What is the nVidia Tesla P4? - Cloud Gaming Server Part 17

Technological Concepts / Product Features Discussed

NVIDIA Tesla P4: identity and hardware

• Small form factor: half-height, passive server GPU.
• GPU core: uses a “full fat” GP104 (Pascal) die with 2560 CUDA cores.
• Memory: comes with 8GB GDDR5 and is described as the same Pascal GPU + memory combination used in products like Quadro P5000 / GTX 1080.

Key P4 constraints vs. consumer / high-end cards

• Power cap: limited to about ~75W (far below typical ~180–200W cards like the GTX 1080).

• Lower clocks by default: around ~1113 MHz out of the box (contrast noted with higher boost on larger cards, e.g., ~1733 MHz).

• Cooling approach: 100% passive cooling, requiring rack/server airflow.

Why it matters for video encoding / cloud

• NVENC blocks: includes two NVENC encode/decode blocks.
• Cloud video positioning: presented as an efficient video encode accelerator—described as the smallest/most efficient encode card in NVIDIA’s 1000-series lineup.
• Virtualization support: claims official NVIDIA vGPU support, enabling partitioned GPU resources.

Performance expectation framing

• The discussion is part of a cloud gaming server series, focusing on how the P4 performs in VM-based gaming workloads, including real-world considerations like cooling and power limits.

---

Review / Guide / Tutorial-Style Elements Emphasized

vGPU support and a crucial limitation

• For vGPU testing, the video references a need for a “vgpu rust unlock” script (promised in the description) to enable custom memory profiles.
• Problem described:
  • The P4’s practical usable memory is about ~7.6GB.
  • Default vGPU profiles are split into fixed even chunks (example: 1024/2048MB), which can leave host memory underutilized.
• Example approach:
  • With the unlock/custom profiles, the tester configures two VMs with 3.5GB each, instead of wasting capacity.

Test methodology / environment

• Comparison: Tesla P4 vs Tesla M40.
• Server platform:
  • 1U GPU server
  • Dual Intel Xeon E5-2698 v3 (16 cores each)
  • 256GB DDR4 ECC
  • Two 800GB NVMe drives
• VM setup:
  • Game storage via iSCSI
  • Each VM uses:
    • 6 CPU threads
    • 16GB RAM
• Workload: GTA V at 1080p, “very high” / worst-case max-like scenario.
• Measured outputs: average FPS, 1% low, 0.1% low, utilization, thermals, and power draw.

---

Core Performance Analysis (P4 vs. M40)

Tesla M40 baseline (virtualized)

• GPU generation: Maxwell.
• Specs noted:
  • 3072 CUDA cores
  • 12GB GDDR5
  • about ~250W TDP
• Single GTA V instance:
  • ~48 FPS average at 1080p very high
  • weak frame pacing:
    • 1% low ~20 FPS
    • 0.1% low ~9 FPS
• Second instance:
  • becomes completely unplayable
  • average drops to ~34 FPS
  • 0.1% low ~9 FPS

Tesla P4 results (virtualized)

Single GTA V instance

• ~60 FPS average
• smoother pacing:
  • 0.1% low ~23 FPS
  • 1% lows stay above ~30 FPS
• Thermals/power noted with one instance:
  • utilization around ~50%
  • peak power about ~40W
  • max temps about ~55°C

Two simultaneous GTA V instances

• Average ~49 FPS (about matching M40’s single-instance averages)
• frame lows remain acceptable:
  • 1% low ~24 FPS
  • 0.1% low ~20 FPS
• Claim: the P4 runs two instances faster than the M40 runs one, while using only about ~50W total power.

Clock tuning “surprise”

• Default P4 boost is restricted to about ~1113 MHz (firmware-limited).
• Using unlock/firmware override allows clocks up to about ~1531 MHz.
• After higher clocks, with two VMs running GTA V:
  • average increases by ~10 FPS (to about ~58 FPS)
  • lows improve:
    • ~28 FPS (1% low)
    • ~24 FPS (0.1% low)
  • power rises modestly to about ~65W total, further weakening M40’s value for this workload.

---

Practical Takeaways / Implied Recommendation

P4’s main advantage for cloud gaming (in this setup)

• Despite low power (~65–75W class) and passive cooling, the P4 provides a better playable experience and more consistent frame timing than a higher-power enterprise GPU (M40), specifically under two-VM concurrent workloads.

Cooling feasibility in non-rack contexts

• Because power draw is low, the video argues it may be cooled using a small fan (mentions ~40mm fan plus a custom/3D-printed setup), making it potentially workable in quieter, more desktop-ish builds than typical passive server cards.

Market price note

• The video notes the P4 has dropped below ~$200 “for the first time in the last month,” positioning it as a strong value target for smaller cloud gaming projects.

---

Main Speakers / Sources

• Jeff — host of Craft Computing (primary speaker running the analysis/tests).
• John Jay — mentioned as a fan who sent a beer (not a technical source).
• NVIDIA — referenced for GPU architecture/features (GP104/Pascal, NVENC, vGPU support).