Vega/Navi Rumors (Updated)

[KompuKare]

Also people aren't dumb, and those looking for more performance can buy an Nvidia card despite what the stock clock speeds are because they know there's more performance left under the hood.

That was true for big Maxwell 980 Ti, but it's far less the case with big (gaming) Pascal 1080 Ti.
As it happens for overclocking W1zzard at TPU lists the clocks of all the variants reviewed.
Taking the best performing 980Ti and 1080Ti, we get this

Source: https://www.techpowerup.com/reviews/ASUS/GTX_980_Ti_Matrix/26.html
Max vs stock was +34.5%.

Source: https://www.techpowerup.com/reviews/NVIDIA/GeForce_GTX_1080_Ti/33.html (yes, for TPU their FE's edition was a slight bit quicker than their faster 3rd party, the Palit).
Stock vs max is +15.7

So while AMD's card ship far closer to their max (at least recently, both 7950 and R9 290 had plenty of overclocking headroom), Nvidia are now a lot closer to their max than they were with the 980 Ti.

 

[Glo.]

https://videocardz.com/69475/amd-radeon-vega-spotted-with-16gb-memory-and-1600-mhz-clock

 

[.vodka]

1.6GHz, that sure looks good for GCN. Now, will this be like on Tahiti, that had another 200Mhz of headroom... or is it another Fiji, already clocked to its maximum out of the box?

Either way, this verifies the ~1.5GHz speed of the MI25 and its 4096SPs. Now... 16GB HBM2 on two stacks? Is that possible? 8-hi stacks?

 

[Krteq]

Hmm, according to DeviceID it's a Vega 10 based card that lies right in the middle - performance wise.

{0x1002, 0x6860, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x6861, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x6862, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x6863, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x6864, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x6867, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x6868, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x686c, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},
{0x1002, 0x687f, PCI_ANY_ID, PCI_ANY_ID, 0, 0, CHIP_VEGA10},

The last one with DevID 687F is that ES we know about since January.

Still waiting for results of top model with DeviceID 6860

 

[redzo]

Acording to this article:
https://www.techpowerup.com/233325/entire-amd-vega-lineup-reportedly-leaked-available-on-june-5th
there is going to be a huge price gap between the cheapest vega and the rx 580(8gb).
I was hopping that AMD will come up with a direct gtx 1070 competitor <= price wise.

 

[Krteq]

Isn't that article based on removed reddit post?

 

[antihelten]

Isn't that article based on removed reddit post?

Yep:

https://www.ceddit.com/r/Amd/comments/6aeat3/full_vega_lineup_and_release_date_revealed/

 

[raghu78]

1.6GHz, that sure looks good for GCN. Now, will this be like on Tahiti, that had another 200Mhz of headroom... or is it another Fiji, already clocked to its maximum out of the box?

Either way, this verifies the ~1.5GHz speed of the MI25 and its 4096SPs. Now... 16GB HBM2 on two stacks? Is that possible? 8-hi stacks?

Yes 16GB is possible with just two 8 Hi stacks. I think the Radeon Instinct MI25 will use 8 Hi stacks (16GB) while the consumer graphics Radeon Rx Vega will use two 4 Hi stacks (8GB). As for headroom I hope AMD's Vega is designed for significantly higher clock speeds (atleast >= 30%) than Polaris.

 

[IEC]

Isn't that article based on removed reddit post?

Not to mention, those product names sound like a CEO told his underlings: Leak this name to X guy, and this name to Y guy. See which name ends up in the media and bring them before me. "You're fired!"

 

[Glo.]

https://videocardz.com/69504/sk-hynixs-updated-memory-catalog-features-hbm2-and-gddr6

 

[Bacon1]

Acording to this article:
https://www.techpowerup.com/bs
there is going to be a huge price gap between the cheapest vega and the rx 580(8gb).
I was hopping that AMD will come up with a direct gtx 1070 competitor <= price wise.

Man TPU is falling behind on the reporting. That was removed a week ago.

 

[ElFenix]

Nvidia is selling products at particular performance TDP envelope, whereas AMD is simply chasing Nvidia at the performance metric only. Nvidia is, at the moment, selling a product with more features. This is why you've got a GTX 1080 85% faster than AMD's RX 480 while not using that much more power. If you don't understand why Nvidia does this, let me give you a hint--majority of these cards will be put in OEM systems, and OEMs are very interested in TDP because that will eat into their bottom line with their custom mobos and possibly requiring higher power delivery specs or more expensive PSUs.


Also people aren't dumb, and those looking for more performance can buy an Nvidia card despite what the stock clock speeds are because they know there's more performance left under the hood.

That was true for big Maxwell 980 Ti, but it's far less the case with big (gaming) Pascal 1080 Ti.
As it happens for overclocking W1zzard at TPU lists the clocks of all the variants reviewed.
Taking the best performing 980Ti and 1080Ti, we get this


So while AMD's card ship far closer to their max (at least recently, both 7950 and R9 290 had plenty of overclocking headroom), Nvidia are now a lot closer to their max than they were with the 980 Ti.

These two posts are going a bit far afield for this topic. Stay on target!

AT Moderator ElFenix

 

[Saylick]

1.6GHz, that sure looks good for GCN. Now, will this be like on Tahiti, that had another 200Mhz of headroom... or is it another Fiji, already clocked to its maximum out of the box?

Either way, this verifies the ~1.5GHz speed of the MI25 and its 4096SPs. Now... 16GB HBM2 on two stacks? Is that possible? 8-hi stacks?

If you have to ask... it's probably the latter, unfortunately. I wouldn't keep my hopes too high on OC potential given that I'd rather have AMD clock these things high out of the gate to be competitive with NV than to leave performance on the table. As you probably know, launch review comparisons are >80% of what sets the expectation for the year that follows, regardless of any driver improvements. The masses simply don't research their purchases as deeply like we enthusiasts do.

 

[Malogeek]

Well the masses aren't usually enthusiast class GPU purchasers either.

 

[GoodRevrnd]

16GB seems like an insane amount for a consumer card. So if it's a pro card, perhaps there's more mhz to be found in the consumer version? Alternatively, maybe it is the consumer card and 1600mhz is pushing the limit and they slapped in 16GB for the spec sheet. All assuming this is a valid rumor.

 

[Saylick]

Well the masses aren't usually enthusiast class GPU purchasers either.

Fair point, but depends on where you draw the line for enthusiast class GPUs. One can argue that it's everything above $300, which is well in the affordability range of your average gamer.

 

[redzo]

Isn't that article based on removed reddit post?

I had/have no idea. Got posted on TPU. Got it almost as real news. Apologies, I can't follow this thread accordingly: not enough spear time due to work: WannaCry is upon us.
... so I guess that I can still hope for AMD pushing for a cheaper 1070 vega gpu?

 

[IEC]

I had/have no idea. Got posted on TPU. Got it almost as real news. Apologies, I can't follow this thread accordingly: not enough spear time due to work: WannaCry is upon us.
... so I guess that I can still hope for AMD pushing for a cheaper 1070 vega gpu?

I would hope that the lowest end Vega still solidly beats a GTX 1070. AMD needs all the competitive edge they can get given how absent they have been from the high-end!

 

[Yakk]

Any guesses at mining performance with these? I doubt they'll be much use, but it's the only way I'd justify spending that much in a GPU anymore.

 

[n0x1ous]

Any guesses at mining performance with these? I doubt they'll be much use, but it's the only way I'd justify spending that much in a GPU anymore.

probably around 30-35 MH/s. I doubt they do much better than Fiji since mining is based more on memory speed latency and the HBM bandwidth doesnt do it any good there

 

[Yakk]

probably around 30-35 MH/s. I doubt they do much better than Fiji since mining is based more on memory speed latency and the HBM bandwidth doesnt do it any good there

Thanks, it's what I thought. Just checking; since I've converted from mining to trading I kinda lost track after selling everything. I would've liked to tinker with a Vega, but it's hard to justify spending that much after years and years of GPUs generating profit. Looks like my leftover 480 CF will be good enough for me for a long while yet it looks like from the forthcoming games lineup.

 

[Samwell]

1600mhz Clockspeed looks great. Seems the only question is now about Hynix with two possibilities:
1. They keep secret that they can produce 2Gbps Memory and 8 High Stacks, then Vega should beat GP102 with 1600mhz.
2. They don't publish it in their datasheet because they have manufacturing problems and the 16Gb Vega in Compubench is just one of the few working samples with 8 High Memory. Slower Memory might limit Vega quite a bit, but otherway they amd can't launch or amd can only launch in very limited quantities.

Fanboys know their sides and the rest of us can speculate
I hope for 1, as it would be pretty sad if amd had a good architecture, but would be strongly limited by hynix.

 

[piesquared]

Unless i missed something, 4 4 high stacks for 1 TB/s would be preferred over 2 8 high stacks for 512 GB/s? The cost of 4 high stacks would presumably be cheaper than 8 high stacks as well. The interposer would have to be slightly larger, but i can't imagine that being the deal breaker.

 

[DisEnchantment]

I find VCZ to be far more reliable than TPU or WCCF.
Taking their values, for the 16GB 6864 HBM part, it would come to 820 GB/s which is respectable. But I would suppose this would be the Instinct SKU.

I would suppose a top Consumer card to be 8-12 GB using 4-12 Hi Stacks for 410 - 615 GB/s considering Koduri mentioning considerable flexibility with the Memory for various SKUs.

 

[Valantar]

Unless i missed something, 4 4 high stacks for 1 TB/s would be preferred over 2 8 high stacks for 512 GB/s? The cost of 4 high stacks would presumably be cheaper than 8 high stacks as well. The interposer would have to be slightly larger, but i can't imagine that being the deal breaker.

Sure, it would be preferable, but it would require two more memory interfaces on the chip (which it's unlikely they added for a single low-volume SKU), a bigger interposer, and all the added part and manufacturing cost that entails. I'd say keeping the interposer as small as possible was a priority for AMD in this case. Makes sense.