AMD Announces Dual-Vega Radeon Pro V340 Card with 32GB of ECC HBM Memory | Hacker News Reader

AMD Announces Dual-Vega Radeon Pro V340 Card with 32GB of ECC HBM Memory

tomshardware.com

127 pointsrerx19 comments

What's more surprising toms hardware actually questions the viability of NVidia 2080... given the recent shocker of 'just buy it'[0]

That being said, I wonder if the card would have solved its memory controller issues, as the desktop Vega draw too much power when it comes to the memory controller. The desktop version pretty much needs water cooling to produce decent results.

[0]https://www.tomshardware.com/news/nvidia-rtx-gpus-worth-the-...

"Just buy it" was an opinion column with a prominent counter opinion column that seems to have been quite effective as revenue clickbait.

I saw this in context, but in the end it doesn't make the piece defensible. The "counter opinion" was that you shouldn't buy (pre-order) the card yet. It's not like it was an NVIDIA hit-piece, just entirely reasonable advice about cards that aren't released yet. "Just pre-order it and don't wait for reviews and benchmarks, by the way here are NVIDIA's claimed performance figures aren't they great" is pretty suspect coming from the editor-in-chief a reviews and benchmarks site.

Edit: also just some really insane lines in there: "When you die and your whole life flashes before your eyes, how much of it do you want to not have ray tracing?" ???

Dear editor in chief for $MAGAZINE would you write an encouraging opinion piece for $BAG_OF_CASH?

I'm not saying this is the case, but it really felt that way seeing the article.

I wish at least TensorFlow was finally running on Vega... Otherwise the card pretty much misses the target when compared to P40, as that one is built specifically for Deep Learning with inferencing. DaaS doesn't seem a large market.

It does, does it not?

https://gpuopen.com/rocm-tensorflow-1-8-release/

I was looking at OpenCL support first:

https://github.com/tensorflow/tensorflow/issues/22

Anyone using ROCm? Is is stable enough/does it support everything in TF 1.8? How is the performance comparing to 1080Ti? Their port of TF doesn't seem to have many users, only 6 issues in TF upstream:

https://github.com/ROCmSoftwarePlatform/tensorflow-upstream

If it were comparable with 1080/Ti and working flawlessly, 32GB would be very enticing!

Anyone know what realistic pricing would be for this? It seems like it could easily be $3500+ (2x Radeon Pro + premium for more VDI seats per server). There's a link about the Tesla P40, but it's not clear to me how comparable that is. I'm also not sure that it's even sold as a separate product as opposed to only being available via system builders.

Will they also make a consumer version? Like the 295x2?

Its TDP is 300W. It has a passive heatsink and relies on the server chassis having big loud fans pushing air from an air-conditioned cold aisle. Dual-slot PC cards seem to top out around 250W, with nvidia's 2080 Ti Founders Edition at 260W with vapor chamber cooling. AMD could go with a fancier cooling system, like the water cooler they put on the R9 Fury X, or they could slow it down to fit in a lower TDP.

Current desktop Vega 64 can go almost 300W in LC versions. It can be pushed more with custom loops, easily 400W.

Source: actually owning Vega FE and tested overclocking using a custom loop. It can easily chew extra wattage when overvolted and overclocking while still staying stable. From performance per watt view, it is ridiculous though. Getting a 10% performance boost by doubling wattage over undervolted stock.

How about just have Vega Cards available at MSRP first?

Yeah, seriously. This coin mining phase has turned finding a decent GPU into a real pain.

Things have gotten much better especially the upper 1080ti and the Titans but the Vega???

  HBM Memory

High Bandwidth Memory Memory? Isn't that like saying ATM Machine? Or DNS System?

But is not perfectly reasonable use of language, if you consider that the language as written there is for humans to read and understand, not for machines to expand acronyms and be confused about semantical errors?

But sure, what you describe has a name, RAS syndrome.

https://en.wikipedia.org/wiki/RAS_syndrome

So just spell out the whole thing. Just say High Bandwidth Memory. Leave out the inscrutable TLA.

Eth mh/s?

TL;DR AMD put two Radeon Pro WX 9100 chipsets on one card and added some form of technology somewhere in the stack to help hypervisors share the resources of these among VMs. The memory figure is the sum of the two 16GB ECC HBM stacks.

Reply on news.ycombinator.com