Senior AMD exec teases Vega GPU launch venue

[HEXUS]

Building anticipation. October isn't that far away.

Read more.

[lumireleon]

so vega 490 will essentially beat the P100, titan x and quadro p6000 in compute performance just like the old days of 6990 on how it shuttered 590? AMD fan boy alert but who cares!

[sykobee]

Well on the one hand, AMD started talking about Polaris forever before we could actually get our hands on it.

OTOH it really doesn't make sense to book a venue until you know you can hit the date accurately.

Basically, Vega is now in production ramp for October (maybe November) launch. Sweet times. GFX 9 (likely GCN5) too, let's see what this brings (hopefully Perf/W on par with Pascal, sheesh).

[Boris]

I'm not sure that's the type of venue you book. You just push your way past the debris and get your tetanus shot.

This venue strikes me as a Battlefield setting; so, I wonder if Vega will be launched in tandem with BF 1.

I'm pleased that the HBM 2 cards are approaching so soon after the current gen.

[anselhelm]

I want to believe, but Polaris was (is) fairly underwhelming. Don't get me wrong: Polaris is a huge step forward for AMD and it's brought better power efficiency, but I guess I just wanted a triple-A performance from AMD rather B+.

That said, I still think the 4GiB RX 480 provides good value for money overall, but I'm less convinced about the RX 470 and RX 460, which need to be cheaper for what they provide. The issue is perhaps that it was so over-hyped that there was no way it could live up to it.

Nvidia are such prats that I really do want AMD to win for once, even if just to force Nvidia to reduce its price gouging, but in reality I'm sadly not expecting a revolution.

[LeBron85]

I want to believe, but Polaris was (is) fairly underwhelming. Don't get me wrong: Polaris is a huge step forward for AMD and it's brought better power efficiency, but I guess I just wanted a triple-A performance from AMD rather B+.

That said, I still think the 4GiB RX 480 provides good value for money overall, but I'm less convinced about the RX 470 and RX 460, which need to be cheaper for what they provide. The issue is perhaps that it was so over-hyped that there was no way it could live up to it.

Nvidia are such prats that I really do want AMD to win for once, even if just to force Nvidia to reduce its price gouging, but in reality I'm sadly not expecting a revolution.

Completely agree.

[chj]

I want to believe, but Polaris was (is) fairly underwhelming. Don't get me wrong: Polaris is a huge step forward for AMD and it's brought better power efficiency, but I guess I just wanted a triple-A performance from AMD rather B+.

That said, I still think the 4GiB RX 480 provides good value for money overall, but I'm less convinced about the RX 470 and RX 460, which need to be cheaper for what they provide. The issue is perhaps that it was so over-hyped that there was no way it could live up to it.

Nvidia are such prats that I really do want AMD to win for once, even if just to force Nvidia to reduce its price gouging, but in reality I'm sadly not expecting a revolution.

Too true. AMD are good at drumming up hype but that just makes the releases underwheling. Need someone to keep Nvidia in check though with their pricing.

[jimbojnr81]

Part of me thinks that image looks fake, maybe showing off the processing power of Polaris.

[jimbojnr81]

Part of me thinks that image looks fake, maybe showing off the processing power of Polaris.

Make that Vega not polaris

[badass]

Polaris = 2.5x perf/Watt.

Except it isn't. RX480=150 Watt (if you're charitable) Performance is slower on average than both the R9-390 and R9-390x - Toms hardware review shows the RX-480 as clearly slower in 4 games, between/similar in 3 games and faster in only one. R9-390/R9-390x = 300 Watt.

More evidence that they missed performance targets and wound up the clockspeed to compensate, hammering power consumption.

AMD claim 110W power for the actual GPU, leaving 40W for the rest of the board. Looking at CAT's graph http://forums.hexus.net/pc-hardware-...ml#post3693307 I did a quick calculation. Remember power consumption increase linearly with clock speed and with the square of the voltage.
At 1 GHz I get a relative power consumption of 720
At 1.25 GHz I get a relative power consumption of 1510
That's more than double for an extra 25%

Of course that graph represents average voltages required so YMMV, however that means that at 1GHz Polaris could have been a 95W card


I'm not holding my breath for Vega. Unless the hype stops now. AMD have a history of hyping underperforming products and being quiet about their good products..

Just look at Barcelona - Massive Hype
Phenom II - Low key, very good
Bulldozer - utter rubbish, massive hype - including claims of higher IPC than the Phenom II!
Piledriver/richland - what bulldozer should have been. faster and more efficient. No hype
Kaveri/Godaveri - big hype, good IPC increase. to make up for that, they lowered clock speeds
Carizzo - low key - massive improvement in efficiency.

Look at the GFX cards. I remember how hyped R600 was. Then there was even more hype when we found out how cheap it was going to be. Then they delivered a card that was priced correctly for its performance and was a noisy power hog. A bit like the RX-480

So what's on the Horizon?
Vega - so far, not enough time for the Hype to start. Lets hope that it stops here
Zen - massive, huge hype. I really, really, really hope it's the exception to the rule but after thinking about it, I'm not holding my breath.

[scaryjim]

Polaris = 2.5x perf/Watt. ...

That's bound to be at a particular performance or TDP level though. I suspect it's running a fully-enabled Polaris 11 at around 950MHz, perhaps with slightly slower memory too, while getting the performance of a stock desktop R7 260X. Remember the perf/watt claims for Fiji - it was 2x perf/watt for Nano, but only 1.5x perf/watt for Fury X - the higher clocks eating a big chunk of the efficiency. Same's bound to be true for Polaris...

[badass]

One more thing - The stream processors in the RX480 in games are no faster per clock than the stream processors in the R9-390. The performance is matched simply by increasing the clockspeed and the lack of memory bandwidth is mitigated by improved colour compression.

Since every leak indicates the first Vega 10 has 4096 stream processors, Vega 10 will (if the 4096SP leaks are correct) be slower than the GTX1080 apart from a handful of titles such as Doom with Vulcan enabled, Ashes of the Singularity etc. Vega 11 is the one to look for in terms of dethroning Nvidia. With 50% more SP's it might be faster than the Pascal Titan.

I did chuckle at the accuracy of rumours though http://hexus.net/tech/news/graphics/...r-says-report/

I'm not having a pop at the reporting of rumours BTW. Merely pointing out the shaky foundations of said rumours, and therefore of the maths behind this post.

[XRoyalT]

Oh im going to regret purchasing two 1080 GTX's if the news of AMD's Vega release being pushed forward to October this year is true!

[watercooled]

That's bound to be at a particular performance or TDP level though. I suspect it's running a fully-enabled Polaris 11 at around 950MHz, perhaps with slightly slower memory too, while getting the performance of a stock desktop R7 260X. Remember the perf/watt claims for Fiji - it was 2x perf/watt for Nano, but only 1.5x perf/watt for Fury X - the higher clocks eating a big chunk of the efficiency. Same's bound to be true for Polaris...

One more thing - The stream processors in the RX480 in games are no faster per clock than the stream processors in the R9-390. The performance is matched simply by increasing the clockspeed and the lack of memory bandwidth is mitigated by improved colour compression.

Since every leak indicates the first Vega 10 has 4096 stream processors, Vega 10 will (if the 4096SP leaks are correct) be slower than the GTX1080 apart from a handful of titles such as Doom with Vulcan enabled, Ashes of the Singularity etc. Vega 11 is the one to look for in terms of dethroning Nvidia. With 50% more SP's it might be faster than the Pascal Titan.

I could be completely wrong and I apologise for repeating myself, it's just a theory of mine, but from the details we have access to it really doesn't seem like 14nm is well-suited to desktop-clocked GPUs. It's fairly safe to assume AMD made at least some improvements to efficiency at an architectural level over the previous generation, so it doesn't seem like they gained very much at all from the move to 14nm. OK so Fiji isn't necessarily a completely fair comparison as it's likely using a fair bit less power for HBM vs GDDR5, but it's also AMD's newest revision of GCN before Polaris.

Nvidia on the other hand gained a lot from the move to 16nm despite having massively increased clocks, so I really do wonder about AMD on 16nm, and they surely must have considered it and probably get quite a bit of data from the Xbox One S SoC.

Given the rumours than Nvidia are using Samsung 14nm for their smaller GPUs, perhaps 14nm is either significantly cheaper, or just better suited to low-power applications. Or maybe volume is limited at TSMC so they're using other fabs for the higher-volume stuff, but then why would AMD/MS be using them for the Xbox One S chip? That chip being made at TSMC on 16nm is really interesting IMO.

As for the performance of Vega 10 - if it's architecturally similar to Polaris and clocks about the same then I agree it's looking more like it will vary between 1070 and 1080 performance depending on the game, perhaps more if Pascal is memory-bound anywhere. However if they get the clocks significantly higher one way or another then it's anyone's guess how it will perform!

I did chuckle at the accuracy of rumours though http://hexus.net/tech/news/graphics/...r-says-report/

I'm not having a pop at the reporting of rumours BTW. Merely pointing out the shaky foundations of said rumours, and therefore of the maths behind this post.

It looks like those 'leaks' were just guesswork, but to be honest I was quite surprised with how quickly they launched the GP102, especially considering it's Nvidia who have historically had catastrophic yields every time they adopt a new node and blame it entirely on TSMC, despite AMD having no complaints and launching earlier. It seems like 16nm worked out really well for them however.

[scaryjim]

I could be completely wrong and I apologise for repeating myself, it's just a theory of mine, but from the details we have access to it really doesn't seem like 14nm is well-suited to desktop-clocked GPUs. It's fairly safe to assume AMD made at least some improvements to efficiency at an architectural level over the previous generation, so it doesn't seem like they gained very much at all from the move to 14nm. OK so Fiji isn't necessarily a completely fair comparison as it's likely using a fair bit less power for HBM vs GDDR5, but it's also AMD's newest revision of GCN before Polaris. ....

2 things:

1) AMD's perf/watt comparisons were all against pre-Fiji GPUs. I believe Fiji had some extra efficiency sauce baked into shaders (which presumably has carried into Polaris), and definitely gained a fair bit from the move to HBM. GDDR5 is hugely inefficient in comparison.

2) I'm not going to repeat my full speculation (from my post in http://forums.hexus.net/pc-hardware-...ml#post3694288) but essentially I suspect there are different versions of GF 14nm optimised for different markets (this was certainly true of 28nm) and Polaris is fabbed on a very mobile-oriented, low-power variant.

[watercooled]

1) Yeah that's exactly what I mean - it's a fairly safe bet that any architectural improvements from Fiji will have carried across to Polaris so at a minimum we'd expect to see static perf/watt even on the same node, allowing for different memory power consumption of course. I wonder just how much Polaris loses out from using GDDR5? It's a tricky one until we have some more HBM cards to use as reference points.

2) I'll leave that on the other thread rather than having the same conversation twice.