RetroFlops
Menu

Graphics

PlayStation 4 custom processor (GPU section)

18 compute units on the processor die, rated by Sony at 1.84 TFLOP/s for graphics, simulation, or any mixture of the two.

Manufacturer
Sony Computer Entertainment / AMD
Architecture
Unified array of compute units
Introduced
2013
Count
18
Record status
approved

Figures

PlayStation 4 custom processor (GPU section): recorded figures

provisional A figure marked provisional is recorded but not accepted: its evidence is unsettled. It may rely on one source, an unconfirmed announcement, or a digit read from a damaged scan. The reason appears under the figure. The catalog shows the figure but never uses it in a multiplier. How records are judged.

Metric Value Evidence Basis Method Source
Peak FP32 rate gpu provisional 1.84 TFLOP/s Confirmed Vendor-rated Theoretical peak Stated by the vendor, before the hardware shipped 18 compute units "collectively generating 1.84 Teraflops of processing power", per the introduction of February 2013. Recorded as single precision, which the source does not state: the same editorial reading the PlayStation 3's figures carry, and the same reason for holding it provisional. Sony also gives no clock, so the figure cannot be checked against the 18 compute units it is attributed to.

Derived comparisons

799 000 000 Hz

Formula
peak FP32 ÷ (lanes × FLOP per lane per clock) clock-from-peak-fp32 v1
Inputs
  • ps4-gpu Peak FP32 rate: 1.84 TFLOP/s
Constants
  • lanes: 1152 Eighteen compute units, which Sony states, at sixty-four lanes each. The lane count per unit is not Sony's: ComputerBase gives the PlayStation 4 Pro's thirty-six compute units as 2,304 ALUs while naming the base console's eighteen in the same sentence, which is sixty-four per unit either way.
  • floating-point operations per lane per clock: 2 The universal convention for a single-precision peak: one fused multiply-add per lane per clock, counted as two operations. It is the same convention NVIDIA states in the GeForce whitepapers this catalog already cites, and every vendor rate here assumes it.
Rounding
half-up , 3 significant digits

Divides a published single-precision peak by lanes × floating-point operations per lane per clock, so it is what the vendor’s own figure implies rather than a frequency anybody stated. Reported to 3 significant digits — the precision of the published rate. Both constants are assumptions about the architecture, cited on this claim: get either wrong and the answer scales with it.

Fitted to

Editorial notes

Sony gives the compute-unit count but no clock, so the teraflops figure cannot be reconstructed from the parts — which is the reverse of the usual problem, where the clock is published and the unit count is not. Both this and the PlayStation 3's RSX figure are provisional; the difference is that here the source at least states what the units are.

Sources