), but does this mean when people were trying to argue this was a better deal than a 2080 (because RT etc features were as yet unrealized), were doing so on the assumption that this new vega had better performance than it does?

FP64 is used in some computation, not gaming. If Radeon 7 offered max FP64 performance it would be equal to Radeon Instinct Mi 50 - a ~3000 USD Card and would kill it sales (plus would dump on Nvidia Titan V which is also a 3K USD computation GPU). Consumer variants of chips used in compute cards have FP64 nerfed, that's the standard (but still FP64 of Radeon 7 is higher than like 2080 Ti - and only specific applications use FP64 - mostly scientific calculation and alike.

[mr moose]

14 minutes ago, riklaunim said:

FP64 is used in some computation, not gaming. If Radeon 7 offered max FP64 performance it would be equal to Radeon Instinct Mi 50 - a ~3000 USD Card and would kill it sales (plus would dump on Nvidia Titan V which is also a 3K USD computation GPU). Consumer variants of chips used in compute cards have FP64 nerfed, that's the standard (but still FP64 of Radeon 7 is higher than like 2080 Ti - and only specific applications use FP64 - mostly scientific calculation and alike.

The people who were arguing with me were trying to insinuate that this vega brought something to the table that made up for the lack of AI and RT hardware. It seems they were assuming it's FP64 performance was bigger than what it was.

[Taf the Ghost]

18 minutes ago, mr moose said:

So forgive my ignorance (I really haven't looked closely at the vega as it doesn't interest me), but does this mean when people were trying to argue this was a better deal than a 2080 (because RT etc features were as yet unrealized), were doing so on the assumption that this new vega had better performance than it does?

Double Precision (FP64) really only matters in a few scientific workloads. Somehow some fake news got around that it was the full Compute capability of the normal MI50. AMD was about to sell out their entire supply of cards to universities by reason of false information. 

 

This does mean, for FP64 tasks, this card is almost 250% more powerful than a RTX Titan, though some of the old Hawaii cards are still floating around and they do pretty well, comparatively. Actually, looking back, normally the "Compute Card as high-end GPU" gets 1:8 from AMD.

[Taf the Ghost]

4 minutes ago, mr moose said:

The people who were arguing with me were trying to insinuate that this vega brought something to the table that made up for the lack of AI and RT hardware. It seems they were assuming it's FP64 performance was bigger than what it was.

For a high-end GPU, which really don't sell in huge numbers, the full FP64 ability, would make it the premiere choice for Universities around the world. Professors & Researchers could swing the 700USD on their budget without much issue. With only 1:8, it still will have that use given it's performance per price, but it won't be quite the hot commodity in that space.

[TheDankKoosh]

I'd like to see a slightly cut down version of this with 56 cu's and 8gb of hbm 2 (still in quad stack) for $550 so that way they have a much better 2080 competitor. Since vega has the front end bottleneck it won't affect performance in games and would only slightly affect compute workloads that don't need massive amounts of memory.

[Canada EH]

On 1/10/2019 at 12:38 AM, DrMacintosh said:

 Nice attempt at controlling the narrative though. 

It is an artform, that is why I like watching CNN sometimes.

[CarlBar]

3 hours ago, mr moose said:

The people who were arguing with me were trying to insinuate that this vega brought something to the table that made up for the lack of AI and RT hardware. It seems they were assuming it's FP64 performance was bigger than what it was.

 

Info has come out indicating that the instinct line at least does have support in it's hardware for machine learning stuff. Weather thats going to port over to this card is unknown, but don't rule it out yet.

[mr moose]

18 minutes ago, CarlBar said:

 

Info has come out indicating that the instinct line at least does have support in it's hardware for machine learning stuff. Weather thats going to port over to this card is unknown, but don't rule it out yet.

But it's still people using an unknown to argue an equivalent without actually having anything to compare.

[riklaunim]

5 hours ago, mr moose said:

The people who were arguing with me were trying to insinuate that this vega brought something to the table that made up for the lack of AI and RT hardware. It seems they were assuming it's FP64 performance was bigger than what it was.

DirectML? But in general it's a standard GPU without any specialized segments. The die is small and that's it. It's not a new generation and it's not even designed to compete on the field of raytracing, AI and so on.

[porina]

7 hours ago, Taf the Ghost said:

Double Precision (FP64) really only matters in a few scientific workloads. Somehow some fake news got around that it was the full Compute capability of the normal MI50. AMD was about to sell out their entire supply of cards to universities by reason of false information. 

Ryan Smith, Editor in chief of Anandtech, had previously tweeted it had full FP performance, based on what he was told by someone from AMD at the time. Incorrect it may be, but he sourced it from AMD. He also got the latest statement, from someone at AMD who hopefully knows more than his first contact.

[Madgemade]

9 hours ago, porina said:

Well this is good news.

According to GPU-db Radeon 7 has 1,728 FP64 with this ratio.

Compared to the 2080 Ti which has 420 GFLOPS FP64. The Titan RTX has 509 GFLOPS. As has been said this puts the Radeon in a good spot, takes sales from Nvidia where buyers want FP64, but doesn't have much impact on their high end cards which still offer more FP64 performance at a pro price.

 

It's important to remember that FP64 Double Precision is irrelevant for gaming. The better this card is at it, the more likely it is to be bought up by Pro users, universities etc. Which will be good for AMD but maybe not for gamers as it isn't going to do anything for prices except push them up. I am (at least for now) a pro user because I have some programming projects which use FP64. This GPU is the fastest consumer FP64 card AMD has produced since the 7970 and it's rebrands which had 1024 GFLOPS. So I might well buy it even though it's not a Double Precision beast it's still a good bit better than everything else under $3000+.

[porina]

9 minutes ago, Madgemade said:

Well this is good news.

According to GPU-db Radeon 7 has 1,728 FP64 with this ratio.

Compared to the 2080 Ti which has 420 GFLOPS FP64. The Titan RTX has 509 GFLOPS. As has been said this puts the Radeon in a good spot, takes sales from Nvidia where buyers want FP64, but doesn't have much impact on their high end cards which still offer more FP64 performance at a pro price.

 

It's important to remember that FP64 Double Precision is irrelevant for gaming. The better this card is at it, the more likely it is to be bought up by Pro users, universities etc. Which will be good for AMD but maybe not for gamers as it isn't going to do anything for prices except push them up. I am (at least for now) a pro user because I have some programming projects which use FP64. This GPU is the fastest consumer FP64 card AMD has produced since the 7970 and it's rebrands which had 1024 GFLOPS. So I might well buy it even though it's not a Double Precision beast it's still a good bit better than everything else under $3000+.

Is there a street price for MI50 for comparison? In a quick search I've drawn a blank. I'm thinking those that care about FP64 wont necessarily want a single cheap card, but want a certain performance level overall, single or multiple cards. In this sense, I'd like to try a DP/$ as well as DP/W estimate. Traditionally I've built up on CPUs to deliver FP64 (especially with AVX2, with AVX-512 support building) in my niche.

 

My feeling for now is, AMD picked this exact rate to give a bit more FP64 for those who want it on the side, but unattractive enough for those who need it exclusively to preserve their higher end offerings.

[Madgemade]

5 minutes ago, porina said:

My feeling for now is, AMD picked this exact rate to give a bit more FP64 for those who want it on the side, but unattractive enough for those who need it exclusively to preserve their higher end offerings.

Exactly that.

There is no street price for any MI50 or MI60 card that I know of. I think they're only available to bulk buyers in server farms and such. Maybe we will see better availability once this launches. The MI25 which has the same FP64 as Vega is hard to find a price for but it looks like it is at least $5000 so I would expect MI50 and MI60 to be close to $10,000!

 

We know the MI50 has 4x the FP64 because it uses at 1:2 ratio at almost the same clocks. But there is no way the MI50 is going for $2796 = 4x$699 it is at least twice or three time that. But for server use it is obviously better to have one card at 300W than it is to have 4 at 1200W and 4 times the space used and heat output.

 

For Double Precision it only makes sense for those who are experimenting and want better performance than the 1:32 ratios of Nvidia and AMDs normal 1:16 ratios, but don't actually need lots of performance and don't have a big budget.

8 hours ago, Taf the Ghost said:

For a high-end GPU, which really don't sell in huge numbers, the full FP64 ability, would make it the premiere choice for Universities around the world. Professors & Researchers could swing the 700USD on their budget without much issue. With only 1:8, it still will have that use given it's performance per price, but it won't be quite the hot commodity in that space.

Exactly this.

[porina]

4 minutes ago, Madgemade said:

There is no street price for any MI50 or MI60 card that I know of. I think they're only available to bulk buyers in server farms and such. Maybe we will see better availability once this launches. The MI25 which has the same FP64 as Vega is hard to find a price for but it looks like it is at least $5000 so I would expect MI50 and MI60 to be close to $10,000!

Ok, that blows the MI50 out of consideration for a FP/$ metric. The Titan V hangs in there, with Radeon 7 about 20% better bang for buck on up front cost. A concern for the Titan V is its ram bandwidth may be more of a limiting factor.

 

I also did a rough estimate based on an i3-8100 and i7-7800X, CPU only, excludes other components required. They're still competitive in that sense, but of course would fall way behind once you do add necessary system parts. They're also competitive on bw/flop. Not enough info to do a good estimate on performance/W but it is probably on the lower end along with the 7, with the Titan V and MI50 in another league.

[cj09beira]

18 minutes ago, porina said:

Ok, that blows the MI50 out of consideration for a FP/$ metric. The Titan V hangs in there, with Radeon 7 about 20% better bang for buck on up front cost. A concern for the Titan V is its ram bandwidth may be more of a limiting factor.

 

I also did a rough estimate based on an i3-8100 and i7-7800X, CPU only, excludes other components required. They're still competitive in that sense, but of course would fall way behind once you do add necessary system parts. They're also competitive on bw/flop. Not enough info to do a good estimate on performance/W but it is probably on the lower end along with the 7, with the Titan V and MI50 in another league.

for workstation use, the Mi line isn't it, you should be looking at the Wx series instead 

[porina]

50 minutes ago, cj09beira said:

for workstation use, the Mi line isn't it, you should be looking at the Wx series instead 

I'm looking at FP64 performance. Wx is comparable to consumer cards and doesn't come close to MI series.

[Madgemade]

1 hour ago, porina said:

I'm looking at FP64 performance. Wx is comparable to consumer cards and doesn't come close to MI series.

The AMD FirePro W9100 is the last GPU from AMD to have 1:2 FP64 @ 2,619 GFLOPS. It also had display outputs the FirePro S9170 and S9100 are identical but are passively cooled (for servers) and have no display outputs. The W9100 is still available for around $2000 new and maybe half that used.

 

Since the launch of Hawaii AMD did not release any GPU (workstation or consumer) that had 1:2 ratio FP64 up until the MI50/MI60 launch in November. The MI25 which is first gen Vega has the same FP64 as Vega 64.

 

Before GCN 2.0 AMD always used decent FP64 ratios on all it's consumer GPUs. For example:

The 5870 had a ratio of 1:5 and had 544 GFLOPS.

The 6970 had a ratio of 1:4 and had 676 GFLOPS.

The 7970 had a ratio of 1:4 and had 947 GFLOPS.

The R9 280X had a ratio of 1:4 and had 1024 GFLOPS. (This was a rebranded and refined 7970 still based on GCN 1.0)

 

I'm glad AMD has got some FP64 chips out because for 4 years they had nothing much.

[cj09beira]

2 hours ago, porina said:

I'm looking at FP64 performance. Wx is comparable to consumer cards and doesn't come close to MI series.

as the post above said thats more due to the chips not having fp64 than a segmentation issue, hopefully they release a Wx card based on the vega 20 gpu soon 

[porina]

16 minutes ago, cj09beira said:

as the post above said thats more due to the chips not having fp64 than a segmentation issue, hopefully they release a Wx card based on the vega 20 gpu soon 

They do have FP64, just not a lot of it. It is a segmentation decision to include or not include a given performance level. We had decent relative FP64 performance up to... whatever the nvidia 500 series was, and the last AMD card I had with decent FP64 was the 280X. It was downhill from there as the 28nm stagnation kicked in, and they nerfed FP64 on consumer devices to redirect that towards expanding other resource. Was kinda hoping the move to 7nm would be a return to those old days, but I guess not.

 

The way things are going, Zen2 might be better value than their GPUs in this area.

[CarlBar]

31 minutes ago, porina said:

They do have FP64, just not a lot of it. It is a segmentation decision to include or not include a given performance level. We had decent relative FP64 performance up to... whatever the nvidia 500 series was, and the last AMD card I had with decent FP64 was the 280X. It was downhill from there as the 28nm stagnation kicked in, and they nerfed FP64 on consumer devices to redirect that towards expanding other resource. Was kinda hoping the move to 7nm would be a return to those old days, but I guess not.

 

The way things are going, Zen2 might be better value than their GPUs in this area.

 

As has been said the issue is server side GPU's. 9+ years ago GPU's hadn't really made a major dent in tehr server compute market, so including full compute capability didn't really affect anything much on the server side sales. Now a full 1:2 ratio would tear their server side market sales of MI50 and even MI60 to pieces, they just can't afford it. Any retail part has to have a significantly lower FP64 that their MI50 and MI60 cards.

[YoloSwag]

They could have fully enabled all features but that would mean they lose sales on their higher margin products. Compute cards don't have movements in units that are huge like in the gaming market but the price margins they have are. Releasing a card that can do well on both markets just spells trouble. For this reason they are making sure premium products stay premium. It's sad. But there you have it.