The full amps - RTX A6000 now Available

[williamcll]

With the full 10752 cores, this deluxe workstation card that is surprisingly priced at half of a RTX 8000, which has less than half of its cores compared.

Quote

NVIDIA has discontinued its Quadro and Tesla series by offering a single series of products known as NVIDIA RTX Axx or NVIDIA Axx. It actually matters if the card carries the RTX branding because these cards will replace the Quadro series, while the non-RTX A40 card is basically a Tesla successor that will accompany the already launched GA100 A100 accelerator. The RTX A6000 is the only graphics card based on GA102 GPU to feature all CUDA cores enabled. With 10752 CUDA cores, the GPU has a single-precision compute performance of up to 38.7 TFLOPs, which is 3.1 TLFOPs more than GeForce RTX 3090.

Unlike the gaming card, however, the RTX A6000 features twice the memory capacity of 48 GB. Such modules are currently only available with GDDR6 memory technology, hence the memory bandwidth is actually lower on the former professional model. NVIDIA RTX A6000 features four DisplayPort 1.4 connectors with no HDMI 2.1 output. It’s a blower type design cooler, so it is not a datacenter card like A40, which only features a passively cooled heatsink. Only two cards can be connected through a new low-profile NVLink bridge. The A6000 is a workstation card that also supports NVIDIA vGPU virtualization technologies.

Source: https://videocardz.com/newz/nvidia-announces-rtx-a6000-workstation-graphics-card-with-full-ga102-gpu-is-now-available
Thoughts: Shiny indeed, I wonder what in the production process made Nvidia to decide to lower down the price? I'm also guessing GDDR6X has no EEC support yet, perhaps there will be a refresh later next year?

[DildorTheDecent]

What a good looking card. That nvlink bridge looks great as well.

[AbydosOne]

6 minutes ago, williamcll said:

How does the top card get any airflow?

[igormp]

6 minutes ago, williamcll said:

I'm also guessing GDDR6X has no EEC support yet, perhaps there will be a refresh later next year?

GDDR6X has a kind of error detection builtin. The problem is that 18Gbit modules are not available yet, so a GDDR6X based GPU with 24 modules caps at 24GiB. A new model when such modules become available would be possible, I guess.

[igormp]

1 minute ago, AbydosOne said:

How does the top card get any airflow?

It's a blower card, it gets airflow through the front and expels it at the back. That's why blowers are better than open-air for multi-gpu configs.

[n0stalghia]

This is completly irrelevant for GPUs that most likely will end up in server racks, but hot damn that industrial design is slick 

[igormp]

Just now, n0stalghia said:

This is completly irrelevant for GPUs that most likely will end up in server racks, but hot damn that industrial design is slick 

The A6000 is a workstation card. The one that'll end up in server racks is the A40 (with no video outputs).

[porina]

22 minutes ago, williamcll said:

With the full 10752 cores, this deluxe workstation card that is surprisingly priced at half of a RTX 8000, which has less than half of its cores compared.

Core functions are not equal. As we saw with the consumer models, the cores seem to be tied to a doubling of FP execution units, but not everything was doubled in that process. So, some things will scale as expected with the core count, but not everything will.

 

22 minutes ago, williamcll said:

Thoughts: Shiny indeed, I wonder what in the production process made Nvidia to decide to lower down the price?

If nothing else it tracks the trend started by the consumer models.

[Belgarathian]

This generation of reference designs from Nvidia is baller. Best looking cards I've ever seen.

[Falkentyne]

So who's going to buy this to play video games on it?  I know you filthy rich people are out there.  Get buying!

[AldiPrayogi]

I was confused at the naming first, 6000-series like when did AMD steal RTX naming haha

[tim0901]

5 hours ago, gabrielcarvfer said:

Radeon Instinct MI100.

Nope. Radeon Instincts are compute cards competing with the likes of the A100 - they have no video outputs and are passively cooled.

Spoiler

The A6000 is a Quadro - AMD's competitor for this card is somewhere in their Radeon Pro lineup (WX9100? I can't say I know much about AMD's workstation cards)

[tim0901]

2 minutes ago, AldiPrayogi said:

I was confused at the naming first, 6000-series like when did AMD steal RTX naming haha

Personally I keep getting confused with Sony's camera lineup - the A6000 is one of their budget mirrorless cameras!

[igormp]

1 hour ago, tim0901 said:

Nope. Radeon Instincts are compute cards competing with the likes of the A100 - they have no video outputs and are passively cooled.

Although I get what you meant (and the current MI100 competes with the soon to be released A40), the price between Teslas and Quadros are usually close enough for such comparison.

[StDragon]

8 hours ago, igormp said:

GDDR6X has a kind of error detection builtin.

Assuming that PAM4 uses FEC (Forward Error Correction) to ensure the integrity of the signaling, that's still only within the DIMM module. Error correction doesn't occur between the GPU and RAM like a traditional ECC scheme would call for.

 

Perhaps full ECC isn't needed if Nvidia was confident that all the errors were in the RAM ICs. It would definitely reduce the workload from the GPU.

[Mark Kaine]

so this is a dual slot card with 48 GB? 

 

because the picture in the OP is kinda irritating, as those are *2* cards glued together apparently... 

 

 

And it's 15% faster than a 3090, hmmpf, hard pass.

[leadeater]

12 minutes ago, Mark Kaine said:

because the picture in the OP is kinda irritating, as those are *2* cards glued together apparently... 

They are two cards, they are showing NVLink connectivity. You don't have to use two, but it's good to know you can as that is still relevant in workstation and compute applications

[igormp]

9 hours ago, StDragon said:

Perhaps full ECC isn't needed if Nvidia was confident that all the errors were in the RAM ICs. It would definitely reduce the workload from the GPU.

Lots of people also disabled ECC due to the perf hit caused by it on your regular workstation cards, so maybe that's another factor.

 

7 hours ago, gabrielcarvfer said:

Not really. They don't have video outputs, but they're meant for virtualized workstations. Cheaper than Nvidia solutions, no stupid licensing shenanigans, etc.

They are also way more limited in usefulness when compared to a Quadro/Tesla, since most stuff relies on CUDA. Instinct cards are mostly used for heavy FP64 simulations/analysis, and now those game streaming platforms.

[Mark Kaine]

25 minutes ago, leadeater said:

They are two cards, they are showing NVLink connectivity. You don't have to use two, but it's good to know you can as that is still relevant in workstation and compute applications

yeah, I figured after clicking the link. 

 

this is just all weird to me, they should have done like the Ares III, just put both in an enclosure with a water cooling block ... much nicer to look at and these blower cards must be running hot as hell? 

 

like I get it, it will work but a confusing product nonetheless, are those 48 GB really important? Because otherwise a 3090 / with or without SLI nvlink  makes a lot more sense imo 

[igormp]

3 minutes ago, Mark Kaine said:

this is just all weird to me, they should have done like the Ares III, just put both in an enclosure with a water cooling block

That means an extra point of failure and shorter lifespan, nobody wants that.

4 minutes ago, Mark Kaine said:

much nicer to look at and these blower cards must be running hot as hell? 

Those cards go into closed boxes with no windows, they're not meant as your regular gaming card. Yes, blowers usually do run hot, but when you have multiple of those they end up cooler than multiple open-air models. All workstation models are blower for a reason.

 

5 minutes ago, Mark Kaine said:

like I get it, it will work but a confusing product nonetheless, are those 48 GB really important? Because otherwise a 3090 / with or without SLI makes a lot more sense imo 

It's not confusing, you're just thinking about the wrong use for those. They are meant to render stuff, do heavy calculations or train ML models, not game. 48GB is actually not enough for most tasks, that's why you get 2~4 of those.

SLI is used only for games, it has no other application. For compute scenarios, you don't even need a nvlink bridge, since the software is meant to distribute the work through PCIe without problems, the benefit of nvlink is when you start to saturate the PCIe link and need more bandwidth for inter-memory comms (usually when you go for more than 4 GPUs).

[Mark Kaine]

2 hours ago, igormp said:

you're just thinking about the wrong use for those

No, I know what the use cases are, I just didn't know how it works exactly and how that much RAM is useful. 

2 hours ago, igormp said:

since the software is meant to distribute the work through PCIe without problems, the benefit of nvlink is when you start to saturate the PCIe link and need more bandwidth for inter-memory comms

I see, I didn't know that. 

 

2 hours ago, igormp said:

Those cards go into closed boxes with no windows, they're not meant as your regular gaming card. Yes, blowers usually do run hot, but when you have multiple of those they end up cooler than multiple open-air models. All workstation models are blower for a reason

Ah, ok didn't know that either.

 

 

2 hours ago, igormp said:

That means an extra point of failure and shorter lifespan, nobody wants that.

*except certain gamer folks apparently

[Blade of Grass]

20 hours ago, n0stalghia said:

This is completly irrelevant for GPUs that most likely will end up in server racks, but hot damn that industrial design is slick 

Man, if companies designed consumer desktop products like this, I would be in love. Instead everything is g4m3r with RGB-everything 

3 hours ago, igormp said:

SLI is used only for games, it has no other application. For compute scenarios, you don't even need a nvlink bridge, since the software is meant to distribute the work through PCIe without problems, the benefit of nvlink is when you start to saturate the PCIe link and need more bandwidth for inter-memory comms (usually when you go for more than 4 GPUs).

NV Link is significantly faster than PCIe 4.0 though, so I would expect any well-written library to use it if possible for inter-card data broadcasts. Not to say I know if it does or not, I don't make GPU accelerated workloads, I just expect it would. 

[porina]

20 minutes ago, Blade of Grass said:

NV Link is significantly faster than PCIe 4.0 though

For rough indication only (values are for a single direction, although not 100% sure for the SLI number):

High Bandwidth SLI connector was 2 GB/s

PCIe 4.0 lanes are 2GB/s each, so 32GB/s if you use 16. Halve that for 3.0.

Consumer nvlink as introduced with 20 series cards is 50GB/s and limited to two devices. I believe the full fat version can go to more devices, but I don't know if their bandwidth is same or higher. A quick search suggests 3090 and A6000 both do 56.25GB/s on their links.

 

To compare theoretical bandwidth dual channel ram running at a marketing speed of 3200 is about 50GB/s, but that can only read or write at a given time, not both at the same time.

[igormp]

37 minutes ago, Blade of Grass said:

NV Link is significantly faster than PCIe 4.0 though, so I would expect any well-written library to use it if possible for inter-card data broadcasts. Not to say I know if it does or not, I don't make GPU accelerated workloads, I just expect it would. 

I can't really say for all scenarios, maybe rendering does make use of it. But for ML stuff (something that I do work with), the benefit of nvlink is less than 10% (in a best case scenario) for up to 10 GPUs. When you go past 8~10 GPUs is where things get interesting.

 

5 minutes ago, porina said:

I believe the full fat version can go to more devices, but I don't know if their bandwidth is same or higher. A quick search suggests 3090 and A6000 both do 56.25GB/s on their links.

The full fat nvswitch can do a total of 928 GB/s bidirectional (up to 18 GPUs, each with ~50GB/s). Link multiple of those with an infiniband link and now you made your own DGX cluster.

[Blade of Grass]

20 minutes ago, igormp said:

I can't really say for all scenarios, maybe rendering does make use of it. But for ML stuff (something that I do work with), the benefit of nvlink is less than 10% (in a best case scenario) for up to 10 GPUs. When you go past 8~10 GPUs is where things get interesting.

Interesting, it definitely depends on how broadcast dependent the workload is, vs. if it's easily solvable as discrete subproblems with a final combination step.

I would expect stuff like FFT, LU Factorization, PageRank, and some other workloads would benefit from the bandwidth. 

 

EDIT: heyo, 5k counted posts