Advice on HP Nvidia Tesla K40 GPUs

[datacomplex]

Hello everyone! First post, long-time oclHashcat / John the Ripper/ CRACK (for DOS, remember that?) user. I bought myself a few K40s. So with that --

I am interested in putting together a 4x HP Nvidia Tesla Kepler K40 GPU, model F1R08A (for servers), rig but have not been able to find much information as to benchmarks of hashcat running on K40s. What little information I have found have been other models of the K40 such as the K40c and some others that I cannot verify my cards fall into. To make it more confusing, when I search the K40 I mostly come up with results for the newer model K80.

Does anyone have any information about hashcat running on Tesla K40 (HP model #F1R08A) and what kind of real-world performance they yield?

If it helps, the specifications/marketing material say the card features 1.4 TFlops of performance, 12GB memory, 2,880 Kepler-generation CUDA cores, and 288GB/sec bandwidth. It has a 745 MHz core that can be over-clocked to 875Mhz by the user, it says (I assume via some option ROM settings).

On a side note, I am not too familiar with building machines that house more than one GPU - and this requires two slots of room for each card - so anyone that has any input on what I'm looking at there would be nice. I have a few rack servers that wouldn't have enough room to house these, save maybe two cards in one of my Dell R900s if I'm lucky (unlikely) although I have a tower server w/ a quad-core Core i5 3.2GHz and 32GB of RAM, and the mobo MIGHT have enough space for four of these but unlikely. I am open to suggestions if anyone has any input.


Regards,

datacomplex

[Chick3nman]

The Tesla K40 is an abysmally slow card by today's standards, especially in hashcat, and coming in at 235W TDP per card, you would be FAR better served getting basically any GPU from the current generation than trying to run those cards. Even the 149$ MSRP GTX 1650 will come in at something like 3 or 4x faster than a K40 and only consume 75W. Wont even need supplemental power for it. If that's out of budget, pretty much anything from the 900 Series and beyond will still be a better choice, simply because the architecture includes instructions that give a crucial boost in speed over the Kepler(700) series cards and previous.

[epixoip]

Yeah, you won't find much out there about the K40. Very few people crack with Teslas since they're slower than their GTX (and now RTX) counterparts at 15x the price. And Kepler is especially garbage for password cracking as it lacks LOP3.LUT; the K40 came out when the AMD R9 290X was king. The only reason you see benchmarks for K80 is because they were the only GPU that EC2 offered at the time. But yeah, K40 performance is pretty abysmal... about 12% slower than its sister GK110 the GTX 780 Ti at ~ 11 GH/s NTLM.

Unfortunately you've already bought the cards... Could I convince you to sell them and buy something else? Looks like they're fetching $150-200 on eBay, so if you unloaded all four you could pick up an RTX 2080 which will be ~ 61% than all four K40s combined with much less complexity and power consumption.

Edit: Dammit Chick3nman, you beat me to it. I knew I shouldn't have stopped typing to get a snack :/

[datacomplex]

@epixoip and @Chick3nman, I will go ahead and return them. I got them for $105 a piece or $420 for all 4, flat rate. Other options were the K10 and GRID M40 also for very cheap.

So to understand clearly, am I better served by grabbing two RTX 2080s? And if so, which edition would you recommend as I see many of them by many different manufacturers. Or should I go for a few of those GTX 1650s? I'm open to recommendations.

Thanks for your input guys!

-data

[epixoip]

You're kind of all over the map here. You just went from 4x $105 GPUs to possibly 2x $700 GPUs or 3x $150 GPUs.

I think you need to:

1. Determine your budget, as this will always be the most impactful limitation,

2. Determine your goals are - what are you cracking, and what are you hoping to accomplish?

3. Do some research and look at some benchmarks. The K40 purchase shows you've not done any research on this topic, as if you had, you would have known to snag those M40s (which have performance similar to the Maxwell Titan X) instead of the K40s!

#1 & #3 combined will help you get the best bang for your buck. I highly recommend that you keep things simple, especially as a newcomer. There's virtually no reason to go for 3 or 4 low-end or mid-range GPUs if they'll be outperformed by one high-end GPU. One GPU will also be easier and cheaper to power, and easier to cool. If you can afford two or more high-end GPUs, that's awesome. But you'll need to do some research first, as it's not anywhere near as simple as "buy computer, buy GPUs, stuff GPUs into computer." The Hardware forum is chock-full of information on building a proper rig built to meet the demands of password cracking workloads, you just have to be willing to put in the work as no one here will spoon-feed you answers.

[datacomplex]

I returned those $105 K40 GPUs.

Per the points:

1) Budget, around $2,000.

2) I am currently using solely JtR, (but use hashcat when using Amazon GPUs, and I'd like to move to hashcat with the GPUs) as I have always just scaled w/ CPU on SMP and multi-core across multiple systems with their various fork and node options (I've used JtR for ages) across a few dozen computers in my rack when resource usage is low. I run many different hashes, but primarily I am running: md5crypt, sha256 / sha512crypt, NTLM, WPA/WPA2 (WPAPSK, WPAPSK-PMK), descrypt in some cases (when I'm lucky), mysql PASSWORD() hashes and the "SSHA" (sha1 seeded) hashes.

3) I couldn't find much info on the K40s but had seen what I thought were decent raw speeds, but it appeared that doesn't translate to much in the hashing. I had to pull the trigger to get the deal, but again I was able to return them so it's no big deal.

In some cases, I've just rented out Amazon p3 instances (esp. with their V100s they're using now) if I have a single or a very limited number of hashes to attack and have had good results over-all if I can craft an appropriate charset and mask that will be best for a given hash. I normally use it when hitting WPAs and the $6$ hashes (sha512, 5k rounds I think?) as those sha512s are incredibly tough when trying to hash on CPU.

I will check out the other posts here on the forum and see what GPUs are being recommended now.

Thanks,

data

[Flomac]

Since the new Super-series just arrives with a significant price drop over its predecessor i'd go for two GTX2070 Super. They deliver 3/4 performance of two normal GTX 2080 at 2/3 of their price. And two cards are always easier to handle in big towers, even if they don't have radial fans.

[datacomplex]

Since the new Super-series just arrives with a significant price drop over its predecessor i'd go for two GTX2070 Super. They deliver 3/4 performance of two normal GTX 2080 at 2/3 of their price. And two cards are always easier to handle in big towers, even if they don't have radial fans.

I assume you mean the RTX 2070 Super? According to Nvidia they release July 9th and, indeed as you said, are priced very nicely at the same price of the 2070 back when it released albeit it with more power. They are claimed to be 24% faster in some cases, with an average of 16% performance increase over-all (see https://www.techradar.com/news/nvidia-ge...s-rtx-2070).

At $499 for the "Founder's Edition" (I am not familiar with this terminology although I've seen this "Founder" label slapped on many gaming-related gear over the past year or so), that'll put two of them in my hands for just $1,000 and also provide me with two copies each of the new Wolfenstein and "Control", a $90 value they claim for $180 total; maybe I can sell these downloads or license keys or CDs or whatever for $15 a piece to someone and recoup $60, effectively bringing price to $940 before tax). 

Thanks for the input!

[Flomac]

I assume you mean the RTX 2070 Super? According to Nvidia they release July 9th and, indeed as you said, are priced very nicely at the same price of the 2070 back when it released albeit it with more power. They are claimed to be 24% faster in some cases, with an average of 16% performance increase over-all (see https://www.techradar.com/news/nvidia-ge...s-rtx-2070).

Yes, correct that's them. More shaders & higher clockrate should deliver (at least) 15% more performance under hashcat.