AMD Vega - Technical discussion

[disi]

The first few posts in this thread are interesting:
https://forums.overclockers.co.uk/threads/the-amd-vega-thread.18781388/

What we know so far:
There will be Vega RX 56 and Vega RX 64
Here is more details and next to specs from 1080, 1080 ti and Fury X:
https://www.pcper.com/reviews/Graph...ga-64-and-Vega-56-Specs-Prices-Power-Detailed
Slower clock speed, but way more cores than 1080 ti.
Peak compute about the same as 1080 ti.
The memory works complete different with HBM2 instead of GDDR5X.

It is not a gaming specific chip, but RX is supposed to be optimized for gaming.
The performance is roughly between GTX 1080 and GTX 1080 ti, closer to GTX 1080.
The price is announced as $599 for the highest model with water cooling (same as 1080 ti).
The normal air cooled version is announced as $499 ($100 less than 1080)

Hopefully some benchmarks soonish

^^

The Vega 64 liquid cooled will hit 345 watts, the air-cooled version hits 295 watts

AMD has told us that RX Vega will be on sale and reviews will go live on August 14th!!

 

[RavenXXX2]

Basically designed for the pro market and mining, gaming took a back seat. At least Nvidia got its priorities in the correct order.

 

[disi]

I like this article which talks about HBM vs. GDDR5:
http://blog.logicalincrements.com/2017/02/types-vram-explained-hbm-vs-gddr5-vs-gddr5x/
GDDR5 cannot be stacked (one layer around the GPU), has a small BUS (smaller connection to the GPU), uses more power because the memory clock speed is much faster.
HBM can be stacked (two or more layers around the GPU), has a much wider BUS, uses less power with a much slower memory clock speed.

With its more recent designs GDDR5X in Nvidia and HBM2 in AMD now.
GDDR5 is more used and way longer in development, HBM is a good idea but not fully optimized yet.

 

[LeMson]

I'm dissapointed vega doesn't have 16GB hbm2 :'(

 

[mmj_uk]

Looking at those tables it's basically just an overclocked Fury X? no increase in ROP's or anything just clockspeed and a decent amount of memory this time.

It looks like a rebadge on a new process to me.

 

[Tute]

The thing that bothers me the most about Vega is another two week wait for benchmarks. People have been patient and now the mickey is being taken.

Apart from that I don’t see the issue with projected performance and price as long as it can dissipate the heat it generates.

Seems to me there are a lot of nVidia shills at work this morning.

 

[chroniclard]

So, are the two Vega 64's identical apart from air vs water cooling?

Theoretically the cheapest air cooled 64 + custom water loop would be similar(if not better) than the liquid cooled one?

 

[Tute]

So, are the two Vega 64's identical apart from air vs water cooling?

Theoretically the cheapest air cooled 64 + custom water loop would be similar(if not better) than the liquid cooled one?

You’d have to be off your rocker to buy the water cooled version. Vega does seem competitive but only at the lower price point.

If you want an AIO buy the cheapest and put your own on.

 

[disi]

So, are the two Vega 64's identical apart from air vs water cooling?

Theoretically the cheapest air cooled 64 + custom water loop would be similar(if not better) than the liquid cooled one?

He says the limited or liquid cooled one also has a higher core clock at: 1677MHz
The boost clock speeds are also the ones expected to run while playing games for example. They are a higher clocked air cooled version with liquid cooling.

I wonder if it is some standard water block you can easily fit into your existing loop...

 

[erazor51]

Tom's hardware:

As an interesting side-note, it sounds like the pixel engine’s Draw Stream Binning Rasterizer, which we introduced back in January, is currently disabled on Radeon Vega Frontier Edition cards. However, AMD says it’ll be turned on for Radeon Vega RX’s impending launch. Don’t expect any miracles from the feature’s activation. After all, AMD is assuredly projecting performance with DSBR enabled. But a slide of presumably best-case scenarios shows bandwidth savings as high as 30%.

http://www.tomshardware.com/news/amd-radeon-rx-vega-64-specs-availability,35112.html

http://www.tomshardware.com/news/amd-radeon-rx-vega-64-specs-availability,35112.html

 

[muon]

GCN

/thread

 

[Rroff]

Looking at those tables it's basically just an overclocked Fury X? no increase in ROP's or anything just clockspeed and a decent amount of memory this time.

It looks like a rebadge on a new process to me.

Even a shrunk Fiji "should" out perform Vega on paper.

snip

That bandwidth saving (and other related savings which reduce load) will most likely equate to 10-15% maybe 20% in some cases power reduction - people hoping to see miracle performance gains from it are likely to be sorely disappointed. It might at least make Vega little closer to the competition in actual game use for power consumption though.

 

[VincentHanna]

Well the WatercooledVega RX 64 TDU is 345 watts

A GTX 1080Ti is 250watts...

 

[AllBodies]

They've basically shrunk the FuryX to 14nm, which should be around 290mm2, and overclocked it thanks to dennard scaling. On the gaming side.

Then added lots of compute-focused logic on top of that, and things like the HBCC, bringing it up to 484mm2.

Result: performs like a hugely overclocked FuryX (so around 1080 performance), but is genuinely very very strong for compute tasks. Especially for the die size.

So bad news for us, good news for compute.

/Thread

 

[CAT-THE-FIFTH]

I saw that on TH.

Also from the AT article on the launch:
http://www.anandtech.com/show/11680...-vega-64-399-rx-vega-56-launching-in-august/3

Talking to AMD’s engineers, what especially surprised me is where the bulk of those transistors went; the single largest consumer of the additional 3.9B transistors was spent on designing the chip to clock much higher than Fiji. Vega 10 can reach 1.7GHz, whereas Fiji couldn’t do much more than 1.05GHz. Additional transistors are needed to add pipeline stages at various points or build in latency hiding mechanisms, as electrons can only move so far on a single clock cycle; this is something we’ve seen in NVIDIA’s Pascal, not to mention countless CPU designs. Still, what it means is that those 3.9B transistors are serving a very important performance purpose: allowing AMD to clock the card high enough to see significant performance gains over Fiji.

Speaking of Fiji, there’s been some question over whether the already shipping Vega FE cards had AMD’s Draw Steam Binning Rasterizer enabled, which is one of the Vega architecture’s new features. The short answer is that no, the DSBR is not enabled in Vega FE’s current drivers. Whereas we have been told to expect it with the RX Vega launch. AMD is being careful not to make too many promises here – the performance and power impact of the DSBR vary wildly with the software used – but it means that the RX Vega will have a bit more going on than the Vega FE at launch.

So nearly 4 billion transistors were used to make the design clock better.

 

[mahius]

The price is announced as $599 for the highest model with water cooling (same as 1080 ti).
The normal air cooled version is announced as $499 ($100 less than 1080)

I did not know that the 1080 and 1080ti cost exactly the same XD

 

[disi]

I did not know that the 1080 and 1080ti cost exactly the same XD

They don't. Maybe my wording is a bit weird

The recommended vendor price for the GTX 1080 ti is $699, same as the liquid cooled overclocked Vega 64. For the GTX 1080 it is $599 and the air cooled AMD Vega 64 is at $499.
I just saw they dropped the price for the GTX 1080 this year down to $499.

They are the same price now, but the 1080 is a year older. In reality all the 1080 are still way beyond £500.

 

[Panos]

I saw that on TH.

Also from the AT article on the launch:
http://www.anandtech.com/show/11680...-vega-64-399-rx-vega-56-launching-in-august/3

So nearly 4 billion transistors were used to make the design clock better.

Anandtech missed some small prints pulling out of the thin air ideas. And if an AMD rep/engineer told them that, then I lost faith to all future AMD GPUs and my advice is anyone who wants buy a 1080Ti now is cheap before prices go up.

FuryX can clock to 1200 core with the AMD UEFI bios, NOT with the stock Bios (with the AIO as is no changes). At 1200 is limited not because of the chip transistors, but because of voltage and power supplied to the GPU. And because the card is limited, if someone wants to run 600mhz ram speed (stock 500) has to drop the core to 1190. Which is a good trade off allowing more performance to come out in total.
If we modify the board to allow bigger power draw, there are people who have achieved 1400Mhz+ core.

So the extra transistors count and extra speed, is a silly argument. At least showing ignorance of what FuryX is capable of to whom ever at anandtech wrote the article or who ever their source was

The whole point of the Vega current benchmark numbers, the thing (GPU) looks stupid. Yes I of the people who supported the idea up to yesterday, being classed as AMD fanboy on the others, writing that.

FuryX stock speed is 1050. By applying 14% OC on core and 25% Ram speed (1190/600 some can do easily 1195/625 if the AMD UEFI Bios is used on custom loop keeping temps bellow 40C) straight away the card is within single digits away (3-4fps minimums on their chart) from the Vega performance at 1600 core. So the Vega IPC is weaker to say the least, than the FuryX. A FuryX at 1600 could be beating comfortably the overclocked GTX1080, the Vega also and inline to start trading blows with the Titan X (Pascal) or the 1080Ti FE. If 4Gb VRAM isn't limiting factor.

So here goes the argument of the "Vega is a shrank Fiji". Is not. Is far worse GPU than Fiji ever was. Imho AMD should have pulled a "Nvidia Pascal". Shrink the Fiji to 14nm clock it higher and sell it with 8GB HBM1 ram.
The chip was more than capable to compete and be a worthy upgrade. Vega is not a worthy upgrade to an overclocked FuryX.

And to put numbers. My (sold) 1080 (with the OLD ram modules not the new one with the faster ones) @2190 was 40% faster than my (sold) Nano @ 1100/550 at 2560x1140 at DX11. At DX12 it tanked (Total War Warhammer for example) compared to the Nano. (Nano was losing 1 fps, the 1080 a whooping 13). Hence sold the 1080, because I was losing Freesync and annoyed me with it's DX12 support (up to November 2016). Since then bought a FuryX and loved it.

And no I am not buying a Gtx1080ti, because Nvidia drivers have issues with the Zen CPUs and doesn't support GPU pass through on Windows VM (over Linux). Something AMD does on both Zen & GPUs.

 

[disi]

Has anyone experience with those AIO liquid cooled cards and what type of connectors they use?
It would be a shame to have to wait for a water block that may or may not fit if they provide with a card that has it already attached and all you need to do it hook it up into your loop?

 

[ubersonic]

Basically designed for the pro market and mining, gaming took a back seat.

Yeah, similar performance to an RX580/1080ti at double the power consumption is a real winner, surprised miners are still buying anything else and not just saving to preorder XD