GeForce RTX 3080 vs. Radeon RX 6800 XT: 30 Game Benchmark

neeyik

Posts: 1,877   +2,191
Staff member
I have two Sapphire R9 Furies with 4GB of HBM1 but I'd much rather they had 6 or 8GB of GDDR5. Let's remember that HBM1 had an astonishing 4096-bit bus. That gave it 16x the bandwidth of today's RX 6800 XT and 8x the bandwidth of the RTX 3080. That didn't stop the GTX 980 from out-performing it with regular GDDR5.
The HBM modules on the R9 Fury were clocked to 1 Gbps, whereas the GTX 980's GDDR5 were at 7 Gbps - so although former's combined bus width was indeed an enormous 4096-bits, giving a peak bandwidth of 512 GB/s, the comparatively slower access speed didn't help matters on a per-ROP/MC basis (and in the case of Fiji, it was two controllers per HBM stack). So if data in the same page is being requested, HBM was slower than GDDR5, although such situations are far less frequent than its normal usage.

Of course, all the peak DRAM bandwidth in the world isn't going to matter very much, if the chip isn't requesting the data very often, and this was GCN's Achilles' heel. This is why the likes of the RX 590 fairs pretty well against the R9 Fury, despite having half of everything (except clock speed).
 

Avro Arrow

Posts: 1,256   +1,386
TechSpot Elite
The HBM modules on the R9 Fury were clocked to 1 Gbps, whereas the GTX 980's GDDR5 were at 7 Gbps - so although former's combined bus width was indeed an enormous 4096-bits, giving a peak bandwidth of 512 GB/s, the comparatively slower access speed didn't help matters on a per-ROP/MC basis (and in the case of Fiji, it was two controllers per HBM stack). So if data in the same page is being requested, HBM was slower than GDDR5, although such situations are far less frequent than its normal usage.

Of course, all the peak DRAM bandwidth in the world isn't going to matter very much, if the chip isn't requesting the data very often, and this was GCN's Achilles' heel. This is why the likes of the RX 590 fairs pretty well against the R9 Fury, despite having half of everything (except clock speed).
I agree with this. The only reason that I bought my R9 Furies was because it was in the middle of the 2017 mining craze and newegg had them for $235USD. It was less than half of the price of even the RX 580 at the time. I was actually up for hours trying to see if there was something wrong with them because to this day, I still don't understand how newegg sold them for so cheap at the time. My guess is that it was because the R9 Fury was a power hog and that's not conducive to good mining profits. The price did go up to $260USD when Greg Salazar did his video on them:
Otherwise, I would have chose the RX 590. There just weren't any that didn't cost less than $500 at the time. I was becoming a bit antsy because with games not supporting multi-gpu setups anymore, I had to upgrade. My twin Gigabyte HD 7970s (aka R9 280X) were getting long in the tooth performance-wise and even worse, only had 3GB of VRAM.
 

Gerald L

Posts: 10   +2
That very much used to be the case, due to increased electrical loading, and while it still can be, a modern CPU handles multiple ranks far better than they ever used to. Only one rank can be read/written to at a time, but with dual or quad ranks, the others can be 'prepped' for use whilst the selected rank is being accessed. Essentially this shaves off a few nanoseconds of latency and for memory intensive situations, even tiny gains can result in notable benefits.
Ahhhh. Thank you; that was very helpful!
 
Of course, just because $999 is the rumored retail price of the RTX 3080 Ti, that doesn't mean that's what most people will be asked to pay for it. Scalpers have been notorious for snatching up 3080 and 3090 cards and then selling them at exorbitant prices. Unless something drastic changes between now and May 26, the RTX 3080 Ti will face a similar fate. Combine that with the leaked hash rate of the 3080 Ti — 118.9Mh/s — and the RTX 3080 Ti will also be a highly sought-after card for crypto miners.