RX Vega Architecture Issues

[Patrick_Bateman]

Hello,
As many are aware, the Vega architecture is not as good at gaming as its theoretical numbers would lead you to believe. For instance, RX Vega has over 12 TFLOPS of performance, while the GTX 1080 Ti stock can do around 11 TFLOPS of performance. However, the Vega 64 barely competes with the GTX 1080 which does around 8 TFLOPS.Theoretically, it should have over 50% more performance than the 1080, but it just ties it in gaming. Is this an optimization issue? Or is it a flaw with the architecture itself. The issue that I can see with the architecture is its dependence on memory bandwidth. For instance, I remember (maybe erroneously) Gamers Nexus mentioning that Vega significantly benefited from increased memory throughput, which lead me to believe that the HBM2 was the issue with Vega. The failure of HBM2 to meet its expectations causes the Vega cores to be starved for resources, which is highlighted by a modified RX Vega 56, which only performs ~2% worse compared to RX Vega 64 at the same clock speed, despite having significantly fewer cores. Couldn't AMD have widened the memory bus to 4096 bits like they had with the R9 Fury to make up for the decreased clock capabilities of HBM2? I am not a CPU architect by any stretch, so could someone more knowledgeable explain what would have been possible by AMD to fix Vega and what the trade-offs are?

TLDR: What is wrong with Vega?

 

[k1114]

And the rx 580 is 6 tflops and competes with a 1060 6gb at 4 tflops but you only mention vega. Hbm2 is helping not hurting. As mentioned, it benefits significantly and memory bandwidth helps more at higher res. It doesn't depend on it like you think. If memory bandwidth was an issue, they'd give it more. How else do you think they decide what the bus width should be? https://www.techpowerup.com/reviews/AMD/Radeon_RX_Vega_64/31.html As you can see the res go up, so does vega in relative performance. I don't know where 2% came from for the vega 56 vs vega 64 as that's much too small. The number of cores never scaled well with performance. Just look at past gpus on both sides. But the theoretical numbers will be linear as it's still the same flop per clock so the higher you go the worse it gets. You really can't compare any theoretical numbers so with that; there's nothing wrong with the 400, 500, vega or any series in the flop topic.

Let's get into why flops is more but it's less real world performance. There's a reason why we ignore theoretical numbers. Especially flops which is calculated by 2 flops per clock no matter the architecture. This means it does not take in the architecture or really anything beyond cores and clocks into account. Check for yourself, it's cores * clocks * 2 = flops. That is useless for real world performance guesstimating which, among other aspects, will be affected for architecture, memory bandwitdh, memory amount, software optimization, etc. Flops are for fp calculations, that's what a flop is. For cpus, flops per clock changes with some arch but it's still inaccurate for real world performance. Haswell and skylake are both 32 flops per cycle but we know the ipc increased.

You can tell something by the numbers; that like intel, nvidia has higher ipc than amd. Too bad there's no specific number to calculate ipc but then again, the real world isn't clear cut either. TLDR, you really should just throw the spec sheet out the window. If you want to know how it compares in the real world, go to the real world. Beyond that, you only need to compare price.

 

[Patrick_Bateman]

MERGED QUESTION
Question from Patrick_Bateman : "AMD HBM2 Underperforms"

Why is it that HBM2 underperforms? HBM1 was great compared to GDDR5 when it was released, but HBM2 makes much less dramatic improvements compared to what Hynix and Samsung were touting. Could AMD have widened the memory bus to 4096-bits in Vega to increase the memory bandwidth at the cost of power? The bandwidth is even lower than HBM1 in vega. Gamers Nexus posted an article and a video about this here: https://www.gamersnexus.net/guides/3032-vega-56-cost-of-hbm2-and-necessity-to-use-it