AMD Vega Hardware Reviews

"Just a "vcore floor" voltage that at this point only exist IMO to make the gpu enforce a higher VDDCI voltage"


Thanks for the link, but apart from answering my mem voltage question, I find his video lacking and not that good TBH. I found out more about my card in few hours I played with it and if someone paid me to do it, you would get better video than that. Shame I haven't got time nowadays ...
 
That's quite impressive for the $100 cheaper 900Mhz slower 1600. If Battlefield 1 was actually fully tested in MP, the 1600 would come out on top overall. GTAV is missing though which would skew it in favor of 7700k and call it a wash.

In fact, theres no reason to dont increase the clock of the 1600 to 3.8-4ghz and it will be overall faster by a good margin. ( yes i know, you can too OC the 7700K, this is not really the point...it just for saying that for 100$ cheaper and 2cores more, the cpu is completely a valid option for gaming if a 7700K stock is a valid option. )
 
Could this be related/impacted by the disabled DBSR? Honest question, I have no idea, but DBSR is related to rendering so...



Did you notice that the information about IPC came immediately coupled with FP16, RPM and INT8? I actually did and Anandtech as well:

"That said, I do think it’s important not to read too much into this on the last point, especially as AMD has drawn this slide. It’s fairly muddled whether “higher IPC” means a general increase in IPC, or if AMD is counting their packed math formats as the aforementioned IPC gain."

I was not expecting higher IPC in all situations. Why? Because if you look at the slide the difference is having two ops instead of one, which for me was a blatant clue that that was referring to RPM. The message was poorly conducted but the image was very revealing.



Given that Vega is still pretty much GCN, plays by its rules and limitations (e.g. maximum of 4 ACEs) and still has 4096 "cores , yes your expectations were too high. Especially when the same number of cores is now responsible for more geometry stuff than before. Didn't you think that that would have a drawback? The same number of units of Fiji but now spread out to more workloads would magically find a way to do more with same, even if running at higher speeds? If you didn't think a drawback would exist, yes that's silo vision because you were thinking about the changing in geometry without thinking about the impacts elsewhere.

I'm starting to think that's the reason for our disagreement. I had my expectations way more in check. It's performance is pretty much where I expected it to be.



That's the job of marketeers. I learned a long while ago the art of smelling bullshit and critical thinking about what they present, not taking anything they say at face value. As you can see from what I wrote above my expectations were balanced by this attitude.



Honestly, the White Paper is not that different from the slides, with the exception of not mentioning IPC anywhere. I guess they realised the stupidity of calling RPM an increase in IPC, which is really what they meant by it. Regarding reasonability of thinking, refer to my answer above about expectations.



Like I said above, it's still GCN. You don't just change how an architecture performs in every single situation overnight just with small tweaks and touches here and there. That's what Vega is, there was not a big overhaul, with the exception of RPM and INT8 which are more due to AI than gaming. Especially when the chip still has the same number of units. With the exception of Pascal consumer, NVIDIA always changed the SM layouts and unit ratios affecting load balancing from Fermi to Kepler and Kepler to Maxwell, looking for optimal performance. I had thought that by now, with 3 iterations of GCN, people would have realised that GCN does not offer the same kind of flexibility NVIDIA architectures have.



Refer to my point about the changes in geometry using more compute resources having a knock out effect on other functions that also use them.
On a brief aside, the number of compute engines has been an unexpectedly interesting point of discussion over the years. Back in 2013 we learned that the then-current iteration of GCN had a maximum compute engine count of 4, which AMD has stuck to ever since, including the new Vega 10. Which in turn has fostered discussions about scalability in AMD’s designs, and compute/texture-to-ROP ratios.

Talking to AMD’s engineers about the matter, they haven’t taken any steps with Vega to change this. They have made it clear that 4 compute engines is not a fundamental limitation – they know how to build a design with more engines – however to do so would require additional work. In other words, the usual engineering trade-offs apply, with AMD’s engineers focusing on addressing things like HBCC and rasterization as opposed to doing the replumbing necessary for additional compute engines in Vega 10.

From Anandtech's review.
AMD spent the resources that could be put into reconfiguration into other things. It's as simple as that.
 
The way I see Vega's shader power, only about 60% to 70% of it at best is being utilized in games, with a few rare exceptions. The massive shader array it has is not there for gaming, but rather for all the other markets it's looking to serve that NVIDIA has specialized GPU's taking care of.

I have no doubts that a gaming part with 3072 shaders but with the same front end and back end as Vega 10 would have performed extremely close to what we have now, while being smaller and taking less power. A part of me wishes that that's exactly what Vega 11 is.

For a GP102 competitor in gaming, two extra shader engines would need to be implemented and have those 4096 shaders spread amongst them.
 
In fact, theres no reason to dont increase the clock of the 1600 to 3.8-4ghz and it will be overall faster by a good margin. ( yes i know, you can too OC the 7700K, this is not really the point...it just for saying that for 100$ cheaper and 2cores more, the cpu is completely a valid option for gaming if a 7700K stock is a valid option. )
The benchmark was the 1600 @4Ghz and the 7700k @4.9Ghz
 
And DDR4-3200. So almost best-case compare for Ryzen.
Best-case scenario for Ryzen would've been a similarly priced 1700X with SMT disabled and overclock on top.

Edited for pedantics.
 
Last edited by a moderator:
Best-case scenario for Ryzen would've been a similarly priced 1800X with SMT disabled and overclock on top.
huh? 1800X is over $100 more. In most of those games, the extra cores wouldn't make much difference and the core clock would be identical.
 
They are not even pushing the CPUs there, when testing CPUs of this caliber a 1080Ti is a must, or maybe some Crossfired Vegas. Single GTX 1080/Vega 64 are not enough to push these CPUs in half of these titles.
 
Best-case scenario for Ryzen would've been a similarly priced 1800X with SMT disabled and overclock on top.
Oh, suddenly 100 $ is similarly priced. Need to remember that for future reference. Good thing, R5-1600 and i7-7700K are similarly priced too.

--
edited later that day:
Best-case scenario for Ryzen would've been a similarly priced 1700X with SMT disabled and overclock on top.

Edited for pedantics.
Now pointing out blatantly biased comparisons are pedantics. It gets better.
 
Last edited:
LMAO! Newegg will now sell a Vega 56 for only $420 BUT you have to buy it as part of a combo that includes a $129 "AMD Gift" of Prey/Wolfenstein.
Newegg need to look up "gift"
Actually people just need to realize that they can't sell your "Radeon Pack"-card as just the cards MSRP, but that particular "Radeon Pack" price. Of course they should be sold as "Radeon Packs" and not "Radeon + gift that costs you", but still people need to stop saying "price is this much higher than it should" when it isn't.
 
Actually people just need to realize that they can't sell your "Radeon Pack"-card as just the cards MSRP, but that particular "Radeon Pack" price. Of course they should be sold as "Radeon Packs" and not "Radeon + gift that costs you", but still people need to stop saying "price is this much higher than it should" when it isn't.

Well, I think they should drop the term "gift" completely. A gift has no cost to the recipient. In this case a more accurate item description is "RX Vega 56 at only $21 over MSRP provided you buy two games at about $10 over new release price."
 
Even though it's not exactly a review, I can say that my retail V56 ran for the past day non-stop under load with 1100 MHz mem clock after flashing the V64 VBIOS without error. Seems like either that V56's memory has some reserves or it is indeed the same memory that's used on V64 as well.
 
Even though it's not exactly a review, I can say that my retail V56 ran for the past day non-stop under load with 1100 MHz mem clock after flashing the V64 VBIOS without error. Seems like either that V56's memory has some reserves or it is indeed the same memory that's used on V64 as well.
What was your HBM clock before flashing Vega 64 BIOS? I might flash my card tonight!

EDIT:

After flashing my Vega 56 I can run 1100 HBM2 clocks!
Quick observations:
- card vGPU floor increased from 0.925V to 1.025V resulting in ETH power usage jumping by around 10W to 20W
- mining efficiency ETH/HBM clock is lower with new BIOS, I need 970MHz HBM to match 930MHz 35MH ETH mining on VEGA 56 BIOS
- GPU-Z now detects Samsung HBM memory instead of Hynix (obviously reading from BIOS)

At 1100HBM and 1007MHz GPU mining ETH using Claymore 10 has reached 38.5MH/s (Cat 17.8.2, Windows 10 64bit).
 
Last edited:
What was your HBM clock before flashing Vega 64 BIOS? I might flash my card tonight!

EDIT:

After flashing my Vega 56 I can run 1100 HBM2 clocks!
Quick observations:
- card vGPU floor increased from 0.925V to 1.025V resulting in ETH power usage jumping by around 10W to 20W
- mining efficiency ETH/HBM clock is lower with new BIOS, I need 970MHz HBM to match 930MHz 35MH ETH mining on VEGA 56 BIOS
- GPU-Z now detects Samsung HBM memory instead of Hynix (obviously reading from BIOS)

At 1100HBM and 1007MHz GPU mining ETH using Claymore 10 has reached 38.5MH/s (Cat 17.8.2, Windows 10 64bit).

Stable max before was 960 (might be a bit more, but 980 definitely was not). GPU-z has been update lately to (more?) correctly detect HBM manufacturer. I have not tests mining efficiency at same clock speeds, but I am getting 38.5 MH/s as well at roughly the clocks you stated (995/1100 MHz, 71/77/81 °C for the three temps with HBM obviously being the hottest, and at 2,000-2,100 rpm fan speed in an rather small desktop case, GPU-z tells me 0,875v, though it might be higher, did not measure directly at card here at home). System pulls 205W at the wall, GPUz says GPU-only power draw is 116W, normal idle for this system is around 35W.
 
Back
Top