Vega/Navi Rumors (Updated)

[Stuka87]

Fury needs all 512GB/s to perform. Vega is fury at significantly higher clocks hence it needs a proportional increase in bandwidth to perform as well as Fury clock for clock.

Not entirely true, as Vega does have better compression than Fury, assuming it is actually working in the release drivers.

 

[Despoiler]

Nvidia was likely paying 2.5x more to fulfill their ORNL order. I doubt 2.5x is still or will remain the price for very long now that we have at least two vendors increasing production.

 

[mohit9206]

So the reviews come out on Aug 14? Should be quite an embarrassing day for AMD.




Threadcrapping isn't allowed.


esquared
Anandtech Forum Director

 

[Elixer]

So the reviews come out on Aug 14? Should be quite an embarrassing day for AMD.

Embarrassing because AMD has no answer to the 1080Ti? Sure.
Embarrassing for gamers? Perhaps. Really can't tell without actual, fair reviews.
Embarrassing for miners? Not for Monero, as for the others, if you buy into the rumors, then, nope, not at all for them.
Embarrassing for non-gamers & non-miners? Remains to be seen.

 

[DDH]

AMD would have to clock down the GPU or have even lower voltages, since they couldn't afford the power envelope that GDDR5/X would be eating.
In other words, Vega would be much slower than it is now, if it used GDDR5/X.

As for HBM2, AMD is getting screwed on price (& performance) as well...
http://www.anandtech.com/show/11690/sk-hynix-customers-willing-to-pay-more-for-hbm2-memory

AMD really needs Samsung's HBM2 production.

"To bring Project 47 to life, AMD worked closely with Samsung Electronics with respect to the HBM2 memory used across the “Vega”-based product lines including the Radeon Instinct MI25 accelerators."

https://instinct.radeon.com/en-us/p47-announcement/

 

[french toast]

Vega suffers from 3 short term failings that should be mostly resolved in 12 months.
-Drivers and/or game optimisations.
-Non suitable low power process. (14lpp)
-Low quality/high voltage/low clocked HBM2, which is also still likely expensive and difficult to produce in large quantities.

If AMD had use of tsmc 16nm FF+ and also mature and expected performance HBM2 I wonder what the effect would be on Rx Vega? 30℅ better perf/watt?
Add in mature driver fine wine magic and game optimisations which utilise Vega specific features (2 x FP16, dsbr, primitive shaders, asynchronous shaders,etc) then it could probably match or beat a gtx 1080ti.

As it happens not tsmc, but 7nm late next year on a proper HP process, hbm 2 should improve greatly in one year and software is likely to make large gains.
Vega uarch might actually be a smart long play by AMD, they are forced to swallow a tough early pill for hopefully a competitive uarch in the long run.

 

[bassof]

....................



Since this was not an actual twitter link I will assume it fake.

It's deleted for trolling.


esquared
Anandtech Forum Director

 

[Krteq]

I'll take this as a fake until you post a link to those tweets.

 

[Veradun]

It's blatanly fake.

And if even if it wasn't obvious enough, reviews are done by reviewers the way they like, and companies give them written guidelines, they don't tweet instructions :>

 

[thecoolnessrune]

The only 3 tweets that page has made on Twitter have nothing to do with reviews. As @Krteq @Veradun stated, it's blatantly fake.

https://twitter.com/AMDRyzen/status/893562259972255744
https://twitter.com/AMDRyzen/status/893622633492774913
https://twitter.com/MAINGEAR/status/893613005799378944

 

[NostaSeronx]

Vega suffers from 3 short term failings that should be mostly resolved in 12 months.
-Drivers and/or game optimisations.
-Non suitable low power process. (14lpp)
-Low quality/high voltage/low clocked HBM2, which is also still likely expensive and difficult to produce in large quantities.

My list of three...
- Didn't launch with the 400 series.
- Wasted time going on FinFETs. (Every generation decreases electron mobility)
- HBM2's issue is that those that use it go onto shoddy nodes. HBM2 has an insane 99.9% good stack ratio for 1.6 GHz. (That is 1 to 100,000 good stacks being bad)

 

[maddie]

My list of three...
- Didn't launch with the 400 series.
- Wasted time going on FinFETs. (Every generation decreases electron mobility)
- HBM2's issue is that those that use it go onto shoddy nodes. HBM2 has an insane 99.9% good stack ratio for 1.6 GHz. (That is 1 to 100,000 good stacks being bad)

Where is this coming from?

 

[NostaSeronx]

Where is this coming from?

Industry sources...


HBM2 is at version 3 currently, "Perfected Yields", version 4 allows for half-channel => 512-bit, and 2.4 GHz op which is 307.2 gigabyte per sec or 153.6 gigabyte per sec,"Low Cost". 😉

Vega 12 w/ 512-bit HBM2... yum.

 

[Elixer]

"To bring Project 47 to life, AMD worked closely with Samsung Electronics with respect to the HBM2 memory used across the “Vega”-based product lines including the Radeon Instinct MI25 accelerators."

https://instinct.radeon.com/en-us/p47-announcement/

Not sure why you posted that, all I said was AMD really needs Samsung's production, not that they aren't using Samsung.

I would love to see a tear-down & voltages used for the MI25 accelerators, though, they also seem to be running at 1.89Gbps, not 2.0Gbps, which seems to indicate HBM2 is so far under performing.

Industry sources...


HBM2 is at version 3 currently, "Perfected Yields", version 4 allows for half-channel => 512-bit, and 2.4 GHz op which is 307.2 gigabyte per sec or 153.6 gigabyte per sec,"Low Cost". 😉

Vega 12 w/ 512-bit HBM2... yum.

So, what do these "sources" tell you on why they aren't running HBM2 @ 2.0 Gbps?

 

[french toast]

Industry sources...


HBM2 is at version 3 currently, "Perfected Yields", version 4 allows for half-channel => 512-bit, and 2.4 GHz op which is 307.2
Vega 12 w/ 512-bit HBM2... yum.

What voltage is Vega Rx HBM2 running at if you are in the know? It is supposed to be around 1.2v, is that correct? I'm sure I read on this forum Vega Fe HBM2 was running over spec at 1.35v?

I don't know what the expected performance of HBM2 was supposed to be for Vega but I'm quite sure it was to be at least 512gb/s and using much less power to do so.

 

[NostaSeronx]

So, what do these "sources" tell you on why they aren't running HBM2 @ 2.0 Gbps?

What voltage is Vega Rx HBM2 running at if you are in the know? It is supposed to be around 1.2v, is that correct? I'm sure I read on this forum Vega Fe HBM2 was running over spec at 1.35v?

- AMD missed the launch for 2 GHz. SK Hynix then prioritized DDR4/LPDDR4/GDDR5/GDDR6 over HBM2 because of low HVM demand. Aka, AMD ruined their own plans by not having a production line already in place. Aka, Fiji Redux like Bristol/Stoney on 28nm w/ only two HBM2 PHYs.
- AMD is overclocking the 1.6 GHz SKUs of last year. (Just a warning, SK Hynix halted HBM2 production to switch to ver 4.)
- It's AMD's own fault for not working well with SK Hynix. Leading to Nvidia to get the gold at both Samsung and SK Hynix.

 

[french toast]

Thanks, if true explains a lot, does mean navi could well have considerably better hbm2 compared to what appears to be out of spec? HBM2 on vega.

 

[CatMerc]

Industry sources...


HBM2 is at version 3 currently, "Perfected Yields", version 4 allows for half-channel => 512-bit, and 2.4 GHz op which is 307.2 gigabyte per sec or 153.6 gigabyte per sec,"Low Cost". 😉

Vega 12 w/ 512-bit HBM2... yum.

Can AMD use four half channel HBM2 stacks to get a maximum bandwidth of 1.2TB/s on existing Vega silicon? Or does the PHY not work in half channel?

 

[Elixer]

- It's AMD's own fault for not working well with SK Hynix. Leading to Nvidia to get the gold at both Samsung and SK Hynix.

Interesting.
I thought AMD & Hynix have a partnership to develop HBM, so, it was said they have/get priority treatment.

I guess that leaves room for Vega RX 64/56 Ti(😉) to get the full 512GB/s bandwidth.

Thanks, if true explains a lot, does mean navi could well have considerably better hbm2 compared to what appears to be out of spec? HBM2 on vega.

Hmm.. actually...

As a bit of background, SK Hynix formally announced their intention to produce GDDR6 in late April, and subsequently added 8 Gb chips rated for 12 and 14 Gbps data rates at 1.35 V to its databook. Meanwhile, 4Hi HBM2 chips with 4 GB capacity have existed in SK Hynix’s product catalog for some time, but recently the company reduced their data rate to 1.6 Gbps.

but...

- AMD is overclocking the 1.6 GHz SKUs of last year. (Just a warning, SK Hynix halted HBM2 production to switch to ver 4.)

so, it seems Hynix has reduced it to 1.6Gbps for ____ reason?

 

[JDG1980]

Industry sources...

Would these be the same industry sources that claimed AMD was still working on Bulldozer-derived architectures for 22FDX?

 

[IntelUser2000]

You don't design for 4 stacks and just change to 2 at the last minute.

It seems most people think HBM/HBM2 is the reason for the state of Vega?

 

[iBoMbY]

- AMD missed the launch for 2 GHz. SK Hynix then prioritized DDR4/LPDDR4/GDDR5/GDDR6 over HBM2 because of low HVM demand. Aka, AMD ruined their own plans by not having a production line already in place. Aka, Fiji Redux like Bristol/Stoney on 28nm w/ only two HBM2 PHYs.
- AMD is overclocking the 1.6 GHz SKUs of last year. (Just a warning, SK Hynix halted HBM2 production to switch to ver 4.)
- It's AMD's own fault for not working well with SK Hynix. Leading to Nvidia to get the gold at both Samsung and SK Hynix.

Only AMD's HBM2 for Vega is 100% from Samsung.

 

[PhonakV30]

But can you really make that claim? Yeah, Navi could be a huge "chip" but will it be competitive at price/perf, perf/watt, etc? Volta will be out by then and I suspect the delta between NVIDIA and AMD will be just as large. AMD simply has a lot of ground to makeup.

My friend look at those Basher's comments.even If Navi becomes better than Pascal in term of Performance/W Or Performance/Price but Worse than Volta , they will call it Fail.PERIOD.
they will tell you :
1) Navi has 20% more transistors but with same Performance
2) Navi Is 20% larger die with 20-100% higher power draw
3) Navi does not provide smooth low latency in DX11 or blah blah blah blah.....
Whatever they say , I'm just saying :Navi won't be competitive

Do you believe XFX GTR RX480 at 1450Mhz does have less power draw than RX 480 reference? i have 4x and I can tell you with my instrument.now How about Vega? We will see.AMD always put too much voltage in other make it works.

 

[ozzy702]

My friend look at those Basher's comments.even If Navi becomes better than Pascal in term of Performance/W Or Performance/Price but Worse than Volta , they will call it Fail.PERIOD.
they will tell you :
1) Navi has 20% more transistors but with same Performance
2) Navi Is 20% larger die with 20-100% higher power draw
3) Navi does not provide smooth low latency in DX11 or blah blah blah blah.....
Whatever they say , I'm just saying :Navi won't be competitive

Do you believe XFX GTR RX480 at 1450Mhz does have less power draw than RX 480 reference? i have 4x and I can tell you with my instrument.

As both a gamer and miner (with several GH/S between ZCASH and ETH) I'm a big advocate for AMD to be competitive. I've owned ~ 50 AMD GPUs over the years and dozens of NVIDIA GPUs. That said, AMD is behind and by a significant margin, especially if these leaks from Vega are true. Is it Global Foundaries? Is it GCN being too dense and compute heavy? I don't know and frankly I don't care, I just want to see competition and a healthy AMD because the healthier AMD is the more competition there is against Intel and NVIDIA and we the consumers benefit greatly.

Navi will be late to the game against Volta and one thing you can count on from NVIDIA is that they will constantly be pushing forward with more streamlined and efficient GPUs throughout their stack. I see the potential in Navi but it's a 1st gen product and AMD's track record since the 200 series doesn't exactly inspire confidence in the roll out being solid.

AMD's only hope for Vega is that it mines exceptionally well both from a pure throughput standpoint and hash/watt. If it doesn't pull some impressive numbers it's a stillborn product that will be sold for little to no profit and it looks terrible for AMD.

 

[tviceman]

My friend look at those Basher's comments.even If Navi becomes better than Pascal in term of Performance/W Or Performance/Price but Worse than Volta , they will call it Fail.PERIOD.
they will tell you :
1) Navi has 20% more transistors but with same Performance
2) Navi Is 20% larger die with 20-100% higher power draw
3) Navi does not provide smooth low latency in DX11 or blah blah blah blah.....
Whatever they say , I'm just saying :Navi won't be competitive

Or you can just look at the facts and call it how it is. Vega is over a year late, bringing similar performance at the same prices already available and doing it with way worse metrics. If Vega had been 6-9 months earlier, if Vega was an improvement in perf/$, or if Vega was able to differentiate itself with better efficiency, etc. then it wouldn't be a fail. None of these issues by themselves makes Vega failure, but when you add it up, Vega is flopping like a fish on it's last gasp before it even exits the pond. Nvidia was roasted by the fire (and for good reason) when Fermi 1.0 came out. But guess what - at least they brought a higher level of performance with it and was only 6 months behind. Vega is 15 months after the GTX 1080 and >=35% slower than the 1080 TI. AMD was always able to at least come within a few percent of Nvidia's best (except with the HD 6970 vs. GTX 580) out of the gate but now they're a solid 35% behind.

Navi doesn't have to win in every metric. But if it loses in most and is 9+ months late, then Navi will suffer the same fate as Vega.

Do you believe XFX GTR RX480 at 1450Mhz does have less power draw than RX 480 reference? i have 4x and I can tell you with my instrument.now How about Vega? We will see.AMD always put too much voltage in other make it works.

All chip makers "put too much voltage" into their chips. It's so they can guarantee a certain percentage of yields at defined clock speeds. I have a i7-7700HQ that I've undervolted by 130mv and dropped temps 10C under firestrike. Anyone can set a power limit with Pascal (or Maxwell) and get a noticeable temp & power usage drop with little or no drop in performance. The practice of undervolting is not limited to AMD hardware. It's nothing new and has only recently become an AMD talking point because AMD has fallen dramatically behind in efficiency.

Let me put it to you this way - if AMD chips were able to be competitive in perf/w with Nvidia, you'd see 5x as many AMD discrete GPU options in notebooks as you do now. But you don't.