Click the card below to follow the " CVer " public account
AI/CV heavy dry goods, delivered as soon as possible
1. Not much nonsense:
This reader benefit is sponsored by AutoDL.com, and the rules of the event:
For the top 100 readers who send emails to [email protected] , AutoDL will reply the 100 yuan voucher activation code directly by email. After registering on the AutoDL official website (www.autodl.com or www.gpuhub.com), you can use the voucher without any threshold. coupons .
The subject of the email can be made by yourself. In order to prevent the black production of mining, the content of the email must include the following:
1. Name
2. School name
3. Research majors and directions
Second, get down to business:
In 2022, how cheap can it be to rent a cloud GPU? How long can you rent a GPU for 100 yuan? Please see AutoDL's answer sheet to the student parties:
(The following calculations are calculated based on the price of AutoDL Level 3 membership. AutoDL will send you a 1-month Level 3 membership by default. There are questions about the membership system below)
(The following configurations are all equipped with local SSD, that is, the host is a pure solid-state server, most of which are NVMe solid-state)
1. TESLA A100 80GB：
The configuration is shown in the figure; the unit price is ￥7.98/h, and the 100 yuan can be used for more than 12 hours.
(This is the NVLINK version, the whole machine is equipped with the standard DGX-A100, and the available video memory for renting 8 cards will be linearly expanded to 640GB)
2. TESLA A40 48GB：
The configuration is shown in the figure; the unit price is ￥2.88/h, and the 100 yuan can be used for more than 34 hours.
(A40 may be the most low-key in this generation of professional cards, with a single-precision computing power slightly higher than 3090 and A100 48%, half-precision computing power higher than V100. AutoDL will soon launch A40 with NVLink bridge blessing, The video memory can be expanded to 96GB, do you really want to try it?)
3. TESLA V100 32GB：
The configuration is shown in the figure; the unit price is ￥2.38/h, and the 100 yuan can be used for more than 42 hours.
(V100, as the flagship professional card of the previous generation, must be familiar to everyone. This is the NVLINK version, which can be expanded linearly by renting multi-card video memory)
4. RTX 3090 24GB：
As shown in the figure, the unit price is ￥1.69~￥2.28/h, and 100 yuan can be used for more than 59 hours at most.
(AutoDL provides a variety of supporting CPU specifications for the contemporary card king: Inner Mongolia A district has ￥2.28/h per card with 15-core AMD 7543, Shenzhen A district has ￥2.28/h per card with 14-core Intel Gold 6330, Beijing A district There are ￥2.28/h per card with 24-core AMD 7642, and some ￥1.88/h~￥1.99/h with Intel Silver CPU in Inner Mongolia A District and Nanjing Novice District. The most cost-effective is the Tidal computing power area, priced at ￥1.69/ h 3090, equipped with 256GB memory, I suggest students to search how to use the memory as a hard disk, manual dog head)
5. RTX A5000 24GB：
The configuration is shown in the figure; the unit price is ￥1.69/h, and it can be used for more than 59 hours for 100 yuan.
(Many students of A5000 may not have heard of it, but it will definitely become popular! Compared with the 3090, it also has 24GB of video memory, but it is only slightly inferior in single-precision computing power. The tidal computing power area has the same configuration and is priced at 1.58/ h's A5000, really fragrant)
6. RTX 2080Ti 11GB：
The configuration is shown in the figure; the unit price is ￥1.08/h, and it can be used for more than 92 hours for 100 yuan!
(I won't go into details on the card king, the point is that you can use CUDA_10 to run the old algorithm)
7. TESLA P40 24GB：
The configuration is shown in the figure, each GPU is equipped with 12-core Intel E5-2650v4, 64GB memory; the unit price is ￥0.99/h, and the 100 yuan can be used for more than 100 hours!
(The computing performance is slightly weaker than the 2080Ti, but with 24GB of video memory, it can run the old CUDA_10 algorithm that consumes more video memory)
8. TITAN Xp 12GB：
Outrageous here! Compared with the unit price of ￥0.52/h, although the CPU and memory configuration are not high compared to the above-mentioned graphics cards, they are irrelevant. 100 yuan can be used for nearly 200 hours!
(TITAN Xp is the flagship desktop of the previous generation, AutoDL positioned it as a low-threshold debugging card for users. After debugging on Titan Xp, students can use the instance migration function to migrate the code to a card with higher performance to accelerate)
The above only lists the more numerous models of the more than a dozen different graphics cards in AutoDL. There are more than 2,000 GPUs of various types on AutoDL, even including Cambrian and Ascend.
3. Regarding everyone's worries and anxiety, AutoDL solves it like this:
1. Wouldn't it be a waste of money to debug code with a cloud server? ——AutoDL provides no card mode, that is, only one CPU core is used to create a container environment without GPU, the price is 0.1 yuan/h, and online debugging is not anxious.
2. If the training set is too large, will it be very laborious to upload? ——AutoDL supports downloading training data from Alibaba cloud disk and Baidu network disk, with a speed of up to 10MB/s. With the card-free mode, data anxiety is eliminated. For very large data, it also provides a physical mounting service for mobile hard disks.
3. Why are servers in the “Tidal Hash Power Zone” so much cheaper than other zones with the same configuration? Is it bad quality? ——The server in the AutoDL tidal computing power area is the periodic idle resource of the B-end client in the large model training task interval, which is as stable and reliable as the server in other areas. The only difference is that the instance data retention period in the Tidal computing power zone is only 7 days, while the other zones are 30 days.
4. The price of Level III membership seems to be more cost-effective, but what if there is no membership? Will there be any pits in the membership system? ——AutoDL is mainly for student parties. New users will get 1-month level III membership. After completing the student certification through the education mailbox, they can obtain long-term level II members. The discount for level II members is only 0.5% lower than that for level III members. The threshold for upgrading to Level III membership is also very low. You can recharge at least 300 yuan within three months (it will be reduced to 150 yuan after June 1). For details, please refer to the documentation: https://www.autodl.com/docs/member/. If you do not have an educational email address, you can also send a private message to customer service to complete the student certification.
5. I have never used a cloud server, how can I get started? - See the AutoDL help documentation, all you need is here:
https://www.autodl.com/docs/
6. Don't know how to choose the graphics card? ——There are different models of graphics card performance evaluation reference:
https://www.autodl.com/docs/gpu_perf/
4. Are you celebrating May Day?
Happy holiday~!
AutoDL has prepared a recharge and cashback event for May Day, from April 25, 2022 to May 5, 2022, recharge:
1. 100~199 yuan: Free membership III is valid for 3 months
2. 200~999 yuan: 10% voucher for the recharge amount
3. 1000+: 15% voucher for recharge amount
There are also red envelopes raining in the AutoDL alchemy user group~ Scan the QR code of customer service WeChat to join the group~
I wish the working people a happy holiday! I wish teachers and students a happy alchemy! Good luck to all CVer research!
Click to read the original text to access AutoDL !