ServaRica Local AI VPS: dedicated Tesla P40 / V100 excellent tok/s for Strata (Qwen3.8-Flash-Next 12

servarica_haniservarica_hani ProviderOG
in VPS

Hi LES!

Today we're introducing our fine-tuned GPU VPS service.

These are large VPS instances with one or more full GPUs passed through to each VPS. The GPU is fully dedicated to your VPS and isn't visible to the host or to any other VPS.

Platform

  • AMD EPYC 7552 hosts
  • Linux or Windows templates with GPU drivers preinstalled
  • Location: [Montreal, Canada]

The specs aren't random. We ran Strata on both cards and built each plan around the RAM and SSD needed to run Qwen3.8-Flash-Next (125B) on a single card at full speed. We also use some of these VPS internally for very large log analysis and chat analysis tasks.

Benchmarks (on our own VMs)

Strata – Qwen3.8-Flash-Next 125B, 128K context

GPU Generation (write) Prompt processing (read)
1× P40 39–41 tok/s ~450 tok/s
1× V100 72–79 tok/s ~1,160–1,450 tok/s

llama.cpp – smaller models run even faster

  • Qwen3.6-35B-A3B: up to 73 tok/s on a P40
  • Qwen3.8-35B-A3B Distill: up to 154 tok/s on a V100

Full benchmark tables, a step-by-step Strata setup guide, and an honest list of what these cards can't do are on each plan page.

Plans (USD/month)

Tesla P40 (24 GB VRAM per card)

Plan Cores RAM NVMe VRAM Price
P40 ×1 10 dedicated 40 GB 500 GB 24 GB $139 Order
P40 ×2 20 dedicated 80 GB 1 TB 48 GB $278 Order
P40 ×3 30 dedicated 120 GB 1.5 TB 72 GB $417 Order

Tesla V100 (32 GB HBM2 per card)

Plan Cores RAM NVMe VRAM Price
V100 ×1 8 dedicated 35 GB 400 GB 32 GB HBM2 $179 Order
V100 ×2 16 dedicated 70 GB 800 GB 64 GB HBM2 $358 Order
V100 ×3 24 dedicated 105 GB 1.2 TB 96 GB HBM2 $537 Order

Bandwidth – every plan lets you choose:

  • 100 TB/month on a 10 Gbps port, or
  • Unlimited transfer on a 250 Mbps port that grows by 1 Mbps per day up to 1 Gbps

Why run this instead of using an API?

  • Privacy: It's your server. Nobody sees your prompts or trains on your documents.
  • Flexibility: Switch models whenever you like.

About servaRICA

servaRICA is a VPS and dedicated server provider based in Montreal, Canada.
We own and operate our own network and infrastructure, delivering reliable service since 2010.

FAQ

  • Terms of Service (TOS) and Acceptable Use Policy (AUP): View Here
  • Accepted Payment Methods: PayPal, Alipay, credit cards, cryptocurrencies (BTC, ETH, etc.)

Thanks

Comments

  • NeoonNeoon OG

    You could try STATA with the lowest tier you offer can post the numbers here? :D

    Free NAT KVM | Free NAT LXC

  • sumosumo

    @servarica_hani said:

    Plans (USD/month)

    Tesla P40 (24 GB VRAM per card)

    Plan Cores RAM NVMe VRAM Price
    P40 ×1 10 dedicated 40 GB 500 GB 24 GB $139 Order
    P40 ×2 20 dedicated 80 GB 1 TB 48 GB $278 Order
    P40 ×3 30 dedicated 120 GB 1.5 TB 72 GB $417 Order

    Tesla V100 (32 GB HBM2 per card)

    Plan Cores RAM NVMe VRAM Price
    V100 ×1 8 dedicated 35 GB 400 GB 32 GB HBM2 $179 Order
    V100 ×2 16 dedicated 70 GB 800 GB 64 GB HBM2 $358 Order
    V100 ×3 24 dedicated 105 GB 1.2 TB 96 GB HBM2 $537 Order

    Bandwidth – every plan lets you choose:

    • 100 TB/month on a 10 Gbps port, or
    • Unlimited transfer on a 250 Mbps port that grows by 1 Mbps per day up to 1 Gbps

    The pricing is reasonable, given the specs and electricity costs. You can do a lot with a single gpu

    I've used variants (lower ram) of both tesla cards locally in the past. They're both quite usable. You can buy used cards to run locally. Just keep in mind the heat generated (you need good airflow), maybe not a bad thing given its getting cold in the northern hemisphere. If you go v100, I've had luck with the server version with an adapter board.

  • AnthonySmithAnthonySmith AdministratorProviderOG

    damn, that's an interesting offer!

    TierHive - Hourly VPS - NAT Native - /24 per customer - DE, UK, SG, CA, USA x4, FR x2, AU, PL, NL, JP
    FREE tokens on sign up, try before you buy. | Static Hosting Free for life: https://tierhive.com/static-hosting/

  • servarica_haniservarica_hani ProviderOG

    @Neoon said:
    You could try STATA with the lowest tier you offer can post the numbers here? :D

    the numbers here and here are actual numbers on these VPS
    https://servarica.com/plan/683/
    https://servarica.com/plan/688/

    Strata – Qwen3.8-Flash-Next 125B, 128K context
    GPU Generation (write) Prompt processing (read)
    1× P40 39–41 tok/s ~450 tok/s
    1× V100 72–79 tok/s ~1,160–1,450 tok/s

    actually the V100 was tested with slightly less ram so we will do the test again with exact ram and update the numbers which i assume will be better

    Strata with Qwen Moe model really gave life to lower GPU tiers

    These numbers are from actual long tasks we currently run the cards on
    and this model Qwen3.8 FN is much better than I initially expected

  • NeoonNeoon OG

    Did you test with all under load? Memory bandwidth still shared on the Host though.
    My assumption is, that if each card is under load with STRATA, the numbers will be lower.

    Free NAT KVM | Free NAT LXC

  • servarica_haniservarica_hani ProviderOG

    @Neoon said:
    Did you test with all under load? Memory bandwidth still shared on the Host though.
    My assumption is, that if each card is under load with STRATA, the numbers will be lower.

    good question
    no I didnt
    I assume the difference will not be big but it make sense to test it
    will do it and report back

Sign In or Register to comment.