News RTX 4000 SFF Ada Generation: Half Height Half Pint Hardware for SFF Workstations & Signage

Full Anandtech Article:

NVIDIA Unveils Ada Lovelace RTX Workstation GPUs for Laptops; Desktop RTX 4000 SFF

www.anandtech.com

Excerpt:

On the desktop side of matters, NVIDIA is launching a single new proviz product for desktops: the RTX 4000 SFF Ada Generation. This part is aimed at the small form factor market, supplanting NVIDIA’s previous flagship low-profile video card, the Ampere-based RTX A2000.

​

The RTX 4000 SFF Ada Generation comes as NVIDIA’s desktop proviz parts are in the middle of a transition of their own. For the desktop market NVIDIA has only launched an RTX 6000 based on their Ada Lovelace architecture, so it’s a bit of a surprise to see them launch a SFF-focused part (and only a SFF-focused part) as their second desktop card. None the less, given the very limited number of half-height cards in the market and the form factor’s hard limits on power consumption and cooling, NVIDIA is betting that cards based on a newer architecture (and smaller process node) will be a welcome sight for the SFF market.

Amazing news, we just got the A2000 and it's already being dethroned:

 

[ChinStrap]

$1250

 

[msystems]

Lmao

 

[Analogue Blacksheep]

So who's doing the first build, lol?

 

[b3rdm4n]

I picked up a cheap A2000 6GB recently, gonna be a long time before these are anywhere near as cheap, but wowza do I want one!

 

[scatterforce]

The original MSRP of the A2000 12GB was $850. Now, you can get one for 600 on Amazon. Used, I've seen folks grab them for half of that.

Give this 2 years and it will be much cheaper. They might even stealth launch a 10GB version of this card at some point.

 

[GuilleAcoustic]

The original MSRP of the A2000 12GB was $850. Now, you can get one for 600 on Amazon. Used, I've seen folks grab them for half of that.

Grabbed mine for 300€ (320€ delivered), with og. box and both brackets.

I'm glad that NVidia is pursuing that dual-slot low-profile form factor .... now we need a non PRO version of this with a better pricing

 

[robbee]

I'm glad that NVidia is pursuing that dual-slot low-profile form factor .... now we need a non PRO version of this with a better pricing

Oh yeah, we need that 75W RTX 4050 again!

 

[ondert]

wow... so it gives rtx 3070ish performance on 70w? It would be perfect setup with HDPlex 250w GaN in a 4 liter case.

 

[smitty2k1]

All I want is an LP GPU with HDMI 2.1 for my 120hz TV...

 

[b3rdm4n]

All I want is an LP GPU with HDMI 2.1 for my 120hz TV...

One of the only downsides of the A2000 and my LG C2 - unless there is any way to adapt DP 1.4a to HDMI 2.1, but my searches all come up with nothing.

 

[GuilleAcoustic]

One of the only downsides of the A2000 and my LG C2 - unless there is any way to adapt DP 1.4a to HDMI 2.1, but my searches all come up with nothing.

Have zero experience with those .... but there seem to be active adatpter that does DP1.4a to HDMI 2.1 : https://www.uptab.com/displayport-1-4-to-hdmi-2-1-8k-60hz-or-4k-120hz-hdr-active-adapter/

 

[robbee]

All I want is an LP GPU with HDMI 2.1 for my 120hz TV...

One of the only downsides of the A2000 and my LG C2 - unless there is any way to adapt DP 1.4a to HDMI 2.1, but my searches all come up with nothing.

I'm using this cable with my A2000, it supports 4K@120hz bandwidth but unfortunately no G-sync: https://www.amazon.nl/-/en/gp/product/B094XDDRL6/ref=ppx_yo_dt_b_asin_title_o00_s00?ie=UTF8&psc=1

 

[smitty2k1]

Have zero experience with those .... but there seem to be active adatpter that does DP1.4a to HDMI 2.1 : https://www.uptab.com/displayport-1-4-to-hdmi-2-1-8k-60hz-or-4k-120hz-hdr-active-adapter/

That looks really nice, if I had an A2000 I'd try it out. Wonder if it supports VRR? My GTX 1650 doesnt have DSC 1.2 though.

 

[wertzius]

All I want is an LP GPU with HDMI 2.1 for my 120hz TV...

You mean something like a RX6400?

 

[smitty2k1]

You mean something like a RX6400?

Yes, good point. In my case it has worse performance than my GTX 1650 since my motherboard is only PCIe 3.0.

 

[Brian_Buckley]

When do we think general availability will be? They said April but not sure if that actually means DIY or just prebuilts since these are usually more enterprise-focused.

I did find one listing on ShopBLT for $1444. Early adopter tax or are they all gonna be that much? (they originally said $1200)

 

[robbee]

When do we think general availability will be? They said April but not sure if that actually means DIY or just prebuilts since these are usually more enterprise-focused.

I did find one listing on ShopBLT for $1444. Early adopter tax or are they all gonna be that much? (they originally said $1200)

I received the official availabity announcement mail from Nvidia at the 3th of April and it has been on ShopBLT ever since. IIRC, it was a bit cheaper at the time though.

If you search for the PNY part number, you'll find a couple other retailers: VCNRTX4000ADALP-PB

 

[scatterforce]

Here is a review of the card's performance.

It's just shy of a 3060ti's performance out of the box. I'm sure with some tweaking, you could get it there... but as expected, it's loud. A solid copper heatsink swap would be great on this.

 

[SFFMunkee]

For what it's worth, a stock RTX A2000 garners about 6,000 graphics points in 3DMark Time Spy.
My modified RTX A2000 managed to pull 9,020 graphics points in the same benchmark (5 mOhm shunt mod front and rear, upgraded VRAM thermal pads, stock cooler, overclock/undervolt via MSI AfterBurner).

While I wouldn't recommend running 24/7 it at full power (estimated at double the standard 70W), it shows the performance overhead available to nVidia's SFF workstation cards.

Given the RTX4000SFF is getting 10,603 graphics points in Time Spy I would be surprised if couldn't easily best the RTX 3060Ti, more likely slightly ahead of an RTX 3070. especially if you're willing to sacrifice a bit of the power efficiency by shunt-modding.

::EDIT::
So I plotted a quick and dirty histogram based on the unique users with TimeSpy results using RTX A2000, keeping only the highest result for each user.

Score Range	Unique User Count
Below 5000	1 (<1%)
5001-5250	2 (<1%)
5251-5500	15 (~4%)
5501-5750	74 (~21%)
5751-6000 (average) (100%)	152 (~43%)
6001-6250 (101-104%)	54 (~15%)
6251-6500 (104-108%)	18 (~5%)
6501-6750 (108-113%)	20 (~5%)
6751-7000 (113-117%)	11 (~3%)
Above 7000 (117-150%)	8 (~2%)

355 Unique RTX A2000 Users
Average Result: 5,991 points
Median Result: 5,884 points

If we applied this same curve to the RTX 4000 SFF... and assume the average is around the 10,603 review score then we could reasonably hope for 12-15% performance increases from overclocking and minor mods, with shunt-modded cards maybe being able to approach 25-50% improvement if we're really lucky.

That would mean..
10600 Average
11024 (15% of RTX4000SFF cards)
11448 (5% of RTX4000SFF cards)
11518 (RTX 3060 Ti)
11978 (5% of RTX4000SFF cards)
12402 (3% of RTX4000SFF cards)
12946 (RTX 3070)
13-15000 (1 or 2 RTX4000SFF cards maybe)

Spoiler: Meh

If you were to look at the graphics score results curve of Time Spy / A2000...

You can see most cards being tested (~741 results) are actually pulling about 5,500 points
These are (probably) your 'normies' - perhaps workstations, cards with limited PCIe connectivity, non optimised settings (e.g. no ReBAR etc), that are just curious to test out their graphics card but haven't modified settings or anything.

The median (~381 results) is around 6,000 points.
These are (probably) cards with stock cooling and settings, or perhaps slightly overclocked / undervolted but happy with what they've got.

There are plenty (~202 results) around 6,500.
This (probably) represents the peak you can achieve without physically modifying the card or cooler. That is, users who have done a fair bit of software tweaking and testing but don't want to risk damaging their card with modifications.

Then... we see about between 15-20 results at each point 7,000 / 7,500 / 8,000 / 8,500 with only a single result at 9,000.
These (probably) represent shunt-modded cards which may or may not also have custom cooling.

Semi-logical analysis to come to estimate overhead available to RTX4000SFF based on 3DMark TImeSpy USERS instead of results.

I think the real question will be, where does the 4060 Ti sit in terms of performance, and how small will they get