Playing with RTX 5000 Ada (AD103): first impressions

Playing with RTX 5000 Ada (AD103): first impressions

Picture from the Internet

Mining and uses for mining are not discussed in this article.

I have a couple of old articles (about the A100, and about the 3090 and the A10) and there is also an old, but still good and relevant comparison of the Selectel settlement cards. About two years have passed since then, and it’s time to write something new and try new calculation accelerators. Yes, this is another article about an Nvidia card, not AMD and not Intel, and not about the Chinese and some fashionable big chips, unfortunately.

Since then, new cards of two new generations have appeared – Ada Lovelace and Hopper. At the same time, Hopper was supposed to replace the very successful generation of Ampere (there is still an old tradition of Nvidia – a super successful generation is followed by a less successful one), but they did not come. But perhaps due to trade wars with China, Hopper generation cards began to be issued only to the right vendors and in the form of assembled systems, and at the same time export to China was prohibited.

In this article, we will analyze the first impressions of the card NVIDIA RTX 5000 Ada Generation (chip AD103) and let’s add a little humor to start.

The elephant in the room

Probably, taking into account the history of the articles, you will ask, why, if I am so impressed with Gigabyte’s 3090 Turbo card (low heat, fast, relatively cheap), this article is not devoted to the 4090 or 4090 Ti card (if it is already out)?

Let’s answer with pictures from the article:

Dimensions of 4090 Ti

Rumors about the size of the 4090 Ti

Rumors about the size of the 4090 Ti

This is of course a joke, 4090 cards only look like this:

Actual size 4090

Okay, here’s a real comparison:

Actual sizes

Probably Founder’s Edition cards

Game options

That is, it turns out that different 4090s occupy from 3 to 4 slots and eat somewhere around 450 watts. There are threads and press releases about “normal” size cards, but in our area I have only seen such options – fully on water and hybrid.

There is little desire to test the water and hybrid ones, it is not clear whether they will be accepted for placement in DC, and somehow it is a bit fast, the product will most likely be produced in small series, like the ThreadRipper Pro processors of its time. There is also little desire to disassemble and do the cooling yourself, you just want to buy and forget.

A few words can be said about the new connector and rumors that the 4090 is on fire. Probably, a new connector was made, because the power supply from the motherboard and the old PCIe 8-pin + PCIe 6-pin connector for 450+ watts was no longer enough, which, together with a sharp increase in TDP, also puts sticks in the wheels of ordinary users. For example, a popular 2 kW power supply unit does not have such connectors in principle, and power supply units that have it are designed for a maximum of 2 cards.

Well… the hint is clear. If you even solve “water” questions, you will have to put two cards instead of four, which is not very attractive, even taking into account the growth of “parrots”.

The culprit of the celebrations

The “culprit” of the celebration

In general, joking aside, it is obvious that Nvidia, against the background of the AI-bubble and the success of the Ampere generation, simply divided the cards into three segments – Hopper for the countries of the golden billion and the ultra-rich, late professional RTX cards of the Ada Lovelace series (news that- they were supposed to be from 2022, but this is not accurate, but they came out as a result at the end of 2023) and game ones, which suddenly became very difficult to use for calculations on basic components, without server hardware or without getting “inside”.

In general, during a more detailed selection of components and communication with suppliers, I paid attention to the following document (perhaps a VPN will be required), which puts all the dots on the i:

A plate from the document

Well, that is, in simple words, the previously voracious Quadro (230 W) and 3090 (350 W) cards, which only needed PCIe 8-pin + PCIe 6-pin power, lived perfectly on the old connector, but here are the new cards that eat from 130W to 300W , they all require the same connector as the 450-watt 4090. Note that at least 70 watts are supplied from the motherboard. Some will say standardization, others will say hand-wringing.

At the same time, there are adapters, but judging by the description, such an adapter is not just a wire, but some logic is sewn into it. There is no photo of such an adapter in the documentation, but they come with video cards.

Hidden text

In the sum we get, in order to use 4090 for calculations equal to 3090, you need to solve the following questions:

  • Physical size. Apart from factory or self-made water or a hybrid, I don’t particularly see tracks;

  • An adapter must be included or purchased, as power supplies with the appropriate plugs are only designed for 2 cards maximum;

  • The maximum power supply limit will need to be set at 300-350 W, which will most likely result in a drop in performance of at least 20-30% (4090 owners please give exact figure);

All this makes the purchase of 4090 for calculations unattractive. We seem to get x2.5 parrots, but lose in the number of cards, or lose 20-30% if we heroically solve all problems, but get moments with support.

Dry characteristics

Yes, well, let’s say Hopper cards cannot be bought for adequate funds because of the US trade war with China. Let’s compare at least a few “interesting” cards of the new generation (for A6000 and L* series cards, the price of one parrot increases several times more) with the old ones, but at the same time we will output the chip identifier:

Judging by the official press releases (and still can’t find Ada on the main PNY A5000 website) (1, 2, 3, 4) and the article on Wikipedia, the information in places contradicts each other, but in general the general direction is clear.

It probably doesn’t matter how many cores are actually there in any card (there are several versions of them and their filling has been changed several times already), the number of parrots can be compared in a real test. Professional cards this time also lagged behind gaming cards, but if A10, RTX 3090 and RTX A5000 were relative counterparts, now it seems that A5000 Ada is much more convenient, despite the cable.

It also seems that for the 4090, they did not add a lot of memory, and all cards higher than the A5000 are already very overpriced, probably this is also a form of “trade” war. Also, I didn’t understand which process is used, 4 or 5nm for some of the latest cards, but it probably doesn’t matter.

First impressions

The card is of a standard size, there is an active centrifugal “hair dryer”, outwardly, PNY’s design decision is somewhat similar to reference cards from Nvidia, previously they looked more modest. The card needs a new power adapter.

Power adapters

All adapters are nearby, from bottom to top – new, Tesla 8-pin, Quadro 8-pin, GeForce 8+6 pin.

The culprit of the celebrations again

The culprit of the celebration is two

The culprit of the celebration is three

The card itself

The map has some kind of design, which is strange. Kind of a mix of PNY and FE design.

Calculation of parrots

The most interesting. To maintain consistency with past tests, we will use a simple utility gpu-burn, which is also very useful for stress testing video cards. Now there is a pre-assembled build through docker, you can not select basic images and do not assemble “yourself”:

git clone https://github.com/wilicc/gpu-burn
cd gpu-burn
docker build -t gpu_burn .

Next, the documentation asks to implement the following command: docker run --rm --gpus all gpu_burnbut it will run the default test on all maps at once and the server will beg to heaven. Let’s try what’s easier first:

docker run --rm -it --gpus '"device=0"' --entrypoint bash gpu_burn 
./gpu_burn 120 
./gpu_burn -tc 120

We run it on the selected card, run the test for a couple of minutes with or without the use of tensor kernels. I couldn’t find any new, super interesting and non-obvious dependencies here (the cards are active), so I’ll present the results in a backwards compatible format with all the cards I’ve ever tested.

I will indicate the maximum parrots and the maximum temperature of the card during the test. No dances with a tambourine in the form of overclocking, limiting the power of the card or working at the speed of a “hair dryer” were performed, everything is stupid out of the box. Tests where you had to dance with tambourines with settings, I simply exclude from the sample.

Parrots

Map temperatures

A100 using MIG technology, i.e. “cutting” the card into virtual “sub-cards”

It is clear that these tests do not reflect any workload, in part it is just a backwards compatible way to compare the cards that I got my hands on and make some conclusions about their qualitative and quantitative characteristics. I will not write obvious things, the trends and comparisons here are quite clear.

Conclusions

In principle, there are no revelations here. Nvidia Corporation, remaining a de facto monopoly in the market for computing accelerators, is more and more efficient in extracting monopoly rent from this market by dividing it into new segments through price discrimination.

And if at the very beginning of the server process with cards like 1080 Ti or 3090 they were little inferior to “professional” solutions and did not add, apparently, only memory, now other numerous levers are already used: direct sanctions, memory restrictions, “innovations” in power supply, vendor – Loki.

The only consolation here is probably that if you divide the ruble price of cards that are not included in the “sensitive” list of top products by parrots, then in the last interesting cards this indicator remains at about the same level, but the “density” of parrots is growing, but to assemble a powerful server, it is no longer enough to simply buy components from the store across the street.

Related posts