There's no real regression, just too high clocked chips as shown by the powersaver modes which lose just a tad of performance while giving leaps and bounds better perf/watt
AMD Vega Hardware Reviews
- Thread starter ArkeoTP
- Start date
Are they compared to R9 Nano, which also does not sit way beyond the sweet spot - contrary to Fury X?
Yes, at least going by TPU's numbers.
They had Fiji/R9 Nano at high resolutions (1440p, 4k) equaling 980 Ti perf/watt, while Vega 56 in default configuration offers 10-15% better perf/watt than 980 Ti. Vega 64 on the other hand is equaling or losing by couple %'s to 980 Ti at default profile & bios, beats it by some 5-10% on 2nd bios default profile, while with the powersaver profiles it beats 980 Ti perf/watt by 26-31% (default bios) and 29-34% (2nd bios)
edit: comparing to Fury X, only default BIOS + Turbo-profile makes Vega 64 lose to Fury X on perf/watt at TPU benches, all the other profiles and 2nd bios are better.
Jawed
Legend
Doesn't GF "fine" AMD for not buying enough wafers?
Well, cross-comparing is all we have at this point. Maybe comparison to Fury X would be more fair? It's and Nanos perf/watt should have evolved more or less the same?
At launch Nano offered 17-22% better perf/watt than Fury X. The default bios powersaver offers around 40% and 2nd bios powersaver around 43% better perf/watt than Fury X today. Default Vega 54 offers 22-24% better perf/watt than Fury X today.
That's why I asked if I missed some data points wrt this topic. Cross-compairing is all fine, when you acknoledge the necessary grain of salt to ingest along with conclusions from that.
Anarchist4000
Veteran
The slide said it was heavily used by the resolve pass, but the context was FP16 optimizations in general, not that it was only 30% on that shader. Many other areas to use FP16 not covered in that presentation that would be passing data to that shader. Not all of them would have been covered. Even if limited to that one shader it's ~5% overall increase given numbers on later slides with the time measured. FP16 isn't necessarily packed math either but packed registers.
What IS a bit disturbing is when you compare to AMDs previous chip on the same process, Polaris. VEGA 64 has more than twice the ALU power, slightly below twice the bandwidth and twice the power draw. So it would be reasonable to expect twice the performance or a bit better thanks to architectural advances, right? And we're seeing that in a number of synthetic tests, yet we're not anywhere near that for either games or graphics benchmarks on average.
I'd really like to have a better handle on why that is the case. We've seen the numbers, next question is the "why".
Actually, at mere 1080p which favors the Polaris Vega 56 manages to match RX 480 perf/watt, lose to RX 470 and beat 570/580, going higher resolutions Vega 56 first leaves 480 dust and at 4K beasts 470 perf/watt too.
Performance/watt is a dubious metric for a chip, because it changes drastically with power draw and clock speed which has pretty much an O(n^3) relationship in the part of the curve where the VEGA64 and full Polaris products typically reside.
Application performance/ALU FLOP, or per unit of bandwidth is probably more useful when comparing VEGA and Polaris architectures. As I said, my "natural" expectation would be that VEGA would perform better per FLOP or unit bandwidth, given its promoted enhancements to CUs and bandwidth saving techniques, not to mention the touted amount of on-chip storage. That's not what we are observing in application benchmarks though.
D
Deleted member 13524
Guest
You'll have to compare the R9 Nano to the Vega Nano, not the high-end/high-consumption cards based on Vega 10.
Total power consumption doesn't depend only on the graphics chip. If you're using a graphics chip that is using less power then you can also use power converter ICs that use less power, for example.
GF is AMD's former fabrication plants who in the meanwhile joined with some of IBM's fabs. AMD having to use GF is probably part of the deal they made when the split was made, in order to prevent GF from going out of clients.
GF has made all of AMD's CPUs so far, but what I don't get is why they decided to extend that obligation to GPUs when the FinFet generation came up. Up until Polaris, AMD's (and ATi's) GPUs were all made on TSMC or UMC IIRC.
DrYesterday
Newcomer
Ryzen seems to do fine on the GF 14nm process. The 1800X, in a Prime95 torture test, only draws 112W @ 3.8GHz. Is it possible that the GF process is good for CPUs & poor for GPUs? Or did AMD just do a poor job with Vega?
Anyone know if Vega ISA supports any level of PlayReady 3? Perhaps @Ryan Smith ?
The short answer is that Global Foundries was AMD. They spun off their fab in order to not have to carry the full cost of process development themselves, and judged that this part of their business would have better odds at staying close to the cutting edge of process technology as a pure play foundry than as a part of AMD.
Part of the deal when they divested that aspect of their business was that they promised that their business would stay with Global Foundries, to lower the risk for whoever took over the fabs. Mubadala who bought GF, also owns just under 20% of AMD. The ties are still very close. And that's where it stands today. AMD last autumn paid GF a large sum in order to be able to somewhat side step the Wafer Supply Agreement. Read Ryan Smiths write-up of that here.
My bad, I misready you saying 56 for some reason instead of 64. RX Vega 64 however does the same, beats the Polaris 10 in perf/watt, as long as you use the powersaver profile instead of default, which indicates they had to push Vega further past it's optimal range compared to Polaris (which also is past it's optimal range, which would be around 1 GHz IIRC)
I believe Vega is currently held back even more than earlier GCNs at launch simply because of the various changes in the architecture. There are several completely new features, which are said to be enabled but can't be controlled by devs or reviewers - they're just there and used when the driver thinks they should be used. That said, devs apparently can tailor their code to fit those new features better even when they have no control over it, which could bring in the improvements you're looking for.
Performance/watt is IMO one of the best performance metrics, since it enables you to compare how much each chip is doing useful work from end user perspective, per watt. You can then compare those to the theoretical numbers with ease, too.
D
Deleted member 13524
Guest
Ryzen fares fine on 14LPP against Intel's CPUs built on Intel's 14FF process.
No one is using TSMC for x86 CPUs, and it's not like 16FF+ is their flagship process anymore, either.
Fully agree that performance/watt is an important metric! It's just that it moves around so much depending on frequency/voltage.
It's interesting to note that Apple when announcing the iMac Pro claimed 11TF for the Vega64 inside, or 1300-1350MHz. They are not compelled to play the numbers game like the PC market, so even back then I took that as an indication that there was a diminishing returns knee somewhere around there for VEGA.
Similar threads
