Low demand for AMD Radeon Pro W6800?

I was comparing the prices of recent AMD GPUs with ECC memory, when I noticed a pricing oddity; the Radeon Pro VII is often selling at the same price or slightly greater than the W6800. Of course, all prices are probably still inflated due to shortages, but in such circumstances especially I would expect people to purchase a newer+cheaper card if available.

The only explanations that come to mind are that HBM2 is that much more valuable than GDDR6, or that so many workstation card buyers use PCPartPicker that its absence there is making people settle for an older card at the same price

Recent (everything older is GCN 1 or 2) AMD GPUs with ECC memory (accelerators w/o display output excluded):

Card Launch Architecture Memory Peak Memory
W6800 2021 RDNA2 32 GiB GDDR6 512 GiB/s
Pro VII 2020 Vega20 16 GiB HBM2 1024 GiB/s
WX8200 2018 Vega 8 GiB HBM2 512 GiB/s
WX9100 2017 Vega 16 GiB HBM2 484 GiB/s
Pro SSG 2017 Vega 16 GiB HBM2 484 GiB/s

Data from AMD’s product comparison table for professional graphics.

the radeon pro vii got gobbled up by miners. it is nearly identical to a regular radeon vii and that is still one of the most efficient ethereum cards on the planet.

The bandwidth and throughput on the Pro 7 is insane! A 4096-bit bus?

ETA: Actually that’s what’s wrong with GPU’s lately. Manufacturers dropped the high bit bus’s(?) in favor of more GB memory. That’s why a 980Ti can keep up with a newer(within a couple gens) “high end” card, it’s got a higher bit-but size and faster speeds on the memory.

But it is only twice the peak memory bandwidth; so is it a wider but slower connection than GDDR6? The latency could be substantially better, since the HBM2 dies are physically much closer than the GDDR6 chips, but how else can a wider bus help if the speed is not improved?

@Zedicus what in Ethereum is the VII or its siblings, the Pro VII and MI50, so well suited to?
Is faster processing on non-HBM, half-bandwith, but double capacity not better?

1 Like

Radeon VII/Pro VII has three times the double precision throughput. It’s very good for mining of course with the memory bandwidth on tap, but the real draw is for the users who need the full stack of half, full, and double precision capabilities without much compromise (except perhaps in data size).

1 Like

basically what @TheSkuffedKnerd said.

Ethereum likes a wider road more than a higher speed limit. the GPU itself is often underclocked and undervolted while the RAM on the GPU has the timings cranked down and the speed cranked up. so starting with HBM2 and an uberwide bus makes them amazing for ethereum. even out of the BRAND NEW AMD and Nvidia cards, nothing beats a Radeon vii in efficiency. some are getting close now though.

also the vii and vii pro have some of the… @Fouquin said it.

1 Like

Ah, ~1 vs. ~6.5 for FP64, I did not notice that; that is a substantial difference.


1 Like

This topic was automatically closed 273 days after the last reply. New replies are no longer allowed.